Академический Документы
Профессиональный Документы
Культура Документы
This book is devoted to the study of Fuzzy Reasoning as applied to decision making and
control processes. It contains a collection of important contributions covering a wide weU-
selected range of topics within the field. The first pioneering work on fuzzy sets and fuzzy
logic was published in 1965 by Zadeh. This work, which has broken down the classical two-
valued logic, was not very much appreciated until Zadeh has published his subsequent
papers in 1971, 1972 and 1974. These works are now considered as the foundation of fuzzy
reasoning and have inspired numerous researchers to work in the field. The first noticeable
practical application of fuzzy logic was done at Queen Mary College of London by
Mamdani. This work concerns the control of a pilot scale steam engine, and has shown for
the first time that fuzzy logic offers the right tool for the construction of a linguistic model
of a given process on the basis of the operators' experience. From that point on, fuzzy logic
theory has found a tremendous attention with a plythora of theoretical and practical
accomplishments.
Today one can distinguish five main areas of research on fuzzy reasoning, namely:
analysis of relational systems (Ostergaad, Rutherford, Perdrycz), fuzzy optimization /
decision making (Gupta, Willaeys), design of fuzzy controllers (Kickert, Mamdani), and
mixed fuzzy logic - neural network techniques (Kosko, Keller, Yager). Most of the current
work on fuzzy sets and applications is currently conducted in Europe and Asia (China,
India, Japan).
TTie present book contains twenty one chapters, written by thirty four distinguished
contributors and divided in the following five parts:
Part 1: General Issues
Part 2: Neuro-Fuzzy Systems
Part 3: Fuzzy Controllers
Part 4: Fuzzy Reasoning and Estimation
Part 5: Applications
Part 1 involves four chapters providing background material together with useful
techniques for the validation of fuzzy knowledge bases and the software representation of
fuzzy sets and fuzzy logic. Part 2 presents an overview of neuro-fuzzy expert systems along
with an important case study, and a neural network model which is suitable for fuzzy
reasoning. Part 3 presents the state of art of fuzzy controllers, including design and
implementation aspects. This part starts with a chapter on the demystification of fuzzy
control and includes critical evaluations of fuzzy controllers, along with new types of fuzzy
controllers such as the sliding mode fuzzy controller. Part 4 involves a chapter on fuzzy
parameter and state estimation, which is of fundamental importance in a variety of
applications, a chapter on fuzzy reasoning as used in rule-based systems, and a chapter on
computing the multivariable shape of an n-D pattern class. Finally, Part 5 presents six
important applications dealing with industrial robotic systems, mechanical systems,
manipulators with artificial rubber muscles, Petri nets, biomedical engineering, and
nondestructive fruit collection.
Although the literature on fuzzy sets and fuzzy logic theory and applications is now
very rich, the editors feel that the present book provides an important addition, since it
presents new angles of attack and includes many new topics and results not available in
other books. Each chapter is self-contained and in many cases involves fresh results and
how-to-do issues. The book would not have been possible without the enthusiastic support
of the contributors. The editors are indebted to all of them for their up-to-date
contributions, and to Kluwer's (Dordrecht) editorial staff members for their care
throughout the editorial and printing process. The editors also express their gratitude to
the Natural Sciences and Engineering Research Council of Canada for support through an
international scientific exchange award which made possible the preparation of their own
contributions.
The book is suitable for the researcher and practitioner, as well as for the teacher
and student in related Master and Doctoral courses.
Spyros Tzafestas
Anastasios Venetsanopoulos
CONTRIBUTORS
GENERAL ISSUES
CHAPTER 1
Fuzzy set theory was initiated by Zadeh in 1965 [1] and permits the treatment of vague,
uncertain, imprecise and ill defined knowledge and concepts in an exact mathematical way.
Throughout the years this theory was fully studied and used for the analysis, modelling and
control of technological and nontechnological systems [2-21]. Actually, our life and world
obey the principle of compatibility of Zadeh, according to which "the closer one looks at
a 'real' world problem, the fuzzier becomes its solution". Stated informally, the essence of
this principle is that, as the complexity of a system increases, our ability to make precise
and yet significant statements about its behaviour diminishes until a threshold beyond
which precision and significance (relevance) become almost exclusive characteristics.
Fuzzy controUers and fuzzy reasoning have found particular applications in
industrial systems which are very complex and cannot be modelled precisely even under
various assumptions and approximations. The control of such systems by experienced
human operators was proved to be in many cases more successful and efficient than by
classical automatic controllers. The human controllers employ experiential rules which can
cast into the fuzzy logic framework. These observations inspired many investigators to work
in this area with result being the development of the so called fuzzy logic and fuzzy rule-
based control [3, 8, 9, 14, 18, 20].
The purpose of this chapter is to provide a short account of fuzzy set and fuzzy
reasoning theory in order to help the unfamiliar reader to study and understand easier the
rest of the book. The reader who is familiar with the fuzzy sets can probably find here a
ready-to-use material for his(her) applications. Section 2 presents the basic concepts and
definitions of fuzzy sets. Section 3 reviews the three fundamental fuzzy logic operations of
Zadeh and section 4 provides a set of other fuzzy operations and relations. Section 5
presents a generalization of the three fundamental operations, and section 6 introduces the
concept of hypercube as used in fuzzy theory. Section 7 presents the representation
theorem, discusses the fuzzy functions (domain of definition and domain of values) and
states the fuzzy extension principle that helps in the fuzzification of mathematical concepts
and laws. Section 8 provides a brief discussion of categories and lattices in the framework
of fuzzy sets, and section 9 examines the theory of fuzzy reasoning (linguistic variables,
3
The concept of "set" plays a fundamental role in mathematics. Actually one cannot define
in a rigorous unique way what is a set, what is number or what is a straight line [10]. These
concepts can be understood better not via definitions but via examples [11].
Let X be a classical set and x an element. Then one of the following holds (how
much this holds will be discussed soon): The element x belongs to X (symbolically xeX)
or X does not belong to X (xtX). This is the so called principle of dichotomy. By
disputing dichotomy, the classical (crisp) set theory breaks down and the fuzzy set theory
naturally emerges [12].
f 1 if xeA
The question now arises whether fi^{x) can take only the values 0 and 1 or any value
between 0 and 1. In other words, why ju^:A-»{0,l} and not ^^-•[0,1] ? This question was
firstly examined by L. Zadeh and is the starting point for the development of fuzzy sets [1].
Definition 2.1
Let X be the reference superset and A a subset of X. Then A is said to be a fuzzy subset
of X if and only if
A = {ix,fi^{x)) I xeX,Ai^U):X-[0.1]}
It is obvious that in the special case where we have {0,1} instead of [0,1] the fuzzy subset
A degenerates to the crisp subset A.
Pictorially a fuzzy set A has the representation of Fig.l.
fuzzy subset A
crisp subset A
A = \uj^(x)/x where the symbols" + ", "Z", "J" denote union and the symbol"/" does
Definition 2.2
Let ^ be a fuzzy subset of the reference superset X. Then
(a) The crisp subset Supp(/1) of X is called support of A if and only if
Supp(-4) = { x e X : / i ^ ( x ) ^ 0 }
with Supp(/4)cX.
(b) The crisp subset L^ of X is called a-cut of A if and only if
L^A = {xeX : n^(x)ia}
with L ^ c X .
Definition 2.3
The fuzzy subset A of X is called
(i) convex if and only if (see Fig.2):
(ii) norma] if and only if there exists at least one element xeX for which n^(x) = \.
\if.W ,, fVW i
(a) (b)
Definition 2.4
Let A he a fuzzy subset of the reference superset X. Then
(i) The quantity | A | = Y,^gx I^Ai'') ^^ '^^^^^ ^^^ cardinality of A
The cardinality measures in a certain way the "size" of a fuzzy set. Using the cardinality one
can compare fuzzy sets that have the same reference superset.
Example 2.1
(a) Let X=N and "A = integers near to 10". Then one possible A is the following:
A = {(7,0.1) , (8,0.5) , (9,0.8) , (10,1) , (11,0.8) , (12,0.5) , (13,0.1)}
= 0.1/7 + 0.5/8 + 0.8/9 + 1/10 + 0.8/11 + 0.5/12 + 0.1/13
and has the pictorial representation of Fig.3.
(b) Let X=R and "A = real positive numbers near 10". Then one possible A is
^2^-ll
7 8 9 10 11 12 13 X
liXx) i
The three fundamental operations of fuzzy sets are based on the use of the membership
function [1].
Definition 3.1
Consider the fuzzy subsets A and B of the reference superset X. Then the section and
union of A, B are defined as
and the complement A'^ oiA as A' = Ux,Hj^,{x)) \ xeX , A'^c(Jt) = l-\x^{x)\
Example 3.1.
Let X = {apartments with 1,2,..., 10 rooms},
A = {apartments suitable for 4-member families},
B= {apartments with large space}.
The sets A and B are fuzzy subsets of X. Suppose that
A = 0.2/1 + 0.5/2 + 0.8/3 +1/4 + 0.7/5 + 0.3/6 and
B=0.2/3 + 0.4/4 + 0.6/5 + 0 . 8 / 6 + l / 7 + l / 8 .
Then
C=Ar\B= {apartments suitable for 4-member families and with large space}
= 0.2/3 + 0.4/4 + 0.6/5 + 0.3/6
D = /\u_B= {apartments suitable for 4-member families or with large space}
= 0.2/1-1-0.5/2 + 0.8/3+1/4-1-0.7/5 + 0.8/6+1/7-1-1/8
A'^= {apartments not suitable for 4-member families}
= 0.8/1 + 0.5/2+0.2/3 + 0.3/5 + 0.7/6
B*-= {apartments with smaU area} = 1/1+1/2+0.8/3 + 0.6/4 + 0,4/5 + 0.2/6
Remark: The fundamental axioms of the Aristotelian logic do not hold in fuzzy logic, i.e.
in general AnA'^^0 and AuA'^^X . This can be easily seen from Fig.5.
Definition 4.1.
Let A and B fuzzy subsets of X. Then we define the following:
(i) Algebraic sum A + B
A+B = [Cx,iJ^^g(x)) I xeX , I^^.BM = fxjx) + n^U) - n^(x)\ig{x)]
Fig. 5. Membership functions of A and A*^.
Example 4.1
Let X= {1,2,3,4,5,6,7} and A = 0.8/3 + 1 / 5 + 0.6/6 , B= 0.7/3 +1/4+0.5/6.
Then
>iuB=0.8/3 + 1 / 4 + 1 / 5 + 0.6/6
/ i n B = 0.7/3+ 0.5/6
A''= 1/1+1/2+0.2/3+1/4 + 0.4/6
AB=0.56/3+ 0.3/6
>i2 = ^ s ^ = 0.64/3+ 1/5 + 0.36/6
A«J3=l/3+l/4+l/5+l/6
v4oB=0.1/3+1/5 + 0.1/6
2/1 = 0.8/9+1/25+1/36
10
4.2. Fuzzy Relations and Fuzzy Cartesian Product
Definition 4.2
Let Xj , X2 , ... , Xj, be reference supersets and A^, Aj , - , A^^ respective fuzzy subsets
of them. Then as a fuzzy relation f{A^ , ... , A^) we define the fuzzy set:
/ M l . A ) = { i(x^,...,xj , fifi4„.A.)(^t, ,xj ) I (x^,...,xJeX^x xX^ ]
where
crup relacion R
f(x.x^=0
Definition 4.3
Let Xp...,X^ be reference supersets and Ap...^„ respective fuzzy subsets of them. Then the
Cartesian product of A^,...yA^ is defined to be the fuzzy set
g(A^,...,A^) = { ((AC,,...,J:J . ^tg^A^,,^/x^,...^^)) | (x^,...,x^)eX^x•••xX^ }
where
fg(A,....A,)(^i'-'''n) = niin{/i^(x,),.„,/i^(xj)
Figure 7 shows a crisp and fuzzy Cartesian product. One can see that the crisp Cartesian
product is 2-dimensional but the fuzzy one is 3-dimensional.
11
^ J (lAlxAl (X1,X2)
'^
fuzzy A 2
cnip
xi
fuzzy Ai
\ '' \ /
fuzzy A-xAo
Example 4.2
Let Xi = X2=(2,4,6) , Ai = 0.5/2+1/4 + 0.6/6 and ^^2= V2 + 0.6/4. Then
AixA2=0.5/(2,2) + 1/(4,2) + 0.6/(6,2) + 0.5/(2,4) + 0.6/(4,4) + 0.6/(6,4)
Definition 5.1
The function T:[0,l]x[0,l]-»[0,1] is called T-norm if and only if Tpossesses the following
properties for all x,y^€[0,l]:
1. ^x,y) = ^y^)
2. T{x,y)i'Jlx,z) yy^z
3. T{x,T{y,z)) = TXTix,y)^)
4. T{x,l) = \
12
Definition 5.2
The function T :[O,l]x[0,l]-*[O,l] is called T-conorm if and only if T possesses the
following properties for all x,y,2e[0,l]:
1. r{x,y) = r{y,x)
2. T (x,y)iT (x,z) \iy<.z
3. 'f{xX{y,zy)-'f{'f{x,y)^)
4. T(x,0) = 0
Definition 5.3
The function N;[0,l]-*[0,1] is called negation function if and only if N possesses the
following properties for all x,ye[0,l]:
1. N(0)=1 , N(1) = 0
2. N{x)<N{y) Vx^y
Throughout the years there have been proposed several T-operators. Some of them are
listed in Table 5.1.
x-y x+y-2-x-y
4. x + y-x-y 1 -x-y \-x Bandler
xy j_ (\-x)-il-y)
6. max{x,y,X) max(l --V, l - y , A ) l-x Dubois
Actually, not any single triple of T-operators is the best in all decision making problems.
The simplest triple is the one suggested by Zadeh, but it is not the most suitable in many
cases. For a detailed discussion of T-operators the reader is referred to [15].
Definition 6.1
Let A = [ (x,fi^(x)) ( xeX , p^(jc):>4-[0,l] } be a fuzzy subset of the finite reference
Clearly, the fit vector contains (gives) all the information that is provided by the set theory
approach. Also, the pair (X,[0,1]") defines a basic metric space in the fuzzy theory.
TTie set [0,l]"=[0,l]x...x[0,l] is called unit/j_>j3erci;be of dimension n, and forn=l,
n = 2 and /j = 3 has the geometric representation shown in Fig.8.
(0.1,1) (1.1.1)
L _ (1.0.1)
/
kL
(0.0) (1.0) (Pfl,0) (1.0,0)
Using the concept of hypercube one can represent both the crisp sets as their
vertices, and the fuzzy sets as internal points.
The above can be clarified by the following example. Let X= {x^,X2} be a reference
superset, and P(x) the set of all subsets of X, i.e. P(x)= {0,{x^^2}'{^i}>{^2}}- Now consider
the following fuzzy subset of X: i4 = {(Xj,l/3), (Ar2,3/4)}. The fit vectors of P{x) are as
follows:
i) The set e> or {(x^fi), (^2,0)} has fit vector (0,0).
ii) The set {^1,^2} or {(Xpl), (^2'0} ^^^ ^^ vector (1,1).
iii) The set {Xj} or {(.yj,!), (%0)} has fit vector (1,0), and
iv) The set {;r2} or {(x^O), (x^j-^)} ^^^ ^^^ vector (0,1).
v) The fit vector of the fuzzy set A is (1/3, 3/4).
Let us now represent the above fit vectors in the hypercube [0,1]^ (here n = 2 since
X has two elements). This is shown in Fig.9. One observes that the crisp sets are
represented at the vertices, and the fuzzy sets at internal points of the hypercube.
14
(X2l=(0.1) Xi.Xi=(l,l)
fuzzy subset A
A=((Xi, 1/3) , (X ,23/4)1
O=(0.0) ix,i=(l,0)
Fig. 9. Representation of the fit vectors in the hypercube [0,1]^.
(0,0) (0,1)
From Fig.lO it is now obvious that the sets A, A^, AnA*^ and AuA"^ occupy symmetric, with
respect to the centre M, positions in the hypercube. If A = (0,1), then A'^ = (1,0),
AnA'^ = (0,0) and AuA'^ = (l,l), i.e. the sets A, A^, AnA"^ and AuA^ coincide with the
vertices of the hypercube. If A = M=(l/2,l/2), then ^"^=(1/2,1/2), AnA'' = (1/2,1/2) and
^uy\*^=(l/2,l/2), i.e. the sets A, A'^, AnA^ and AuA'^ all lie at the centre of the hypercube.
Thus for minimum uncertainty (i.e. 0,1) the four sets coincide with the hypercube vertices,
and for maximum uncertainty (i.e. 1/2) the four sets coincide with the centre M.
15
The centre M of the hypercube behaves in a singular way since it possesses the
paradox property: M=M^=MnM^=MuM^ !!! Thus the centre M is considered as the
"black hole" of the fuzzy set theory. In other words, at the midpoint of a hypercube nothing
is distinguishable, while at the vertices everything is distinguishable. Actually, one can
easily verify that the logic paradoxes discussed in Sec.2.1 which according to the classical
logic contain contradictions are solvable with the fuzzy logic and geometrically lie at the
centre of the hypercube [16]. Indeed in these cases, if S is a proposition and S^ is its
complement we have S='S^ and S^=*S i.e. S'^S^. Now, if t{S) is the truth value of 5 then in
the classical logic we have t(S) = 1(8*^ i.e. 0 = 1 (contradiction). But in the fuzzy logic we
have t(S) = t(S^) or f(S) = l-f(S^) whence t(S) = t(S^= 1/2. Thus S and S^ are half true and
half false and the paradox ceases to exist.
Ml = E A ' . W
which coincides with the Hamming-distance and measures the distance of the fuzzy set A
(in the hypercube) from the vertex (0,0) as shown in Fig. 11.
(X2)=(0,l) lx,x^=(l,l)
fuzzy subset A
A=((x,,l/3),(x^/4))
O=(0,0) ix,)-,.,-,
F*ig. II. Geometric representation of the cardinality.
Definition 6.2
The generalized distance of order p ^ l of two fuzzy sets A and B is defined by
L''{A,B) = EIM^.)-/'B(*<)K
16
For p=2 the distance L\A,B) is the fuzzy Euclidean distance between A and B, and for
p= 1 the distance 0(A,B) is the fuzzy Hamming distance between A and B. In the special
case where A is a fuzzy set and B=« (the empty set), then
E(A)
• fuzzy set A
A=((x,,l/3),(x ^/4))
II 1 ^1
O=(0,0) |x,|=(l,0)
Fig. 12. Geometric representation of L'(A,Apgj,f) and L^(A,Aj-3j).
Theorem 6.1
The fuzzy entropy E{A) of the fuzzy set A is equal to
EiA) - l^^-^^l
IAUA^I
17
Proof
The proof will be given with the aid of the hypercube concepts. For simplicity we use n=2
(Fig. 13). By definition E(A) = a/b. Due to the symmetric position of the fuzzy sets A, A'^,
AnA^, AuA'^ one can see that a = a'and b=b', i.e. a'= |>\rL4'^| and b'= \AuA'^\. Thus
J5(A) = £ = i ! = l^^^'^l
b hi \A\JA'
Remark: \{ A is a crisp set then obviously Ar\A'^=0, i.e. |>4n4'^|=0 and A<JA'^=X, i.e.
|>\uA'=|=n and £(A) = 0/n = 0.
(1.0) (1,1)
— b
A first idea is to approximate the fuzzy set with a sequence of a-cuts where Osa^l. Then
it is obvious that LQA=X and a s ^ •• L ^ ^ ^ .
The approximation problem is formulated as follows. Let A, B, C,... fuzzy sets and
L^, L^, L^C,... their a-cuts. As a first step in the approximation one works with L„A,
L^, L^C,... which are crisp sets, and from them produces a set, say ^ , for some ae[0,l].
This is repeated many times, and so from the given families of the a-cuts L^, L^, L^C,
... a sequence of results ^ is produced for all a, ae[0,l]. The question now arises: Does
there exist a fuzzy set Z which would give the above family of sets ^ for all a, Osasl in
the a-cut sense? The answer to this question is positive as stated in the following theorem.
Theorem 7.1
Let ^ , Osasl be the family of crisp sets with the following properties:
i)oZ=X
ii) a s ^ - ^ i ^
CO
For n= 1 the extension principle coincides with the concept of "image of a fuzzy set" (see
sec.7.2) i.e.
y= {iy,^iYiy)) I y ^ " ^ . M;f(y):"5^-[o,i] }
where
It is noted that / i s always applied upon the elements of the crisp set X.
20
8. CATEGORIES A N D LATTICES
8.1. Categories
Category is a complex mathematical concept which consists of (i) a class of objects <£, and
(ii) a set of morphisms from A to B, for every two objects A,Be^ (symbolically (it{A,B) ).
The object of class ® may be sets, groups, topological spaces and so on. In the special case
where the objects are sets, the category consists of; (i) the class ® of sets, and (ii) the set
of aU possible functions from A to B (symbolically &iA,B) or f:A-*B) for every two sets
A,Be^. The following table shows a sample of objects and their respective morphisms and
structures.
8.2. Lattices
The interval [0,1] and its properties plays a central role in the concept of fuzzy set.
Definition 8.1
Lattice is defined to be a partially ordered set L that is associated with an ordering
operator possessing the properties of reflexivity, antisymmetricity and transitivity and for
which for every two elements a,beL there exist two new elements (aAb), (aVb) eL, which
are called minimum and maximum respectively, such that a/\b<.a, b^aVb.
The set [0,1] is a lattice according to the above definition under the ordering
operation <. and the usual min and max concepts.
Definition 8.2
Let L be a lattice and X a crisp set. Then as L fuzzy set we define the set
A = {(Ar,^^(x)) \xeX,fi^{x):A^L }
Clearly, the L-fuzzy set is a generalization of the standard fuzzy set where the lattice
i s L = [0,l].
21
At this point a generalization of the fuzzy set concept at higher level will be
provided [18].
Consider a fuzzy subset of X, namely
A = { (x,p^(x)) I x€X , ^iAix):A^[0,l] }
The main issue is that the fuzzy set is fully characterized by the membership function [ipfiC)
which for every xeX returns a number.
An obvious generalization would be to assume that fi^(x) has the form ^^(x):X-*S
where S= {set of sets}, i.e. the standard //^(x) returns a number from the point set [0,1]
whereas the generalization /i^(x) returns a set from S.
Definition 8.4
The set A is called C-fuzzy set (or type-2 fuzzy set) if and only if
A = { (x,ju^(x)) I xeX , p^(x):A-^S={set of sets} }
Example 8.1
We are given that: "The number 3 is small with degree of membership 'about 0.8'". Here
there is uncertainty not only on the elements Xj but also on the degrees of membership
^^(A-y), because the statement 'about 0.8' is a fuzzy set.
9. FUZZY REASONING
Definition 9.1
Linguistic variable is a variable the values of which are not numbers but words or sentences
or propositions in a natural or artificial language.
Definition 9.2
Linguistic variable is the quintuple <x,T(x),U,G,M> where x is the name of the linguistic
variable, T(x) is the set of its values {term set), U is the reference hyperset upon which
T(x) is structured (recall that the values of a linguistic variable are fuzzy sets), G is a
syntactic rule that generates the names x, and M is a semantic rule that gives sense
(meaning) to the names.
Example 9.1
Consider the linguistic variable X-"age" with T(x) = { very young, not very young, pretty
young, a little old, pretty old, not very old, very old }. Then a meaning of the fuzzy set "old"
is M„,d={(u, t x „ , » ) , ue[0,100]}, where
, «e[0,50]
f^ou("^ =
1+
^r , «e[50.100]
The linguistic variable "age" and its values are pictoriaUy depicted in fig. 14.
linguistic variable
values of the
linguistic varial
15 20 25 30 35 40 45 50 55 60 65 70 75 80
reference superset
Fig. 14. Linguistic variable "age" and its "linguistic values".
Definition 9.3
A linguistic variable x is said to be structured if its set of values T^x) and the set of
meanings M(x) that are given to them can be determined algorithmically.
Example 9.2
Consider the linguistic variable "truth" with the following set of values: T(truth) = {true, not
true, very true, not very true, false, not false, very false, not very false}. It is clear that TJx)
can be determined algorithmically, and so the variable "truth" is a structured variable.
23
Example 9.3
Let A be an element of the term set of a linguistic variable. Clearly, A is fuzzy set. Then
the most common linguistic modifiers applied upon A are:
i) very(A) = con(>\)
ii) more or less (^) = dil(yi)
iii) slightly(^) = Int[(y4) and not (very (A))]
The grammatical correspondence of linguistic modifiers are the adverbs and some
adjectives. Classical examples are the following: very, very much, pretty, nearly pretty, not
at all, almost not at all, etc.
Let us now permit the fact not to coincide with the antecedent of the implication,
but to be "alike". For example
Fact: Tomato very red
Implication: IF (Tomato red) THEN (Tomato mature)
Conclusion: ?
Of course a conclusion can be drawn by a strict application of the MP rule, but it is a real
life conclusion that now the tomato is very mature, something which does not strictly
coincide with the conclusion, but it is "alike" the MP conclusion. On the basis of the above
observation one can state a Generalized Modus Ponens (GMP) rule as:
Fact: A'
Implication: IF A THEN B
Conclusion: B'
where now A', A, B' and B are fuzzy sets.
It is true that the GMP rule mimics much better the human type of reasoning than
the classical MP rule, since it can work not only under idealy identical conditions but also
under general conditions of similarity (alikeness). This is very useful! since reduces very
much the number of rules that are required to have in the knowledge base of the
application.
Definition 9.4
Let A be a fuzzy set on X, B a fuzzy set on Y, and R a fuzzy relation on XxY, i.e.
A = { (x,fiAix) I xeX , ^ ^ ( x ) : A - [ 0 , l ] }
B= { {y,^B(y) |yey,^B(y):5-[o,i] }
Example 9.4
Consider the foUowing: (a) X={1,2,3,4}, (b) A = "x smaU"={(l,l), (2,0.6), (3,0.2), (4,0)},
(c) R = "x nearly equal to y with matrix
25
R X
1 2 3 4
1 1 0.5 0 0
2 0.5 1 0.5 0
3 0 0.5 1 0.5
y
4 0 0 0.5 1
It is desised to find what happens with the variable y, i.e. to determine the fuzzy set B="y'.
To this end, use of the max-min composition rule of inference is made, i.e.
B=A°R, where
Thus, IF "x smaU" and "x nearly equal to y THEN "y nearly small".
d(£)
PA(^) = 1
supd
d/i^(x)
X/z^(x)[l-/i^(x)]
dx
which upon integration gives the membership function
1 + exp(a-bx)
The constants a and b are determined by the other data of the problem.
_ MA(^I)
We now form the matrix P=[Py] which possesses the following three properties:
•P«=landp^.= l / p . .
• All eigenvalues of P are zero except one which is equal to the cardinality n of A.
• The eigenvector w that corresponds to the nonzero eigenvalue n has as its
elements the values of the membership function corresponding to the elements of
A, i.e. w=\ii^{x;)\ = \ii^{x{) ... iiA{x„)V-
Of course, this method is applicable to discrete finite fuzzy sets, not to continuous fuzzy
sets that have an infinite number of elements.
Here a brief description of the structure of the basic fuzzy control loop will be provided.
Fuzzy control uses linguistic variables (this is why it is sometimes called linguistic control)
and mimics the human action more closely than traditional control. For comparison, Fig. 15
depicts the architecture of both the traditional and the basic fuzzy control loop.
The problem in the traditional control loop is to design a controller which will
accept the error e=r-y as input and will give an output (control signal) u such that the
output y of the overall system follows the set point (reference input) r as nearer as
possible.
27
r ^ Controller Process
(a)
Rule Base
inference mechanism
(b)
terms terms
defuzziflcation fuzziflcation
Z^ Process
The problem in the fuzzy control loop is to design an inference mechanism (fuzzy
controller) which will mimic the human type of syllogism and will ensure a desired
performance of the overall system. The fuzzy control loop involves the following blocks: a
rule base, an inference mechanism and the fuzzification/defuzzification blocks. The
fuzzification/defuzzification blocks have been discussed earlier. The fuzzy knowledge base
contains the rules that are to be used for the control of the process. These rules are usually
the result of interviews with the expert operators (very rarely come out of mathematical
analysis or simulations) and have the form IF-THEN. In the general case the rules have
many inputs and many outputs (MIMO). However, it can be shown [20] that a set of
MIMO rules can be transformed to a set of MISO (multiple input - single output) rules.
As an example here we consider rules with two inputs and one output, i.e. rules of
the form
Rji IF "x is A{ and "y is B{ THEN "z is C{
A rule of this type has two assumptions (premises) 'x is A{ and "y is B{ which are
defined on the cartesian product t/xV. The whole rule constitutes a fuzzy relation in the
fuzzy Cartesian product [/xVxW, i.e. fi^^ = /'(^^^a, nwi c,/"'^-^) where U, Vand
W are the respective fuzzy sets.
An easy to use method for the calculation of the relation R makes use of the min
operator, namely
28
R= U Rj
Suppose now that at a certain instant of time we have observed in the process that 'x is
A" and "/ is B". The problem is to combine this fact with the rules of the knowledge
base such that to produce a suitable control y. To this end one must use the max-min
composition scheme, i.e.
n n
(A',B')O U /?,- = U (A',B^)'>RI
j=l /=I
The right hand side of the above identity suggests that instead of applying the fact (A',B')
n
to the knowledge base U Rj as a whole, one can apply {A',B') to each rule Rj of the
1=1
knowledge base separately. The proof of the validity of the above identity can be found in
[20]. In this way, for each input, one can evaluate the individual contribution of each rule
to the final result. This facilitates very much the control process, reduces considerably the
computational load and helps to have a better monitoring which rules are fired and how
much are fired.
11. CONCLUSIONS
Scientists and engineers recognise the fact that "fuzziness" is inherent everywhere in the
real human life. This is particularly true whenever the human factor has a direct and active
effect (social issues, economical issues etc.) Actually, the effort to solve uncertain problems
by classical (deterministic) mathematical theory may lead to "deadends". In general,
mechanistic type systems can be treated by quantitative analysis, while humanistic type
.systems are too fuzzy to be amenable to such exact mathematical methodologies. An
elegant approach to handle such uncertain situations is the fuzzy set theory which has now
arrived at a very mature state of expansion and application.
In this chapter an effort was made to give within a strictly limited size an
appreciation of fuzzy set and fuzzy logic theory. Further concepts and developments can
be found in the references and in the rest of this book.
REFERENCES
1. L.A. Zadeh: Fuzzy Sets, Inform, and Control, vol.8, pp.338-353 (1965)
2. L.A. Zadeh: Fuzzy Algorithms, Inform, and Control, vol.11, pp.323-339 (1969)
This page intentionally blank
29
3. S.S.L. Chang and L.A. Zadeh: On Fuzzy mapping and Control, IEEE Trans. Sys.
Man Cybern., vol. SMC-2, pp.30-34 (1972)
4. L.A. Zadeh: Outline of a New Approach to the Analysis of Complex Systems and
Decision Processes, IEEE Trans. Sys. Man. Cybern., vol. SMC-3, pp.28-44 (1973)
5. L.A. Zadeh: A Computational Approach to Fuzzy Quantifiers in Natural
Languages, Comp. & Maths., vol.9, pp. 149-184 (1983)
6. L.A. Zadeh: Fuzzy Logic and Approximate Reasoning (In Memory of Grigore
Moisil), Synthase, v.30, pp.407-428 (1975)
7. L.A. Zadeh: Commonsense Knowledge Representation Based on Fuzzy Logic,
Computer, pp.63-65, Oct. 1983.
8. R.M. Tong: A Control Engineering Review of Fuzzy Systems, Automatica, vol. 13,
pp.559-569 (1977)
9. P.M. Larsen: Industrial Applications of Fuzzy Logic Control, Int. J. Man-
Machine Studies, vol.12, pp.3-10 (1980)
10. M. Black: Reasoningwith Loose Concepts, Dialogue, vol. 2, pp.1-12 (1963)
11. R. Bellman and M. Giertz: On the analytic Formalism of the Theory of Fuzzy
Sets, Information Science, vol.5, pp.149-156 (1973)
12. B. Gain: Precise Past - Fuzzy Future, Intl. J. Man-Machine Studies, vol.19,
pp. 117-134 (1983)
13. S. Haack: Do we Need Fuzzy Logic?, Intl. J. Man-Machine Studies, vol.11,
pp.437-445 (1979)
14. R.M. Tong: Analysis and Control of Fuzzy Systems Using Finite Discrete
Relations, Int. J. Control, vol.32, (1977)
15. M.M. Gupta: Theory of T-norms and Fuzzy Inference Methods, Fuzzy Sets and
Systems, vol.40, pp.431-450 (1991)
16. B. Kosko: Neural Networks and Fuzzy Systems: A Dynamical Systems Approach
to Machine Intelligence, Prentice Hall, 1992.
17. C.V. Negoita and D.A. Ralescu: Simulation, Knowledge-Based Computing and
Fuzzy Statistics, Van Nostrand Reinhold, 1987.
18. C.V. Negoita and D. Ralescu: Applications of Fuzzy Sets to Systems Analysis,
Birkhauser Verlag, Basel 1975.
19. L.A. Zadeh: Fuzzy Lo^c, IEEE Computer, pp.83-93, April 1988.
20. CC. Lee: Fuzzy Logic in Control Systems: Fuzzy Logic Controller (Part I),
IEEE Trans. Syst. Man Cybern., vol. SMC-20, n.2 (1990)
21. T. Takagi and M. Sugeno: Fuzzy Identification of Systems and Its Applications
to Modelling and Control, IEEE Trans. Syst. Man Cybern., vol. SMC-15, pp. 116-
132 (1985)
CHAPTER 2
Abstract
1 - Introduction
31
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 31-49.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
32
and compatible with the statement "X is A'". The fuzzy set representing the available
knowledge is interpreted as a [0,l]-valued possibility distribution which restricts, in an
elastic way, the more or less possible values of a variable, or more generally of a tuple of
variables. If this possibility distribution is not normalized, it means that the available
knowledge is (partially) inconsistent by itself since there does not exist any fully possible
interpretation (i.e. possible at the degree 1). For instance, in possibilistic logic (Dubois,
Lang and Prade, 1989, 1991b), the complement to 1 of the height of the possibility
distribution representing the whole knowledge base defines the degree of inconsistency of the
knowledge base. Although inference under partial inconsistency can still be performed in
possibilistic logic, checking the consistency of a fuzzy knowledge base remains an
important issue for validating the base. The validation of fuzzy knowledge bases has received
little attention until now, with the exceptions of Hall et al. (1988), Yager and Larsen
(1990), Larsen and Nonfjall (1989,1991), Turksen and Wang (1992), Kinki6ia6 (1992), and
Dubois and Prade (1991b).
which represents the contents of fK-. Then by projecting TT^ over X j , . . . , Xj,, we obtain
the restriction (if any) which is induced by n^ on the input variables ; this is called
"reflecting on the input" by Yager and Larsen (1991). The knowledge base fPO will be said
to be coherent with respect to input data if this projection does not induce any genuine
restriction on the value of the input; namely, ^ is said to be coherent if
Vui, ..., Vuk, supvj . .v^ 'txj Xk, Yi Y / " 1 . •••' "k- ^l- •••. v^) = 1. (3)
Indeed, if (3) does not hold, it means (in case of finite or bounded domains that there exist
input data (u^j u \ ) such that TCX, Xu. Y, YA(U^1 " \ ' ^ 1 ' •••> v.?) < 1,
Vvj, ..., Vv^, i.e. that there exist input data that together with 7i<j^ make a partially
inconsistent knowledge base, since the corresponding possibility distribution is no longer
normalized. Note that (3) entails that, for each X,, proj(7t^,Xj) is the possibility
34
distribution uniformly equal to I on the domain of X;; but the converse is false, it is not
sufficient to have the property for each projection on Xj, i = 1 ,n to insure (3).
In practice, any tuple of input values is not always allowed. For instance, let us imagine
that the variables Xj, ..., Xj^ pertain to the description of people, e.g. Xj = age, X2 =
weight, then the data Xj < 1 year, X2 > 20 kg are not feasible. For detecting input data
which would not be meaningful, integrity constraints describe the tuples of values which
can be encountered and accepted. Let IC be the subset, defined on the Cartesian product of
Xj, ..., X|(, of the tuples of values which are acceptable. Let us assume first that IC is an
ordinary subset. Then the coherence of ^ with respect to input data which satisfy the
integrity constraints is naturally expressed by modifying (3) into
V (ui u^) e IC, supvj„..v^ JtXi X^X»l "k- ^i, ••-, v^) = 1. (4)
It may be natural to allow for flexible integrity constraints, which enables us to distinguish
between input data which are completely possible from input data which are not completely
impossible but are more exceptional, letting IC be a fuzzy set. Then we may think of
extending (4) in two different ways :
- by requiring that (4) holds for all tuples {\i\,.... uj^) belonging to the support of IC, i.e.
In this case, the introduction in ^ of input data, even if they are almost (but not totally)
impossible according to IC, will not create any inconsistency, but the fuzziness of IC is
not taken into account by (5). However the fuzzy nature of IC may still be used for
making clear to the user how much the input looks exceptional and has to be checked ;
Vul,...,u,^,^lIC(ul Ufc) < supvj^ v^ ^Xi XfcCui. •••. Uk' ^ i . . . - v^). (6)
With this condition, which is weaker than (5), it is allowed that the low level of
possibility of encountering some tuple of input data can be already acknowledged by JK
itself provided that the level of possibility induced by X is at least equal to the one
given by the integrity constraints. However, with (6) there may exist input data with a
non-zero degree of membership to IC which make ^ partially inconsistent (when added
to^).
35
2.3 - Impossibility Qualification for Specifying Integrity Constraints
The (fuzzy) set IC can be obtained as the result of the logical conjunction (using min
operation) of elementary integrity constraints. Usually elementary integrity constraints are
Slated in terms of impossibility rather than in terms of what is possible, like for instance,
"X is A is forbidden, is impossible", or more generally "X is A is almost impossible"
(where X is a variable or a vector of variables).
The integrity constraint C "X is A is impossible to the degree a" can be translated as a
constraint on the induced possibility distribution Jt;;^ describing the allowed values of X (the
conjunction of such possibility distributions leading then to the definition of m^).
i.e. (when a = 0 there is no constraint on ir^ and when a = 1, all the values in A are
excluded as expected)
ii) starting with the usual definition of the possibility of a fuzzy event leads to write
i.e. Vu, 7:'c(u) < [HA(U) -»(I - «)] with a - ^ b = ( ' ' f ^ - ' '
lb if a > b.
Allowing the greatest possibility degrees compatible with this constraint yields the
possibility distribution Jt't; pictured by a dotted line in Figure 1, i.e.
't'c=I^A^(l-")• (8)
36
When a = 1, n'l^ is nothing but the complement of the support of A (the set of
elements with non-zero membership grades)
1-a
* • X
ijtc
Figure 1
Figure 1 exhibits the difference between n^= I - min((iy^, a) andrt't^= n^^ —> (1 - a).
In the translation of "X is A is almost impossible" by n'l^, the gradual nature of A is
lost (even if n't;^ does not correspond to the membership function of an ordinary subset
when 0 < a < 1) ; Tt'i; corresponds to a rather drastic solution which relates the
(ordinary) subset concerned by the impossibility to the level of impossibility a, via n^.
We may thus prefer, to K'I^, the first interpretation n^, which reflects the gradual nature
of A.
In Dubois and Prade (199la) a typology of fuzzy "if... then..." rules of the form "if X is
A then Y is B", where A and/or B are fuzzy sets has been proposed. A distinction is made
between
- gradual rules whose intended meaning is of the type "the more X is A, the more Y is B"
and which are represented by a possibility distribution Kyix obeying the inequality
.v«(v.«)..A<.)-.B(v).!;;'(i:fr^B':> m
- certainty rules whose intended meaning is of the type "the more X is A, the more certain
Y is B" and which are represented by a possibility distribution JCyiX
which indeed expresses that "Y is B is |iA(u)-certain" (according to the result already
applied in (7));
- possibility rules whose intended meaning is of the type "the more X is A, the more
possible Y is B" and which are represented by a possibility distribution Ttyix obeying the
inequality
rtYix(v,u) > min(^A("). M-B(V)) (11)
which indeed guarantees, for any input UQ, that the possibility that Y = v for all the
values V such that \i^(y) = 1 is at least equal to HA("O)-
Thus, we can never establish that (3) does not hold (i.e. that 3u, supy 7tYix(v>u) < 1)-
For simplicity, we first consider the case of two parallel rules of the form "the more X
is Aj, the more certain Y is Bj", i = 1,2. We have
Then
provided that B j and B2 are normalized (i.e. supy llg. (v) = 1 = supy [I'Q^iy]) and where
The necessary and sufficient condition of coherence (14) for two rules can be equivalently
written
core(B j) n core(B2) ^ 0 or support(A j) n support(A2) = 0 (14A)
with supporl(Aj) = (u, 1x^(1) ^ ^) ^nd core(Bj) = (v, |1B.(V) = 1}. It expresses that the
condition parts of the rules have to be completely disjoint or the conclusion parts of the
rules have to be fully consistent. It can also be expressed as : whenever the supports of A j
and A2 intersect, the cores of B j and B2 should intersect too.
A simple example of certainty rules are rules expressed by possibilistic formulas (see,
e.g. Dubois, Lang and Prade, 1989) of the form "if p then q with certainty a" where p and q
are classical propositions. Such a certainty rule is represented by the possibility dismbulion
Tt on the set of interpretations 9,
where 1J.M(') denotes the {0,1) -characteristic function of the set of models of a proposition.
Thus a possibilistic certainty rule is a particular case of a fuzzy certainty rule, identifying
l^M(p) ^ ' * I^A 3"d fn3x(|J.f^^(q\, 1 - a) with jig. Then for two rules "if pj then q; with
39
certainty a;", the condition (14) only holds as soon as M(pj) n M(p2) ^^ 0 entails M(q]) n
M(q2) * 0. For instance the two rules "if p then q with certainty a" and "if r then —iq with
certainty 3" cannot be coherent together as soon as p and r are not mutually exclusive since
then cons(Bi,B2) = niax(l - a, 1 - p) < 1 (except if min(a,3) = 0).
Let us now consider the case of a knowledge base % made of n parallel rules. Then
Let P be the partition induced on the domain of X by the supports of the Aj's, i = 1 ,n.
Each subset in the partition is of the form (Hie j support(Ai)) o (Rig i support(Aj)) where
I c [l.nl. Thus if we consider an element u belonging to such a subset, we have
Moreover, noticing that 1 - IIAUFC") = I => 1 - HA(") = ' ^ ^ that hgt(B) > hgt(B n G),
VG, it can be checked that the only way to have proj(% ; X)(u) = 1 for
u e (Hie I support(Ai)) n (Djg I support(Aj)) is to have hgtdHie I ^i) = 1- Indeed in order
to have (15) equal to 1, we only have to consider the terms such that 1 - \i.[\. J A ( " ) = ^
which entails that the index set J is a subset of f, the complement of I in CI ,nl, but these
terms are of the form
min(l - n y . ^ jAi(u). hgtCflkg J B^)) <min(l - u y , ^ JAJC"). hsKflie I Bj)) = hgt(ni£ J Bi),
40
using (16). Thus we have the following proposition.
Proposition : A necessary and sufficient condition for a base of parallel certainty rules of the
form "if X is Aj then Y is Bj", i = l,n to be coherent is that
This proposition provides a simple procedure for checking the coherence of a family of
parallel certainty rules.
More explicit conditions can also be given when n is sufficiently small. For instance,
forn = 3, (15) writes
hgt(Bi 0 8 2 0 8 3 ) = 1
or hgt(Bi o B2 o 83) < 1, hgt(Bi o B2) = hgt(B2 o B3) = hgt(Bi o B3) = 1,
A] o A2 o A3 = 0
or V i #j 9t k, hgt(Bi o Bj) = hgt(Bj o Bj^) = 1, hgt(Bi o Bj^) < 1, Aj o Aj^ = 0
or V i ^ j ;* k, hgt(Bi o Bj) = 1, hgt(Bj o Bj^) < 1, hgt(Bi o B|^) < 1, hgt(Bi(^) = 1,
(A; u Aj) o Ajj = 0
or V i ^ j ^ k, hgt(Bi o Bj) = 1, hgt(Bj o Bj^^) < 1, hgt(Bi o B^.) < 1, hgt(Bij) < 1,
Ak = 0
or hgt(Bi o B2) < 1, hgt(B2 o B3) < 1, hgt(Bi o B3) < 1, hgt(Bi) = hgt(B2) = hgt(B3) = 1,
(Aj u A2) o (A2 u A3) o (A) u A3) = 0
or V i 5* j ;i k, hgt(Bi) = 1, hgt(Bj) = I, hgt(Bi o Bj) < 1. hgt(Bk) < 1,
(Aj o Aj) u Aj^ = 0
or V i ^ j * k, hgt(Bi) = 1, hgt(Bj) < 1, hgt(Bk) < 1, Aj u Aj^ = 0
where fuzzy set intersection and union are defined using min and max respectively. As we
can see the condition is already much more complicated than in the case n = 2, even if it can
be made simpler with the natural hypothesis that the Bj's are normalized, i.e. Vi,
41
hgt(Bj) = 1. Namely, in the nonnalized case, the above condition expresses that if the cores
of 81,62,83 have not a common part (hgt(Bi 0 8 2 1 ^ 83) * 1), the three parallel certainty
rules are coherent if and only if Aj n A2 n A3 = 0 , and as soon as hgt(8j n B^^) < 1,
Ajpi Aj. = 0 holds, Vj, Vk, or if we prefer, ni=i,3 support(A;) ^0=» ni=i,3 core(Bi) *
0 and support(Aj) n support(A|j) * 0 => coreCBj) n core(Bi^) * 0. We recognize the above
Proposition applied to three rules.
with |iA-(u) -* M.Bj(v) = 1 if VLA^^) ^ \^Bi<y) ^nd ^lAj(u) -> ^Bi(^) = I^Bi(v) otherwise.
Thus (17) is equivalent to
with I*y = [i, HA(U) > 0}- T^'S 'S still equivalent to
VJcIl,nl,Vaie(O,l].niej(Ai)ai^0=*niej(Bi)aj#0 (19)
where A^, B^^ denote the a-cuts of the fuzzy sets A and 8 (AQ( = {u, ^ A ( " ) - "D- These
coherence conditions have been ateady pointed out in (Dubois et al., 1988).
This is a necessary condition for coherence (already suggested in Dubois and Prade (1982)).
However it is not sufficient as shown on Figure 2.
such thai Vi=l,n-1, Vu, ^ A ( " ) "•• l^A- i(") = ^ (which entails hgtCAj n Aj^j) = 1/2) are
used on the domain of X and the Bj's make a similar fuzzy partition.
Figure 2 : Condition (18) does not hold : {v I (0.3 (v) > IIAI (")) '^ t'^' l^Bj^"^) - M-A7(")l = 0
The case where the Aj's and the Bj's have monolonically increasing (or decreasing)
membership functions is much simpler. If ^y^ , ji^^, jig , and lUg, are all increasing, then
the two corresponding gradual rules are coherent (provided that the Bj's are normalized). Let
us consider the situation where JJ-AI > M-B > I^Ao ^ ^ increasing while jig^ is decreasing. In
this case hgt(Ai n A2) = I if Aj and A2 are normalized. Then due to (20) it is necessary to
have hgt(Bi n B2) = 1 in order to have the coherence, this is also sufficient to insure (18)
here, as it can be easily checked. Indeed the two rules "the larger X, the larger Y" and "the
larger X, the smaller Y" can be coherent only if Y increases and decreases in different
subdomains.
4. Redundancy
Indeed if (21) holds, we have min(minj_2 ^ TIJ, n) = TC^ and thus n brings nothing new to
% . Moreover if % is coherent in the sense of (3), the addition to ^ of a redundant piece
of knowledge leaves the knowledge base coherent obviously. Let us examine the meaning of
(21) in case of knowledge bases made of parallel certainty, or of parallel gradual rules.
First, consider the case of the redundancy of one certainty rule "the more X is A, the
more certain Y is B", with respect to the other certainty rule "the more X is Aj, the more
certain Y is Bj". (21) writes
and then, provided that A is normalized, we get ^ 3 - < p-g and provided that 3v, |a.B(v) = 0,
we get lO-A - HAI • ^-^^ redundancy is, as expected, equivalent to A c Aj and B | c B. If
infy Hg (v) = X > 0, the redundancy conditions write Bj c B and (1^1 - m'n(UA'' " ^)'
or equivalently, n ^ ^ (1 - ' ' • ) - * M^Ai ^^^ a - > b = l i f a < b and a ^ b = b otherwise.
Let us now consider the redundancy of a certainty rule with respect to a set of two other
certainty rules, i.e. (21) writes
Vu, Vv, mini=i_2 max(l - HAJ(U), |iBj(v)) ^ max(l - n^Cu), PB(V))- (23)
we are certain that the rule is not redundant. Besides, if the rule is already redundant with
respect to one of the rules in the base, the rule is obviously redundant with respect to the
base. Moreover the two rules "the more X is A;, the more certain Y is Bj", i=l,2, entail, on
the one hand, that "the more X is Aj and A2, the more certain Y is B j and B2", and on the
other hand, that "the more X is Aj or A2, the more certain Y is Bj or B2", since it can be
easily proved using distributivity and obvious boundings that
Thus, if the rule "the more X is A, the more certain Y is B" is redundant with respect to one
of the two above rules which are consequences of the knowledge base, the rule will be
redundant with respect to the base (by transitivity of <).
Thus the redundancy of a certainty rule with respect to one of the two certainty rules in a
knowledge base, or with respect to the rules generated from this base (either by intersection
of the conditions and of the conclusions, or by their union), entails the redundancy of the
certainty rule with respect to the knowledge base in the sense of (23). However the converse
is generally wrong. The situations is better for rules with non-fuzzy condition pans. Indeed
if A J and A2 are crisp subsets, the content of the knowledge base is represented by
liBinB2(v)''""e Ai 0A2
, , ,UBi(v) ifuG Ai nA2
JtX(u.v)= \ -7- ^
mB2(v) if u e Ai n A2
U if u 6 Ai n A2 .
Assume moreover that A is normalized and M.B(V) = 0 for some v. Redundancy writes Vu,
Vv, 7t:^(u,v) < max(l - HA(U)' l^B^^)) ^"'^ ^^^^ '' imp'ies A c A] u A2. Let us assume
that A is normalized. Thus, i) if A c A] n A2 redundancy implies B | n B2 c B, ii) if A c
Aj and A 5^ A] n A2, i = 1 or 2, it implies Bj c B, and iii) if A c Aj u A2 but A ^ A^,
i=l,2, then Bj u B2 £ B provided that 3 u' e A] n A2, |iA("') = ^ ^^^ ^ u" e A] n A2,
|J.A(U") = 1- In this case, the rule is redundant with respect to one of the two rules "if X is
Aj, then Y is Bj", or with respect to one of the two other rules obtained from the two
previous rules via intersection or union operations. But it may happen that either u' or u"
does not exist, i.e., 0 < supyg ATOAT I^A(") "^ I^A(" ) *- ^ ^'^^ instance ; then we only
obtain Vv, [13 j(v) < H B W ^nd [i^^'^'^) < max(l - |aA(u"), HB(V)). '•£• liBiuB2 ^ ^ " 0 -
|JA(U "). HB)- ^ " ' ^^^^ '^'' ^^^y special situation.
When Aj or A2 are fuzzy, the redundancy with respect to the base is not equivalent to
the redundancy with respect to one the four rules mentioned above (those of the base or
45
those generated via intersection or union). Indeed, let us consider the case where B j and B2
are ordinary subsets. For a fact X = UQ, the knowledge base yields Tty =
min(max(n3 1 -ttAi("o))' "i^xCja-g-, 1 - HJI^JUQ)). Suppose that for instance ^ A I ( " O ) -*
IIA^CUQ) > 0. Then clearly ny < max (jig , ^ 3 ^ , 1 - ji^^(uo)) in general, the upper bound
being the result obtained by means of the rule, "the more X is Aj n A2, the more certain Y
is Bj n B2". Thus, as soon as the rules have fuzzy conditions, there does not always exist a
rule which is redundant with respect to a rule base and which can produce the same
conclusion as the rule base for a given input.
In order to make a first step to the redundancy checking of gradual rules, let us examine
conditions under which a rule of the form "the more X is A, the more Y is B" is redundant
with respect to another gradual rule. A necessary and sufficient condition is given by
applying (21), namely
It is easy to check that a—»b>c-^d<=> ( a < b ) o r ( a > b and b > c ^ d) for Godcl
implication. Hence (24) is equivalent to
• HA(") - I^B(^)' ^u>v, implies that B = V (the whole domain of v) since A is normalized.
This is the trivial rule "if X is A, then Y is V" which is redundant with anything ;
• The other condition leads to ^B(V) - I ^ B I ( ^ ) ' ^ ^ ^^'^^ ^^^' >M^Bi(^)' choosing u so that
^iA](u)=l-
Moreover, (25) implies that V v £ core(Bi), HB(^) - ^"Pu 1^A(")' '•^- I^B(^) = 1 ^'"'^'^ ^ •**
normalized. So (25), (26) imply that B 3 Bj. Now, given A, Bj and B 3 B], (25) implies
When B = BJ, (27) leads to Aj 2 A, which is a rather expected result. So, the rule "the
more X is A, the more Y is B" is redundant with respect to the rule "the more X is Aj, the
more Y is Bj" if and only if B a Bj and A satisfies (27).
46
The case of redundancy with respect to a set of two gradual rules is more tricky and left
for further research.
5. Concluding Remarks
This paper presents preliminary results on the coherence of knowledge bases made of
parallel fuzzy certainty rules or parallel fuzzy gradual rules. The study of more general
knowledge bases would require a detailed study of the "transitive chaining" of fuzzy rules,
i.e. the generation of new rules through chaining. However in the particular case of so-called
possibilistic knowledge bases made of weighted classical logical formulas (which
encompasses the case of certainty rules with non-fuzzy conditions and weighted conclusions
(i.e. Jig. = max(p.3'., 1 - a) with B'j non-fuzzy)), the coherence problem corresponds to the
research of "no goods" in a possibilistic Assumption-based Truth Maintenance Systems,
where assumptions are the input literals; see (Dubois, Lang and Prade, 1991a ; Benferhat et
al., 1992). This enables us to extend coherence checking procedures available in the case of
classical logic knowledge bases (e.g. Beauvieux and Dague, 1990 ; Mcseguer, 1991 ;
Loiscau, 1992) to possibilistic knowledge bases.
Let us consider a collection of fuzzy gradual rules of the form "the more X is Aj, the
more Y is Bj" where the Aj's and Bj's have triangular membership functions. Let a; (resp.:
b[) denote the peak and [a'i.a'j] (resp.: [b'i, b'j]) the support of A; (resp. : Bj). Sec Figure 3.
Figure 3
47
We further assume that g'i ^a'i+i anda'j ^a'i+i.h'i S h'j+j and b'j < b'j^] (i.e. a fuzzy
set only overlaps with its immediate neighbours). Then the necessary and sufficient
condition for coherence (18) writes
or equivalently
Vi, Vu. (Bi)^^,(„) n (Bi+i)^^^^^(„) ^ 0.
(«L_|2U|2^.2-^'l+b-,-b'2>0. (A)
This condition should hold for any u in the interval [max(aj,a'2). niin(a2^'i)]- Moreover the
necessary condition for coherence hgt(Ai n A2) S hgt(B j n B2) writes
hgt(Ai n A2) = "^ " ^ ^ ( a ' 2 - a ' l ) < ^^ P^(b'2 - b'l) = hgt(Bi n B2). (B)
a2-«i fc-Pi
Then it can be easily checked that hgt(Ai n A2) 5 hgt(Bj n B2) becomes a sufficient
condition for coherence if _
tt) _ a 2
since then conditions A and B coincide. This expresses the respective proportionality of the
slopes of the increasing and decreasing parts of the Aj's with the ones of the Bj's. A
particular case where this latter condition holds is when cTi = - a j and Pi = -p2. ie- when
the Aj's have symmetrical and identical membership functions, and a similar condition is
satisfied by the Bj's; this situation is in agreement with the idea that Aj models "close to
48
a^", which entails the symmetry and the identity of shapes of the Aj's if we want to have the
same modelling of the notion of closeness everywhere.
References
Ayel M., Rousset M.C. (1990) La Coherence dans les Bases de Connaissances. Cepadues-
Edilions, Toulouse, France.
Beauvieux A., Dague P. (1990) A general consistency (checking and restoring) engine for
knowledge bases. Proc. of the 9th Europ. Conf. on Artificial Intelligence (ECAr90)
(L.C. Aiello, ed.), Stockholm, Sweden, Aug. 6-10, 77-82.
Benferhat S., Dubois D., Lang J., Prade H. (1992) Hypothetical reasoning in possibilistic
logic : basic notions, applications and implementation issues. In : Advances in Fuzzy
Systems : Applications and Theory Vol. I (P.Z. Wang, K.F. Loe, eds.), to appear.
Bourrelly L., Chouraqui E., Portafaix V. (1992) Les topoi pour la validation structurclle
d'une base de connaissances. Joum6e Francophone de la Validation et de la Verification
des Systfemes h Bases de Connaissances, Dourdan, France, Apr. 16.
Dubois D., Lang J., Prade H. (1989) Automated reasoning using possibilistic logic :
semantics, belief revision, and variable certainty weights. Preprints of the 5th Workshop
on Uncertainty in Artificial Intelligence, Windsor, Ont., 81-87. Revised version in IEEE
Trans, on Data and Knowledge Engineering, to appear.
Dubois D., Lang J., Prade H. (1991a) A possibilistic assumption-based truth maintenance
system with uncertain justifications, and its application to belief revision. Proc. of the
ECAI Workshop on Truth-Maintenance Systems (J.P. Martins, M. Reinfranck, eds.),
Stockholm, Aug, 6, 1990, Lecture Notes in Computer Sciences, n° 515, Springer
Verlag, Berlin, 87-106.
Dubois D., Lang J., Prade H. (1991b) Fuzzy sets in approximate reasoning — Part 2 :
Logical approaches. Fuzzy Sets and Systems, 40,203-244.
Dubois D., Martin-Clouaire R., Prade H. (1988) Practical computing in fuzzy logic. In :
Fuzzy Computing (M,M. Gupta, T. Yamakawa, eds.). North-Holland, Amsterdam,
11-34.
Dubois D., Prade H. (1982) Towards the analysis and the synthesis of fuzzy mappings. In :
Fuzzy Sets and Possibility Theory — Recent Developments (R.R. Yager, ed.),
Pergamon Press, New York, 316-326.
Dubois D., Prade H. (1991a) Fuzzy sets in approximate reasoning — Part 1 : Inference with
possibility distributions. Fuzzy Sets and Systems, 40, 143-202.
Dubois D., Prade H. (1991b) A note on the validation of possibilistic knowledge bases.
BUSEFAL (IRIT, Univ. P. Sabatier, Toulouse, France), 48, 114-116.
This page intentionally blank
49
Dubois D., Prade H. (1992) Gradual inference rules in approximate reasoning. Information
Sciences, 61(1,2), 103-122.
Hall L.O., Friedman M., Kandel A. (1988) On the validation and testing of fuzzy expert
systems. IEEE Trans, on Systems, Man and Cybernetics, 18, 1023-1028.
Kinki61cl6 D. (1992) Detection des incoherences potentielles dans les bases de connaissances
floues : vers un modele conceptuel. Presented at "leres Rencontres Nationales des Jeunes
Chercheurs en Intelligence Artificiclle", Rcnnes, France, Sept. 7-9.
Larsen H.L., Nonfjall H. (1989) Modeling in the design of a KBS validation system. Proc.
of the 3rd Inter. Fuzzy Systems Assoc. (IPSA) Congress, Seattle, Aug. 6-11, 341-344.
Larsen H.L., Nonfjall H. (1991) Modeling in the design of a KBS validation system. Int. J.
of Intelligent Systems, 6, 759-775 ; Erratum : 7, 1992, p. 391.
Loiseau S. (1992) Refinement of knowledge bases based on consistency. Proc. of the 10th
Europ. Conf. on Artificial Intelligence (ECAr92) (B. Neumann, ed.), Vienna, Austria,
Aug. 3-7, 845-849.
Meseguer P. (1991) Verification of multi-level rule-based expert systems. Proc. of the 9th
National Conf. on Artificial Intelligence (AAAI'91), July 14-19, 323-328.
Nguyen T.A., Perkins W.A., Laffery T.J., Pecora D. (1985) Checking an expert systems
knowledge base for consistency and completeness. Proc. of the 9th Inter. Joint Conf. on
Aruficial Intelligence (IJCAI'SS), Los Angeles, CA, 375-378.
Nguyen T.A., Perkins W.A., Laffery T.J., Pecora D. (1987) Knowledge base verification.
AI Magazine, 8(2), 69-75.
Turksen I.B., Wang Q. (1992) Consistency of fuzzy expert systems with interval-valued
fuzzy sets. Proc. of the Abstracts and Summaries of the 1st Inter. Conf. on Fuzzy
Theory and Technology (FT&T'92) (P.P. Wang, ed.), 225-230.
Yager R.R., Larsen H.L. (1991) On discovering potential inconsistencies in validating
uncertain knowledge bases by reflecting on the input. IEEE Trans, on Systems, Man and
Cybernetics, 21,790-801.
Zadeh L.A. (1979) A theory of approximate reasoning. In : Machine Intelligence,
Vol. 9 (J.E. Hayes, D. Michie. L.I. Mikulich, eds.), Elsevier, New York, 149-194.
CHAPTER 3
Sozo Yamamoto
Abstract
Problems and their solutions in representing fuzzy sets and logic in software sys-
tems are discussed in this article.
Fuzzy set theory is getting to be widely used as a tool for managing uncertainty
in complicated systems. Interactions of fuzzy set theory and information processing
is called 'fuzzy information processing', where software representation of fuzzy sets
and logic is an important subject. Fuzzy information processing is an important
area of research but is not fully investigated. This is because it has some problems.
For one thing, a fuzzy set can be represented with various kinds of complicated
d a t a structures. Another problem is that there are effectively infinite number of
operations defined on fuzzy sets.
Some fuzzy logic based systems have been proposed, like fuzzifications of Pro-
log, fuzzy control shells, and specially designed languages for fuzzy set processing.
But they are not fully acceptable as a uniform platform of fuzzy information pro-
cessing. T h e trade-off of flexibility, convenience and performance remains.
Object-orientation can be a key to solve these problems. Because object-
orientation has the ability of d a t a abstraction and information hiding, it is suitable
for fuzzy information processing which needs manipulation on complicated data
structures. An object-oriented fuzzy set manipulation system named F O P S was
developed on such ideas. Two basic classes for fuzzy sets, ArrayedFuzzySet and
PairedFuzzySet, are provided and they can be used interchangeably. W i t h its
support for fuzzy logic and development environment, F O P S can serve as a good
starting point of fuzzy logic based software. Outline of the system and internal
d a t a structures are discussed in this article.
51
S.G. Tzafeslas and A.N. Venelsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 51-68.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
52
1. Introduction
Great progresses have been made in the application of fuzzy set theory and fuzzy
logic. Most remarkable area of application is 'fuzzy control', where fuzzy logic
was first applied to plant control systems and its use is expanding to consumer
products. Most of fuzzy control systems uses fuzzy inference with max-min or
max-product composition, similar to the algorithm that first used by Mamdani
in 1970s. Some algorithms are developed to refine fuzzy controls systems but the
main part of algorithm stays the same.
Triggered by the success of fuzzy control systems, other ways of applying fuzzy
set theory are also investigated. They are usually referred t o as 'fuzzy expert sys-
t e m s ' , and their purpose are to combine the idea of fuzzy theory with AI based
approach toward knowledge processing. These approaches can be more generally
viewed as 'fuzzy information processing', t h a t is to bring fuzzy idea into informa-
tion processing systems.
These systems have variety of algorithm and each of t h e m are built on some
software basis like Lisp or C language. But if a system is based on fuzzy set theory
or fuzzy logic, the system must have some part t h a t is implementing fuzzy set
representation and fuzzy logic. So there will be a need for a concrete basis for fuzzy
information processing, which enables users to treat fuzzy sets or fuzzy concepts
in the same manner with other components without implementing software for
processing fuzzy set. Notations of floating point numbers like '1.23E4' or '5.6e-
7' are understood by most of common computer languages. Likewise, if fuzzy
sets can be treated conveniently in programming languages, it will help software
development and researches related to fuzzy logic.
In this article, methods of representing fuzzy sets in software are discussed and
a brief survey is made on software systems for fuzzy knowledge processing and fuzzy
set representation. After t h a t , F O P S , which is an object-oriented software for fuzzy
set manipulation, is introduced to show the effectiveness of object-orientation for
fuzzy set processing and fuzzy information processing.
A^ I fiA{u)/u (1)
Fuzzy sets in functions(Fig. l-(a)) One simple way is to describe the fuzzy
set in a function which express the membership function fi^. For example, a
membership function for bell-shaped fuzzy set can be expressed as
1 + p{u — ay
This method is achieved with less memory than other ways of representation, but
is less general because the result of operation can not necessarily be represented
in the same manner.
Pointwise Definition(Fig. l-(b)) Another approach is to describe a fuzzy set
with a polygonal line, giving ea<:h bending point with an element in universe of
discourse and corresponding grade of membership. A fuzzy set with triangular
membership function can be described a^
{0/0-0/2.5-1/4-0/5.5-0/10} (3)
fuzzy sets. This kind of fuzzy set is called 'level-2' fuzzy set and is difficult to be
described in ottier ways.
In paired representation, fuzzy sets are described like,
or more generally,
n
A = Ylnilui. (6)
This method is flexible enough to permit almost all kinds of fuzzy sets.
And a fuzzy set can be represented with a family of these a level sets as
A / ccA, (8)
^ €>€(0,ll
(0,1]
and if the fuzzy set is convex, each a level set becomes an interval in U. So a fuzzy
set are described with a set of numeric intervals.
With this representation, fuzzy number operations can b e handled in a simple
way. For example, a two operand operation * between two fuzzy sets A and B can
be executed as
A*B=I {A-kB)c= I A^*Ba (9)
•^ae(o,il -'oe(o,i]
and the right hand side can be computed as operation on intervals.
3.1 S o f t w a r e for M a n a g e m e n t of U n c e r t a i n t y
T h e first category is software systems for management of uncertainty. One of the
earliest works is PRUF[1] which was proposed by Zadeh. T h e main aim of P R U F
was to represent the uncertainty in natural language in a formal manner. P R U F
has its hash on possibility distribution rather than t r u t h value. Uncertainty or
vagueness in natural language is represented in terms of possibility distributions
and fuzzy propositions. T h e main interest in this system was the representation
of natural language on computer systems, using fuzzy logic as a tool for managing
uncertainty.
55
<1
• • • I•
• • • ; : ;•
• • • ; ; I ! !•
• ! ! ! : ! : • •
T '1'l»»»<
U
(c) Vector Representation (d) Paired Representation
• •
5.1 R e p r e s e n t a t i o n of F u z z y S e t s
I n t e r n a l R e p r e s e n t a t i o n s Two basic classes. ArrayedFuzzySet and PairedFuzzy-
Set, and other derivations are provided to represent fuzzy sets in F O P S .
0 0.3 0.7
0.5 0 1
1 2
a'l 5
block
fuzzy set
expression empt'i
\ • grade — ^ v / y ~ ^ elements -<
O
uraverses
^-^ ON: - ^
fr universe
«£
elements ^ element
W^-^ element
fi£ a -H)
O t h e r T y p e s of Fuzzy Set As stated before, the two kinds of fuzzy set repre-
sentation are ready for use in the system, but other ways of representation can be
easily added on to the system.
Newly created cla^ss should suffice following conditions;
1. Is one of subclasses of ArithmeticValue,
2. Operations are executable between two objects in that class,
3. Have coerce: method to convert into fuzzy set representations with higher
generality,
4. Objects with lower generality can be converted into instances of the class.
For example, to define a fuzzy set whose membership function is given by a func-
tion, it can be represented as a object that consist of universes of discourses and
a block for the function.
Some special kinds of fuzzy sets can be represented by assigning some specific
class of objects to grades, because operations on grade values are interpreted by
each objects representing grades. For example, representation and operations of
L-fuzzy sets can be accomplished by defining a class for partially ordered objects
(like tuples) and assigning them to grade values of fuzzy sets. Also, assigning
interval values to grades makes the fuzzy sets can be applicable to fuzzy interval
logic or inverse problems of fuzzy relations.
- Time (msec) -
10.00-! >
;
yii'"'
; 5^
2 4 8 16 32 64 126 2S6
Nuntier of elBments '
© AFS A Pf S
7. Conclusion
Software for fuzzy information processing is a important but not sufficiently in-
vestigated area of research. Its need is indicated with the fact that each system
that is based on fuzzy set theory is implementing diiferent set of codes for fuzzy
set processing. Some systems have been developed, but most of fuzzy set based
software are still built on conventional computer languages.
It is a natural way to treat fuzzy sets as d a t a types, as with other compo-
nents in software. Fuzzy set manipulation system F O P S was aimed to utilize the
merits of object-oriented paradigm for fuzzy set manipulation. By integrating
two different types of fuzzy set representation and treating fuzzy sets as numbers,
treatment of fuzzy sets in programming language became much clearer. Two fuzzy
set representations, ArrayedFuzzySet and PairedFuzzySet, are implemented and
interchangeably used in F O P S . These two have different characteristics in memory
usage and computation load, so users can choose appropriate d a t a structure for a
given application.
It is expected that such fuzzy set representations are standardized into some
form and get to be included in many computer languages, to promote researches
in fuzzy information processing or fuzzy knowledge processing.
This work was performed in part at the Laboratory for International Fuzzy Engineer-
ing Research(LIFE). The author wishes to thank Dr. Seiji Yasunobu and Mr. Yoshifumi
Inoue for their collaboration at LIFE and Professor Umano of Osaka University for
helpful discussions.
References
[1] L. A. Zadeh: " P R U F - A Meaning Representation Language for Natural
Languages", Intl. J. of Man-Machine Studies, 10, 395-460(1978).
[2] J. M. Adamo: "L.P.L. A Fuzzy Programming Language: 1. Syntactic As-
pects", Fuzzy Sets and Systems, Vol.3, 151 - 179(1980).
[3] J. M. Adamo: "L.P.L. A Fuzzy Programming Language: 2. Semantic Aspects
", Fuzzy Sets and Systems, Vol.3, 261 - 289(1980).
[4] T.P. Martin, J . F . Baldwin, B.W. Pilsworth: "The Implementation of
F P R O L O G - A Fuzzy Prolog Interpreter", Fuzzy Sets and Systems, Vol.
23, 1 1 9 - 129(1987).
[5] M. Umano: "Fuzzy-Set Prolog", Preprints of 2nd IFSA Congress, 750-
753(1987).
[6] W. Siler: " F L O P S : A Fuzzy Expert System Shell", Preprints of Second IFSA
Congress, 848-850(1987).
68
(*)
Department of Electrical Engineering
University of Toronto, Toronto, Ontario, Canada M1551A4
(**) Intelligent Robotics and Control Unit
National Technical University of Athens
Zografou Campus, Zografou 15773, Athens, Greece
1. INTRODUCTION
This chapter deals with fuzzy dynamic systems paying attention to three important issues,
namely analysis, control and identification. Analysis is concerned with the study of input-
output descriptions which are based on fuzzy relational equations. One of the models that
played a central role in the analysis of classical (crisp) dynamic systems is the state-space
model. This model can be extended to cover the case of fuzzy dynamic systems, and is
studied in the present chapter. Control is concerned with the problem of determining a
control sequence which brings the system state, in one or more time steps, from the present
fuzzy value Xj^ to a desired final fuzzy value Xp This problem is equivalent ot solving the
relevant equation Xj-= UjpXjpR with respect to l/^. The fuzzy system identification
problem solved here consists in estimating the system fuzzy relational matrix R from a
given set of fuzzy input-output pairs (X,-,y,), i= 1,2,...,JV. The no-noise case is considered,
since the case where noise is present is stiU a challenging open problem. An algorithm is
provided for determining good upper and lower bounds that confine R. Also, the state
prediction problem of fuzzy state space models is considered. The chapter closes with a
review of three practical fuzzy controllers which have been applied with success to several
industrial systems. These controllers are: (i) the fuzzy self-organized controller (SOC), the
fuzzy PID supervisor, and the fuzzy PID incremental controller.
To understand what actually inspired and pushed L-Zadeh to initiate fuzzy system theory
through his pioneering article of 1965 [1] one must resort to his earlier studies. In one of
them Zadeh introduces the concept of abstract object in order to define what is a system.
In his own words «... A system is designed as a collection of ordered pairs, representing
inputs and outputs, and defining abstract objects. This way, one may treat various systems
as a whole by their causal relationships only, thus dealing with a large variety of problems
in a unified way». The next step is of cource to locate an input-output relation with the
69
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision ami Control Systems, 69-96.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
70
property that it generates input-output pairs characterizing the system. From that point the
problem falls within the framework of the classical systemic approach and leads to the
development of concepts like the transfer function, frequency response, state space model,
etc. Tliroughout the above arguments it was tacitly assumed that the inputs and outputs of
the system can be measured quantitatively (nunierically). This is true for technological
systems which do not involve the "hum.an" as part of them. However this ceases to be true
when the "human" is entered as a component of the system control loop. This is because
the human accepts stimulations (inputs) and replies with responses (outputs) that are
expressed better by words (e.g. adjectives) than by numbers.
The above are best illustrated by the following example, where the human acts as
a controller that regulates the proportion of hot and cold water in order to achieve a
desjred warm water. For Convenience it is assumed that the flow of water remains constant
and only the relative proportion of hot and cold water is changed through a valve. The
control rules of the example are:
IF "water very hot" THEN 'Valve pretty right"
IF "water a little hot" THEN "valve a little right"
IF "water warm" THEN "do nothing"
IF "water a little cold" THEN 'Valve a little left"
IF "water very cold" THEN "valve pretty left"
In input - output pair form the above rules are written as:
(very hot, pretty right)
(a little hot, a little right)
(warm, do nothing)
(a little cold, a little left)
(very hot, pretty left)
It is clear that in this case, one cannot find a suitable function that produces the input -
output pairs, since these pairs are no more numeric but linguistic (here adjectives).
According to Zadeh [2] the linguistic description of a system is much more effective, but
less specific, than the mathematical (numerical) description.
The modern systems theory has been structured and developed around the concept
of state space. A system is described by three quantities: input, output and state, (i) If the
present state and input determine exactly the system output, then the system is called
deterministic, (ii) If the next state and the output (given the present state and input)
cannot be described exactly but only through their probabilistic and statistical properties,
then the system is called stochastic. (Hi) If for the next state and output (given the present
state and input) the only that can be said is that they are elements of a set with fuzzy
bounds then the system is called fuzzy.
A system acts as a mathematical operator, i.e. as a mathematical function that maps
inputs to outputs. In the triad (input, system, output) the system participates with its
mathematical model. Thus, in principle, knowing any two of them one can find the third.
Fuzzy systems and neural systems are mainly used in estimation problems. They possess
an important common advantage over the classical statistical estimation mechanisms and
the adaptive control techniques, respectively. Th^y are model free estimators. That is, fuzzy
and neural systems can find how an output depends on the input without the need to use
a mathematical model. Instead, they use an alternative approach, namely they learn from
examples. If X is the input space, Y is the output space, then the pair (Xj, jy) con.stitutes
71
an element of the Cartesian product XxY. In neural systems the sample (Xj, 7,) is a pair of
numbers (e.g. (1, 5)), whereas in fu2zy systems this sample is a pair of fu2zy sets (e.g. (fast
motion, more green) or (hot water, valve shiightly left)).
y = Xo«
where "o" is Zadeh 's compositional rule of inference (max-min composition) [see Ch. 1,
Section 9.4].
Using the membership functions n^x), /iy(y) and jU^(jf,y) of X, Y, R respectively
the above relational equation takes the form
My(y) = max{mm{n xix),tiif{x,y))
A more general relational equation results if the operator max is replaced by the
operator sup and the operator min by a T-norm (see Ch.l, Sec. 5}, namely
^ y ( y ) = sup{^^(Ar)T/ij?(jf,7)}
ii) Measurement Problem: Given Vand R find X, i.e. Knowing the output and the system,
find the input.
To precisely formulate the above two problems a new operator "«" is required
which plays the role of the Inverse T-norm used in Zadeh 's composition rule. This need
can be understood by the following argument. From the relational equation Y=XoR one
can see that the operator "o" is similar to the operator "•" of the matrbc multiplication.
This is so since the operator "o" produces Y (a vector) from the product of X (a vector)
and R (a matrix). However, it possesses the fundamental difference that it does not allow
the definition of the inverse, i.e. if Y=XoR then it is not true that X= YoR'^. Therefore
it is absolutely necessary the introduction of an operator"«" which is the inverse of the T-
norm used in the operator "o".
Definition 3.1
An operator (function) «:[0,l]x[0,l]-»[0,l] is defined to possess the properties:
i) X " max(y,z) > x«7, x^z
ii) X « (xTy) < y
iii) X « (x^y) > y
For Zadeh 's T-norm i.e. the operator "min" the corresponing "-operator is
Operator «: a«b=l or b if asb or a>b
The operator « can be applied both between two fuzzy sets X and Y, and between a fuzzy
set y and a fuzzy relation R. In the first case:
As we have already seen, fuzzy systems are described by fuzzy relational equations. Our
purporse here is to introduce the concept of fuzzy state equations on the basis of fuzzy
relational equations. We shall describe fuzzy systems of first and higher order, and discuss
the relation of relational equations with the classical difference equations.
^k*p = ^k*p°^
where
o: Zadeh 's compositional rule of inference
t/j^: fuzzy control variable at time instant k
^k+p' ^k+p-i<-' ^k- fuzzy state variable at times k+p, k+p-l,...,k
R, ^ fuzzy relations that characterize the dynamics of the system and
are defined as: p
R : UxXx...xX •* [0,1]
S : XxY - [0,1]
The first equation is the state equation that combines the fuzzy state at time Jt+p with the
fuzzy states at times k+p-l,...,k and the fuzzy control at time k. The second equation is the
output at time k+p with the fuzzy state at the same time instant.
The above model is of order p and can be considered as a generalization of the
difference equation model of order p. In the following discussion it will be assumed that
X= Y (the outputs coincide with the states) and that the relation 5 is diagonal. Thus
without loss of generality we can work only with the first equation.
where Uf^ Xj^ and R have the usual meaning. In terms of membership functions the above
equation takes the form
The pth-order fuzzy state space model can be converted to a set of Ist-order equations by
setting
R' - ^k.\'^k.2'-'^k.p-l'R
^ k*p = ^A+p
s' = s
where R' is the fuzzy relation R':UxXx...X-^[0,l] and S' the fuzzy relation 5':Xxy-*[0,l].
The resulting Ist-order fuzzy model is:
Thus one can work with a fuzzy state-space of 1st order without any loss of generality.
where a, b, Xj^, Xj^^j and Uj^ take real values. These values can be regarded as fuzzy
numbers with the following membership functions:
M[/ (") = 1 if U = Uj^; =0 otherwise
i"x (^) ~ ^ if X = Xj^; =0 otherwise
Hjf{x,y,u) = 1 if y = ax+bu; =0 otherwise
Thus if we consider the quantities t/^, X^, X^ +j and R to be fuzzy ones, one can form the
relational equation:
75
MXk^,(y) = max|min|Ai^(u),;i;^i^(x),^^(Y,y,u))|
Introducing the respective substitutions we find that here
MA" (y) ^ ^ if y = aXj^ + bU]^ ; =0 otherwise
The above veriiy (at least in this example) that the relational equations constitute an
extension of the usual concept of difference equations when the variables involved are fuzzy
sets.
Let us now consider a first-order fuzzy system Xi^= Uf^oXfOR which has the block diagram
shown in Fig. 1. For this system we shall discuss here three particular fuzzy control
problems, namely:
i) One step control without or with constraints.
ii) Multistep control.
iii) Construction of the controller relational equation.
U\ = E,^X,^{X,OR)«X,
where t/j^'is the largest from all possible controls (if they exist) that satisfy the equation
Xj.= 140X1,0/?.
If a solution does not exist, i.e. if the abvoe solution gives to some point of l ^ '
negative membership value, then one can use some distance criterion and try to find a
control Uj^ that brings the quantity L/^oXj^o/? as near as possible (with respect to this
criterion) to the desired final state X/.
In this case the problem is to find a Uj^' such that the fuzzy distance d{Uj^oX^^oR,
Xf) is minimized, i.e.
where e.g. d may be Hamming fuzzy distance (See Ch. 1, Sec. 6.3).
A measure of "how good" is the control U^.' is provided by the e-reachability
concept which is defined as follows.
Definition 5.1
Let e{Uf^") be the quantity
e{U'\) = diU'\'>x^oR,Xf)/supd{U^oX^''R,Xf)
where d is the Hamming fuzzy distance.
Then X^is said to be EQ-reachable from Xj^, if e{Ui^")<Eg. Clearly, if e^=0, then
the state Xf is exactly reachable from Xj^.
X j = t/joXjoft = Ui'>(U^oX^oR)oR
= t;,ot;„=x„oj?2
Xf=Xp= Up,,oUp_2^...oUioU^oX^oRP
Xf= Uo(XgoRP)
U' = {X„oRP)o.Xf
X, U,
= X^^X^^RoG
The control problem for the closed-loop system is to find the fuz2y relation G of
the controller such that one can apply to it (i) single step and (ii) multi-step control.
Clearly, the closed-loop equation can be written as Xj^^^ = CoG which has exactly
the same form as the equation Xf= Uj^oE^ of the single step control. Thus, the one step
solution is readily applicable here too.
The identification problem of fuzzy systems consists in estimating the relation R that
describes the system, from fuzzy input- output pairs (X,, V,), / = 1,2,..., N. Here this problem
will be treated in the no-noise case. The case where some kind of noise is present needs
the introduction of two additional concepts, namely the probabilistic sets and the fuzzy
clustering, and will not be considered here.
/^x = >[^x(^/)].
• •
A'' y =' [^• r(y,)j.
' '
P R• = [f^Ri^vYi)]'' . , ' ""'
j=i,z,...,m
Fy(7j) = max|min(/j;y(.v,),^;j(;yy,7;))|
Then the problem of estimating R is formally stated as follows:
Given: The fuzzy vectors \}i)^x>i\, lJUy(y,)].
Find: The fuzzy matrix [}iji{Xj, y)].
The information contained in the input - output pairs (X,-, Vj), i= \,2,...,N is not sufficient
for the unique determination of the relation /?, but only for the determination of a lower
and an upper bound {R, R respectively) such that
These bounds can be found by using the operators « and « of Sanchez [4].
Theorem 6.1 _
An upper bound R of R is given by the formula:
f^ni^i'yj) = Mx(^i)"i"y(y,)
1. MA-C-*',) ^ t^yiyj)
79
Theorem 6.2
A lower bound 5 of R is given by
The upper bound /? satisfies the relational equation Yj=XjOR, which is not satisfied by
the lower bound, i.e. XpR^Yj.
An algorithm for the estimation of the fuzzy relation R (i.e. an upper and a lower
bound for it) is the following:
Initialization: Put
UpperR*-\, LowerR*0
Counter fc<-l
Number of Iterations N
Step 1: Read the fuzzy input - output sets X, and Yj
Step 2: Compute X;« Y, and X;« Y,
Step 3: Update the UpperR and LowerR as
UpperR*- UpperRriXj« Yj)
LowerR'^LowerRu(Xj^ Yj)
Step 4: If k<N then increase the counter by \: k+ l*-i and go to step 1. Otherwise
terminate the algorithm.
Clearly, at the end of this algorithm R will be confined between LowerR and
UpperR. Taking as R one of the above bounds UpperR or LowerR, respectively, we make
an error. A measure of this error is the distance between R and its estimator UpperR or
LowerR which is defined as
When r= 1 this distance reduces to the well known Hamming distance (Ch. 1, Sec. 6.3).
^k*m
m
Suppose that the fuzzy state at time A^ has yi^ =[0,...,0,1,0,...,0] where the 1 appears at
the jth position. Then a normalized measure of the uncertainty of the fuzzy state X^^. j is
(Ch. 1, Def. 2.4):
Using n^ =[0,...,0,1,0,...,0] , the fuzzy relational equation X^^j = XjjOG, and the
quantity rj, = max/ij^(xy^;), one can compute IXj^ + il and find
l^jt.il =[E/^G(^i.^y)]/'-o
Thus the normalized measure of fuzziness of X^^.j is greater than that of Xjj. Similarly the
fuzziness measure for X^^.2 is greater than that of Xj^^j, and so on. Thus after a certain
number of time instants the prediction is no more useful.
It is desired to find an estimate of R on the basis of the following fuzzy input - output
pairs
81
UpperR =
Hamming (L/pper/?,/?) = 9
Hanunlng(UpperR,R)
Hamming(R,LowerR)
20 4-
Pictorially the sequences of the upper and lower bounds are shown in Fig. 4 where
their Hamming distances from R are depicted.
Example 6.2: Fuzzy state prediction
Consider the system X;t+i=^ik°<^*ith'^Jt = n 0-2 0.1 0 0 0] and
1 0.8 0.6 0.5 0.3 0.1
0.9 1 0.8 0.7 0.4 0.2
0.7 0.9 1 0.8 0.5 0.2
G = 0.7
0.6 0.7 0.9 1 0.8
0.4 0.5 0.7 0.8 1 0.9
0.2 0.3 0.5 0.7 0.8 1
Carrying out the prediction calculations one finds the following sequence of future fuzzy
states:
Time horizon
1 ^k*l ° [[1 0.2 0.1 0 0 0]
2 ^k*2 = [[1 0.8 0.6 0.7 0.5 0.5]
3 ^k.3 = [[1 0.8 0.8 0.8 0.7 0.7]
4 ^k*4 ° [[1 0.8 0.8 0.8 0.8 0.7]
5 ^k*5 " [[1 0.8 0.8 0.8 0.8 0.8]
6 ^k*6 = [[1 0.8 0.8 0.8 0.8 0.8]
Fuzzy state A
Sets X 1 i
Reference superset
Fig. 5 Pictorial representation of the fuzzy states Xj^+i- ^Jt+2> ^Jt+3' ^Jt+4' -^A+5' ^"'^
^k\6
86
These fuzzy states are represented pictorially in Fig. 5.
The size (normalized measure of fuzziness) versus prediction time horizon of the
state X^. {k = k, k+ 1, ..., k+6) has the form shown in Fig. 6.
"Size" of .
fuzzy sets X ' '
6 •
4 5 6
Prediction time horizon
Fig. 6 Size of predicted fuzzy states X versus prediction time horizon
e(kT)
incremental r ( k T )
Hertorinance
meausure ., _ , Model
Ruxe mcxii-
fier
Contro]
GE
Set i(kT
p.oint<i>r Output
Process -r*
TICK^ u
Gi:
C(kT-mT)=F{c(kT-mT)} (change-in-error)
U(kT-mT) = F{u(kT-mT)} (uncorrected control)
U{kT-mT) = F{u{kT-mT)+iikT)} (corrected control)
where F{-} represents the fuzzification process, mT is the time interval between successive
control corrections, and r(kT) is the control correction given by (3).
b) Determine the uncorrected and corrected (modified) relation matrices R' and
R" from the corresponding implications
E{kT-mT) -* C{kT-mT) -* U(kT-mT)
E{kT-mT) -* C(kT-mT) •* U{kT-mT)
That is
R\kT) = E{kT-mT)>^C{kT-mT)y^U{kT-mT)
R'(kT) = E(kT-mT)xC(kT-mT)xU'(kT-mT)
c) Express the control relation matrix modification linguistically, i.e.
89
/=! 1=1
where Vy is the geometric center of the component lif^{v) in the membership function
m
1=1
90
(a)
1
e K i. ^^©^16—P
'uzzi- De
fier fuzzi-
_e__ Fier
(b)
Fig. 8 (a) Supervised PID controller, (b) The architecture of the supervisor
Other defuzzification rules that en be used are the "center-of-gravity" rule and the
"equal-surface" rule [6].
The performnace index is in line with the human judgment, and uses the rise time
T^ the overshoot h and the ITAE criterion to qualify the overall performance of the
controlled system as bad, satisfactory, good and excellent. The selected fuzzy subsets for
T^ h, ITAE, and the overall performance index (P/) are shown in Fig. 9 from which the
following linguistic rules are obtained:
0 10 15 20 25 2 3 4 5 P.I.
ITAE {V.sec^} Performance Index
Fig. 9 Fuzzy subsets for h, ITAE, T^ and performance index
Rl: Jf h is big OR ITAE is big
THEN the performance is bad
R2: If h is medium AND ITAE is medium
THEN the performance is satisfactory
R3: If /j is medium AND ITAE is small AND T, is big
THEN the performance is satisfactory
R4: If /i is small AND /TA£ is medium AND T^ is NOT big
THEN the performance is good
R5: If h is medium AND /TAH is smaW AND T^ is NOT big
THEN the performance is good
R6: If h is small AND ITAE is medium AND T^ is NOT b/g
THEN the performance is good
R7: If b is sma/i AND ITAE is smaii
THEN the performance is excellent
As usual, the performance index is computed by the fuzzy cartesian product of the
above linguistic rules and the weighted sum defuzzifier rule.
Slx-
H. -M -S -0 +0 + S +M +1.
+ u iO +S -M -L -L -u -1. -L
•fM +s -0 -S -M -M -M -u -L
+S +M +s -0 -S -S -S -M -U
+ 0 4M •IM +s +0 -0 -S -M -M
Jf*o
-0 -fM +M +s +0 -0 -s -M -M
-s +L >M +s •t-S +s +0 -5 -s
-M +L +L +M 4U +M +s +0 +0
-L +L +L +L +u +L -fM +0 +0
the number of the projected parts and reduces the fuzziness of the control policy (Fig. 11).
The analysis of the projections of the boundaries can be regarded as a study of the variance
of the neutral point and the amplitude of the fuzzy function /u^. MacVicar-Whelan has
asserted that using this analysis one can predict the resolution of the projections of the
boundaries. An alternative solution would be to start with a small resolution and slowly
increase it until we get the desired performance.
This matrix was not used by others, probably because it implements the fuzzy
controller in a direct way, and so for each particular case with special requirements it
should be varied. Another drawback is the small number of quantization levels used.
However, we believe that the logic and structure of this matrix are valuable, and thus we
used it for improving existing analog and digital PID tuning procedures as described in the
following.
The rules used for shlightly changing the controller parameter settings during
transient are empirical and are involved in a fuzzy 14x14 control matrix. Th eproblems
which must be faced refer to stability, computation time, cost, quantization, and
generality of the approach.
The main issues of our control rules are as follows:
a) since the integral term is responsible for the overshoot, shlightly decreasing it
at the moment when the system response exceeds the value 1, one can reduce considerably
the overshoot. On the other hand, a small increase of the integral term during the rising
of the response, leads to a 10-20% improvement of the rise time.
-6 -5 1 ""*1-^ -1
! • '
-0 +0
1-1 +2 + 3 + 4 +5 +6
-6 -6 -6 -5 -5 -4 -4 -3 -3 -2 -2 -1 -1 0 0
-5 -6 -5 -5 -4 -4 -3 -3 -2 -2 -1 -1 0 0 0
-4 -5 -5 -4 -4 -3 -3 -2 -2 -1 -1 0 0 0 +1
-3 -5 -4 -4 -3 -3 -2 -2 -1 -1 0 0 0 +1 +1
-2 -4 -4 -3 -3 -2 -2 -1 -1 0 0 0 +1 +1 +2
-1 -4 -3 -3 -2 -2 -1 -1 0 0 0 +1 +1 +2 +2
-0 -3 -3 -2 -2 -1 -1 0 0 0 +1 +1 +2 +2 +3
+0 -3 -2 -2 -1 -1 0 0 0 + 1 + 1 +2 +2 +3 +3
+1 -2 -2 -1 -I 0 0 0 +1 + 1 +2 +2 +3 +3 +4
+2 -2 -1 -1 0 0 0 + 1 + 1 +2 + 2 + 3 + 3 +4 +4
+3 -1 -1 0 0 0 + 1 + 1 +2 +2 +3 +3 +4 +4 +5
+4 -1 0 0 0 + 1 + 1 +2 + 2 + 3 + 3 +4 + 4 +5 +5
+5 0 0 0 +1 + 1 +2 +2 +3 + 3 + 4 +4 + 5 +5 +6
+6 0 0 +1 +1 +2 +2 +3 + 3 + 4 +4 + 5 + 5 +6 +6
b) Since the derivative term is responsible for the flatness of the step response, a
small increase of it during rising and in the steady state eliminates the small oscillations
that usually occur.
c) Increasing the proportional term leads to reduced rise time and increased
oscillations. Thus finally this term should be decreased to avoid oscillatory behavior.
All these issues are experential observations of human controllers. Our problem was
to accomodate them into a matrbc and find the criteria of actually carrying out these
changes. The amount of increase / decrease of the proportional, integral and derivative
terms is expressed by three coefficients k^ k^ and k^
The fuzzy PID control matrix is shown in Fig. 12. The basis for the control rules is
the error E=X^-X {X^ is the desired and X the actual system output) and its derivative
AE=A(Xg-X). The sampling of E and AE should be made at small time interavals of the
order 0.1-0.2 sec.
This table involves 14 quantization levels for both E and AE. This ensures a closer
follow-up of the unit step response.
The linguistic code is
+ •» positive change - — negative change
6 •• extra large 4 •• big 1 — extra small
5 •• large 3 •» medium 0 •» zero (null)
2 •• small
The changes of the three terms, which showed good success, are given by the
expressions:
P = P*CV{E^E} X ^1 (Proportionar)
I = I + CV{EAE} X k2 (Integral) (13)
D = D + CV{EAE} X k^ (Derivative)
The parameters k^, k2 and k^ play an important role in the whole procedure, since
they determine the range of variation of each term. For example, if some tuning method
ensures very small rise time and large overshoot, then the integral term should have a large
range of variation, whereas the other terms can remain unchanged. This range of variation
should be matched with the stability interval in order to guarrantee stability. Thus the
values of Jtj, k2 and k-^ are determined from both the stability analysis and the particular
characteristics of the closed-loop response. In general the parameters k^ /cjand k-^ provide
large flexibility and can be used in conjunction with all available PID tuning algorithms. We
have also used the fuzzy control matrix of Fig. 13. The only difference with the matrix of
Fig. 12 is that here the Neperian logarithm of \CV{E, AE] \ is used. This ensures greater
stability around the set point and smoother transition from one value to another. The
results obtained using this table are also satisfactory, but the values of /cj, Aj and k^ are
found to be larger.
The stability can be tested by various techniques. In our studies we used Routh 's
table in the continuous-time case and the corresponding table, obtained through the
bilinear transformation z=(w+l)/(w-l) in the discrete-time case. The most serious
problem faced in our experiments was that of quantization. The right quantization of E and
AE ensures a good evaluation of the controller performance. We found the logarithmic
quantization to be most successful. We have used the following quantization levels:
95
^E{X„-X)
-2 +0 +4 +6
^ • ' - -^ - - " " - ^'
-6 -1.79 -1.79 -1.60 -1.60 -1.38 -1.38 -1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00
-5 -1.79 -1.60 -1.60 -1.38 -1.38 -1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00
-4 -1.60 -1.60 -1.38 -1.38 -1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00
-3 -1.60 -1.38 -1.38 -1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00
-2 -1.38 -138-1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0,00 +0.69
-1 -1.38 -1.10 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 +0.69 +0.69
-0 -1.10 -1.10 -0,69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 + 0.69 + 0.69 + 1.10
+0 -1.10 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 + 0.69 0.69 + 1.10 + 1.10
+1 -0.69 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 +0.69 +0.69 + 1.10 + 1.10 + 138
+2 -0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 + 0.69+ 0.69 + 1.10 + 1.10 + 138 + 138
+3 0.00 0.00 0.00 0.00 0.00 0.00 0.00 +0.69 +0.69 + 1.10 + 1.10 + 138 + 1.38 + 1.60
+4 0.00 0.00 0.00 0.00 0.00 0.00 +0.69 + 0.69 + 1.10+ 1.10 + 1.38 + 1.38 + 1.60 + 1.60
+5 0.00 0.00 0.00 0.00 0.00 +0.69 +0.69 + 1.10 + 1.10+ 1.38 + 1.38 + 1.60 + 1.60 + 1.79
+6 0.00 0.00 0.00 0.00 +0.69 +0.69 + 1.10 + 1.10 + 138 + 1.38 + 1.60 + 1.60 + 1.79 + 1.79
O^a^=0.02, DQ^^^O.02. Let us see in detail the logarithmic quantization. If the error E
is greater than O^^x/^' ^^^'^ ^' ^^ quantized in a value between + 1 and +6. l{E<-0^^/2,
then it is quantixed in the interval -6, -1. If 0<E<O„,gj[/2, then it is quantized to +0. If
-Oj^g^/2<E<0, then it is quantized to -0. Analogous quantization is applied to AE. It
should be noted that these quantization levels are sufficiently large to cover the presence
of noise. However this quantization was behaved quite well.
8. CONCLUSIONS
This chapter was devoted to the study of fuzzy dynamic systems. In particular after a brief
exposiiton of fuzzy relational equations and their properties, a general state space model
of fuzzy systems was presented which makes use of such relational equations. Then, the
particular class of first-order fuzzy systems was examined, followed by a discussion of the
relation between classical difference equations and fuzzy state-space equations. Then the
control problem of fuzzy dynamic systems was considered, and the one-step and multi-step
cases were presented, followed by a remark on the construction of the controller equation.
The identification problem of fuzzy dynamic systems was then treated which includes the
estimation of the fuzzy relation R and the fuzzy state prediction. Finally, three important
fuzzy controllers were reviewed which have practical value and have been applied with
success at least in simulation experiments. The topic of fuzzy dynamic systems is still open
and much has to be done in several directions e.g. structural analysis, input - output
96
analysis, stability analysis, sensitivity analysis, parameter and state estimation, and
controller design.
REFERENCES
1. L.A. Zadeh: Fuzzy Sets, Inform, and Control, Vol. 8, pp. 338-353, 1965.
2. L.A. Zadeh: Outline of a New Approach to the Analysis of Complex Systems and
Decision Processes, IEEE Trans. Syst. Man, Cybern., Vol. SMC-3, pp. 28-44,
1973.
3. W. Pedrycz: An Approach to the Analysis of Fuzzy Sytems, Int. J. Control, Vol.
34, No.3, pp. 403-421, 1981.
4. E. Sanchez: Eigen fuzzy Sets and Fuzzy Relations, J. Math. Anal. Appl., Vol. 81,
pp. 399-421, 1981.
5. E.H. Mamdani, and TJ. Procyk: A linguistic self-organizing process controller,
Automatica, Vol.15, pp. 15-30, 1979.
6. H.R. Van Nauta Lemke, and W. De-Zhao: Fuzzy PID supervisor, 24th IEEE
CDC. Paper WP9-5:15, pp. 602-608. Ft. Lauderdale, FL. 1985.
7. S.G. Tzafestas, and N. Papanikolopoulos: Intelligent PID control based on fuzzy
logic, Proc. IFAC Symp. on Distributed Intelligence Systems Methods and
Applications (DIS '88), Varna - Bulgaria, June 1988.
8. P.J. MacVicar-Whelan: Fuzzy sets for man-machine interaction, Intl. J. Man-
Machine Studies, Vol. 8, 687-697, 1976,
9. P.J. King and E.H. Mamdani: The application of fuzzy control systems to
industrial processes, Proc 6th IFAC World Congress, Boston 1975,
Automatica, Vol 13, pp. 235-242, 1977.
This page intentionally blank
PART 2
NEURO-FUZZY SYSTEMS
CHAPTER 5
1. Introduction
The concepts of fiizzy sets and fuzzy systems proposed by Zadeh [18] [19] have many
possibilities of being applied to the description and analysis of this real fiizzy world. In
the application of these concepts to knowledge engineering, there are two problems, i.e.,
learning of fuzzy knowledges and automatic inference of unknown variables. On the
other hand, one can identify any complex functions by means of perceptron-type neural
networks [12] and the back-propagation error method [13].
The purpose of this chapter is to propose some methods of neural network
applications to fuzzy systems (and not of fuzzy system applications to neural networks
[10]). As mentioned above, the concept of neural networks is very useful for learning
knowledge and automatic inference. In this chapter, besides usual neural networks, a
couple of quasi-neural network methods are proposed.
In the field of engineering planning and design, evaluation and optimization are
performed by using many kinds of variables (e.g., load, size, weight, strength,
intensity, deformation, cost and so on) and constraints (e.g., design formula,
mathematical equation, function, code, criterion and so on). Due to inevitable
uncertainties, in general, these variables and constraints can be described with fuzzy
sets and fuzzy relations [17][18], respectively, which can include rationally usual crisp,
deterministic and probabilistic expressions as special states.
As conditioned probabilities are used in probability systems, so conditioned fuzzy sets
[1] are used as states in fuzzy systems composed of inputs, states and outputs [19]. In
engineering problems, such conditioned fuzzy sets which belong to fuzzy relations are
very useful enough to discriminate uncertainties in variables firom ones in constraints
(or functions) [3].
The final purpose of this paper is to propose a paradigm of intelligent and neural
fuzzy networks in which learning, identification, evaluation and optimization can be
performed for engineering planning and design by expanding fiizzy systems and
employing a couple of models similar to neural networks [10].
99
S.G. Tiafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 99-119.
© 1994 Ktuwer Academic Publishers. Printed in the Netherlands.
100
R(Y/X)=R(XA'), (2)
the system becomes undirected (solid and broken allows). For the purpose of the max-
min operation in fuzzy systems, fuzzy relations are described with matrix expression.
Consider a two-dimensional matrix fj. ^ (0 ^ /ij^^ I) as shown in Fig.2, and this matrix
means a state (i.e., fuzzy relation) R. If R is valid for both the cases of given x and
given y, the undirected fuzzy system can be illustrated without allows as shown in Fig.3.
y y y
1 2 3
A typical neural network is shown in Fig.4 [13]. Input-, hidden- and output-layer
unit values, Oj, Oj and oj^, are given by
f(u)=l/{l+exp(-u)} (8)
in which
0^a,yS,ri^l,(i=1.2,..-,n) (13)
The topological effects of the coefficients a and ^ on the linear combination are
illustrated in Fig.6. This fuzzy confluence rule can be illustrated as a kind of neural
network with hidden-layer units, u^, jUp, fi^jSni ^ g such as shown in Fig.7.
/ZR(x,y)^R
10 © Fuzzy
Confluence Rule
In engineering planning and design, when we search for such an optimal solution of
systems as it needs plenty of time or money to calculate, we have to predict it based on
103
limited results of analysis. Aiming at solving these problems, the authors already
proposed a multi-variable fuzzy identifier shown in Fig.9 [7].
the coefficients of Eq.l4, a 1, Q ; 2 ' ' " ' '^n+l ^^ ^ determined by solving simultaneous
linear equations.
When the following point in the n-dimensional space is given as an input,
Here, let the minimum b and the maximum b among bj, i=l,«.-,n, in Eq.l5 be yomin
and yomax' respectively, and one can illustrate a triangle-type membership function
[yOmin. yO' yOmaxl ^ shown in Fig.lO.
1- A^^bi ^^%^
function fx yo' '•^•» * ^ ^"^' output of the proposed fuzzy identifier are obtained as
shown in Fig. 10, where average means algebraic one with respect to membership values.
This multi-variable fuzzy identification can be illustrated as a kind of neural network
with two hidden-layer units, i.e., linear equation and avarage as shown in Fig.ll. The
authors already applied it to structural planning [2].
(J-yp^ R(Y/X)
Avarage
Linear Equations 1 , t
Fuzzy Fuzzy
f3 Identifier Identifier
^1 ^2 ^3 ''n y
a 10 320 ^30 ^nO (a) Input X (b) Input y
Fig.ll Multi-variable Fuzzy Identifier Fig.l2 Identification of R
as a kind of Neural Network m by Fuzzy Identifier
The simplest fuzzy identifier with input x is shown in Fig.l2(a), where output fi y/x is
conditioned membership function and corresponds to conditioned fuzzy set R(XA'). On
the contrary, when y is put into the fuzzy identifier, /i ^ A ' ^"'^ RCXA') can be obtained.
Therefore, the fuzzy identifier can be regarded as undirected fuzzy system as shown in
Fig.3.
2 - 5 - 1 . Neural Network
^ Y "•" 1
Fig.lS Safety Evaiuation with Respect to // x and |i y by Questionnaires
Output /i J
&i\A 4<"23f31
Z\%lJ\j wKaWi
hm
llf21''Or^ -yi^!2% ./'• 0
Ills X ^
Input- n
A
m In p«t-/f^(y)
Fig. 16 Assumed Neural Network Fig. 17 Safety Evaluation with Respect to
IIX and /i Y by Neural Network
106
To identify n ^'m. Fig. 15, a fuzzy confluence rule model is assumed as shown in
Fig.18 [16]. The paramctcis, a , jS and j are determined by the method of least
squares and showH in Table 2. The final output /^ g is plotted in Fig.l9 which sliows
very nice agreements to Fig.15.
Output /ig
The fuzzy identifier shown in Fig. 11 can present the optimal value yo whose
membership value is the maximum in /£ yo- Therefore, it enables us to examine its
exactness as a function [7]. Assume a function given by Eq.l8
y=(xl/x2)+x3 (18)
107
shown in Table 3 in which b j (i=l, - 1,5) is derived from Eq.18 and assumed an, a2i
anda3i(i=l,»»»,5).
Xl %! l~7(pitch:0.6)
^2 ^2 2~10(pitch:0.8)
^^3 %i 0.5~12(pitch:1.15)
1331 Cases
By using the fuzzy identifier model shown in Fig.ll and the initial data sets shown in
Table 3, the final optimal values of y arc estimated in the 1331 combination cases of
input data shown in Table 4. The comparison between the calculated values y^al by
fuzzy identifier and the real values Jfc-^i by Eq.18 is shown in Fig.20. In this case study,
they coincide well with each other.
17.5 ^real
12.5
ycal
17.5
Fig.20 Comparison between Real Vaie by y=(xl/x2)+x3 and Calculated ones
by Fuzzy Identifier
3 - 1 . Fuzzy Networks
CTa(Stress)
Now, let's go back to fuzzy networks as shown in Fig.21 where states R^, R2,»'> and
Rjjj are identified by learning given initial data. The next subject is to obtain necessary
or unknown variables Xj, Xj, etc. by fuzzy network reasoning. The most perfect
method is to get the total state Rj whose variables X]^,» • ^Xji are all unknown as shown
in Fig.23.
R,= n Rv
RXY(X,Y)ORYZ(Y,Z)=RXZ(X,Z), (19)
In this section, two kinds of practical and intelligent solving methods of fuzzy
network are shown aiming at object-oriented systems.
The first is frame knowledge representation method [6]. Table 5 shows common
frame knowledge representations, i.e.. Frame 10 for inputs and outputs and Frame S for
states of fuzzy systems. In these frames, "knowledge" represents fuzzy set, and D
shows input, state or output in sub-frames such as shown in Table 6. For Example, if
fuzzy systems A and C as shown in Fig.26, sub-frames, i.e., Frame IB, SB and OB are
shown in Table 6. Actually, sub-frames have to be constructed according to fuzzy
networks such as shown in Fig.21.
lA SA OA IB SB OB ]C sc oc
Fuzzy System A Fuzzy System B Fuzzy System C
Fig.26 Serial Fuzzy Network
110
By using Tables 5 and 6, when a certain variable or state is denoted as a final goal,
the most terminal variables or states are searched out (backward inference). When the
most terminal variables or states are given, the goal is obtained (forward inference).
algorithm is shown in Fig.28. When the most terminal variables are given, the goal is
obtained according to the max-min operation (forward inference).
Tenninal X^_i
iOj is 'Input'
iOj is deternined.
iOj and iSa are deleted from KBi.
The first method is not substantially different from the second one. In the first
method, however, the variables and states are not supposed to be given. So, it is
possible to denote even a state as a goal and even states can be pointed out as necessary
inputs. Only when all the states are given from the beginning, the second method
becomes usefUl.
Fig.29 shows an intelligent fuzzy network for aseismic structural evaluation and
optimum design in which directed fuzzy systems are employed. R, V, D and S means
the fuzzy sets which represent the likelihood of earthquake return period, maximum
ground velocity amplitude, seismic structural damage and structural safety, respectively.
V/R, DA' and S/D/R mean conditioned fuzzy sets which represent the likelihood of
seismicity, aseismic structural characteristic and criterion of aseismic structural safety,
respectively. S/D/R is described with non-fuzzy parameters D and R and given from
structural, economical and social points of view.
A B
S~ o
> V <-^ V D, > D
^/R 7V
^ ^/D/R
'^ R ^
R ;Retum Period of Seismic Load
V :Ve]ocily Amplitude of Ground Motion
D ;Damage to Structure
S :Safety
Fig.29 Intelligent Fuzzy Network for Aseismic Structural Evaluation
To evaluate aseismic structural safety, aseismic structural safety S is called initially.
Calling and reasoning flows (hollow and solid arrows) in the case of calling structural
safety are shown in Fig.30. According to the frame knowledge representations shown
by Tables 5 and 6, R, V/R, D/V and S/D/R are demanded of users. When the fuzzy sets
shown in the dottcd-line rectangles in Fig.30 are put into this network, the aseismic
structural safety S shown in the double-line rectangle is finally obtained.
When aseismic structural characteristic D/V is called initially, one can perform
optimum aseismic structural design. Calling and reasoning flows (hollow and solid
arrows) in the case of calling D/V are shown in Fig.31. According to the frame
knowledge representations shown by Tables 5 and 6, R, V/R, S/D/R and S are demanded
by users. When the fuzzy sets shown in the dottcd-line rectangles in Fig.31 are put into
the network, the aseismic structural characteristic D/V shown in double-line rectangle is
finally obtained. The membership function (/ Yi/w of D/V represents the optimum
degree of aseismic structural characteristics.
In these case studies, fuzzy inputs and fuzzy states are calculated by solving fuzzy
relational equations [14][11].
113
1 0.3 0 0 0
; Input InDut : : Call /y : : 20 0.3 1 0.3 0 0
: SI AT Input Input ;r=100 10 0 0.3 1 0.3 0
;L
L
f V/ '
> /R< flte ' D/ 1
&
3D ri AT
S/
AT
:s
;(Year) 5
;
i
2
(ci)
0 0 0.3 1
0 0 0 0.3
"S/D/100
0.3
1
In this case study, a standard type of reinforced concrete (R/C) transfer substations of
Kansai Electric Power Co. Ltd. is employed for aseismic structural planning. This
building is assumed to be located at Kobe City in Japan. Aseismic optimization and
evaluation are performed in the X-longitudinal direction. Fig.32 shows a fuzzy network
composed of undirected fuzzy systems for seismic planning of R/C building. • and [Q]
mean variables and states, respectively. The variables are defined as shown in Table 7.
The states arc already given from seismic and structural points of view [9].
Aseismic optimization is performed with respect to optimal number of shear wall
units. If N-WALL in Fig.32 is nominated for the final output, then the fuzzy network is
transformed into a fuzzy hierarchy such as shown in Fig.33. Considering the properties
of the R/C structure and using the inputs described with fiizzy sets shown in Fig.34,
calculated variables in structural optimization are obtained as shown in Fig.35. Fig.36
shows the region of optimal numbers of shear wall units illustrated with membership
functions given by maximizing decision [1], which means the optimal number is about
9-10.
STATHEfflil
MS-I
IM-WAG
fOTfj^M-CEQ
d= m MB^Mm
^^IMWIMI mm
: STATE
MM-MM-^M
Fig.32 Fuzzy Network for Structural Planning
iKMETMTlFMffllPERTillEM
• INPUTS
Fig.33 Transformed Fuzzy Hierarchy for Structural Optimization
„ ->—I—.—S.HEIGHT 0 FUNCTION
350 ICB) (oml (cm) D.n.4 0 i D i l . D
U
10' 10'
(21 U
SrATE9 (5) -"jn STATE5 '^' ^i-tEsiw STATEIQ " " ^ - « L L STATEll
10
\**^^ , FUNCTION
I 0 •ECONOMY
SAFETY
0 N-WALL
5 »N 13
Fig.36 Decision of Optimal Number of Shear Wall Units
...... STATE J
/^ ^ 1 •com EiGiir
10' lO' lO'
I
^WEIGHT nl—•—L .N-COLM 0 J—.—>D-COLM 0 .HEIGHT
2200 (t) 22 60 (cm) 350
•lAlL /^PERIOD
ID- , 10 LO-
I
^.
^IK-NECE 0 .N-ALL
58 50 62 (xiO^cnl 20
4. Discussions
To identify states (i.e., fuzzy relations) in fuzzy systems, in this chapter, neural
network, fuzzy confluence rule and fuzzy identifier are proposed. We have to pay much
attention to the problem that each of the models has its own merits and demerits. Neural
network is available for any type of functions, but it is a "black box". Fuzzy confluence
rule is effective for human judgements, and it is not a "black box". Fuzzy identifier is
appropriate for engineering problems in which design equations and functions are given.
Anyway, it is desirable for us to have many alternatives for such an identification.
In this real world, almost events and judgements may be transformed into fuzzy
networks as shown in Fig.21, where some variables and states are given and others are
not given. The calculation procedure of states as shown in Fig.31 can be called
"learning", and the calculation procedure of variables as shown in Figs.30, 33 and 37 can
be called "inference". By means of local identifications of states in fuzzy systems
and/or macroscopic leaning of state in fuzzy networks, every state in fuzzy networks can
be determined, which enables iis to infer unknown variables. When these learning and
inference are performed automatically, fuzzy networks can be called "intelligent fuzzy
networks", or "ncuro-fuzzy networks".
5.C0NCLUSI0NS
Acknowledgement: The parts of this research are performed under the financial supports
by Grant-in-Aid for Scientific Research (C) (The Ministry of Education, Science and
Culture, Japanese Government), Kansai Electric Power Co. Ltd., Newjec Co. Ltd., and
the Secom Science and Technology Foundation. The authors are grateful to all the
graduate students who coworked in the authors' laboratory.
6. References
[7] Kawamura,H., and Tani,A-, "Multi-variable Fuzzy Identifier," Proc, 6th Fuzzy
System Symposium, Tokyo, Japan, pp.179-182: Sept.1990. (in Japanese)
[8] Kawamura,H., Tani,A., Kambara,H., and Yamada,M., "Intelligent Fuzzy Network
for Optimum Structural Planning and Design," Proc., 7th Fuzzy System Symposium,
Nagoya, Japan, pp.99-102: June 1991. (in Japanese)
[9] Kawamura,H., Tani,A., and Kambara,H., "Aseismic Structural Planning System by
Fuzzy Network," Proc. 10th World Conference on Earthquake Engineering, Vol.10,
Madrid, Spain, pp.6271-6275: July 1992.
[10] Kawamura,H., Tani,A., "A Paradigm of Intelligent Fuzzy Networks," Proc, The
IEEE International Conference on Systems Engineering, Kobe, Japan, pp.159-164:
Sept. 1992.
[11] Papis,C.P. and Sugeno,M., "Fuzzy Relational Equations and the Inverse Problem,"
Fuzzy Sets and Systems, 15, pp.79-90: 1985.
[12] Rosenblatt.F., "The Perccptron:A Probabilistic Model for Information Storage and
Organization in the Brain," Psychological Review, Vol.65, No.6, pp.386-408: 1958.
[13] Rumelhart.D.E., Hinton,G.E., and Williams,R.J., "Learning Representations by
Back-propagation Errors," Nature, 323-9, pp.533-536: Oct. 1986.
[14] Sanchez,E., "Resolution of Composite Fuzzy Relational Equations," Information
and Control, 30, pp.38-48:1976.
[15] Yamada,M., Kawamura,H., Tani,A., and Yamamoto,K., "A Determination Method
of Hierarchy Models for Decision Making in Aseismic Structural Design by Fuzzy
Confluence Rule," Proc., The 12th Symposium on Computer Technology of
Information, Systems and Applications, Kyoto, Japan, pp.223-228: Dec. 1989. (in
Japanese)
[16] Yamada.M., Kawamura,H., and Tani,A., "Research on Quantitative Formulation
and Objective Determination Methods of Multi-Objective Decision Making Process
in Aseismic Structural Design (Comparison between Fuzzy Confluence Rule and
Neural Network)," Proc. of Annual Meeting, Structural Division., Airchtectural
Institute of Japan, Kinki Branch, pp.161-164: May 1991. (in Japanese)
[17] Yao,J.T.P, Safety and Reliability of Existing Structures, Pitmann Publ. Inc.,
Boston, London, Melbourne: 1985.
[18] Zadch,L.A.,"Fuzzy Sets," Information and Control, Vol.8, pp.338-353:1965.
[19] Zadeh,L.A., "Toward a Theory of Fuzzy Systems," in Aspects of Network and
System Theory, (Eds. Kalman,R.E. and DeClaris,N.), Holt, Rinehart and Winstcr,
Inc.: 1971.
CHAPTER 6
N E U R O - F U Z Z Y E X P E R T SYSTEMS: O V E R V I E W
WITH A C A S E S T U D Y
I. Introduction
Artificial neural networks or connectionist models [1,2,3] are massively parallel
interconnections of simple neurons that function as a collective system. They are
designed perhaps as an attempt to emulate human performance and function intel-
ligently. An advantage of neural nets lies in their high computation rate provided
by massive parallelism, so that real-time processing of huge data sets becomes
feasible with proper hardware. Information is encoded among the various connec-
tion weights in a distributed manner! .The multilayer perceptron (MLP) [2] is a
feed-forward neural network model consisting of multiple layers of simple, sigmoid
processing elements (nodes) or neurons. Alter a lowermost input layer there are
usually any number of intermediate or hidden layers followed by an output layer at
the top. The learning procedure has to determine the internal parameters of the
hidden units based on its knowledge of the inputs and desired outputs.
An expert system [4,5] is a computer program that functions in a narrow domain
dealing with specialized knowledge generally possessed by human experts. Such
programs are very useful due to the usual shortage of qualified human experts in real
life. The primary characteristics of an expert system are a knowledge base designed
with the help of a human expert, a narrow problem domain, and a performance
on par with a human expert. The knowledge base is a problem-specific module
containing information that controls inferencing. Traditional rule-based expert
systems encode this information as If-Thtn rules while the connectionist expert
system [6] uses the set of connection weights of a trained neural net model for this
purpose. The inference engine is problem independent while the user interface links
the external environment to the system. Connectionist expert systems are usually
suitable in data-rich environment. They help in minimizing human interaction
and associated inherent bias during the phase of knowledge base formation (which
is time-consuming in case of traditional modeb) and also reduce the possibility
of generating contradictory rules. The rule generation phase of such connectionist
models are usually completely automated. An expert system is expected to be able
to draw conclusions without seeing all possible external inforrfiation. It should be
capable of directing the acquisition of new information in an efficient manner and
also be able to justify a conclusion reached.
The knowledge base is generally a major source of uncertain information. Besides,
121
S.G. Tzafestas and A.N. Venelsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 121-143.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
122
Output
OR
AND
Input
Figure 1 :, A three-layered neural network implementation And and Or logic functions
at successive layers.
123
This fuzzy logical model broadly performs two main tasks. First we construct
the three-layered fuzzy logical network for classifying multi-class patterns. Then,
the trained network is used to generate rules. The connection weights in this stage
constitute the knowledge base (in embedded form) for the classification problem
under consideration. The model is now capable of inferring the output decision
for complete and/or partial inputs along with a certainty measure and querying
the user for the more essential missing input information. If asked by the user,
the model is capable of justifying its decision in If-Then rule form (in terms of the
salient features) with the antecedent and consequent parts produced in linguistic
and natural terms. The effectiveness of the model is demonstrated on a set of
speech data.
A. Involvement of uncertainties
However the knowledge base itself b a major source of uncertain information [5] in
expert systems, the causes being :
• unreliable information
t imprecise descriptive languages
• inferencing with incomplete information
• poor combination of knowledge from different experts
expression of rules in formal language. Therefore the meaning of the facts have to
be approximately matched with those of the premises. Thirdly, when the available
information is incomplete the system accepts the value unknown while evaluating
the premise's degree of certainty during the approximate pattern matching. Finally,
generation of a consensus knowledge base by combining the views of multiple ex-
perts is often difficult, especially when the experts have contradictory viewpoints.
Moreover, problems occur as human beings generally do not have uniform levels
of expertise throughout a domain. Some conflict resolution strategies need to be
followed, like separately weighting the knowledge of each expert, to calculate a
composite conclusion.
Good expert systems are expected to be capable of handling uncertainty, as in
most cases the data are inherently inexact, incomplete or unmeasurable. Some
of the existing methods in this direction include the subjective probability theory,
Dempster-Shafer theory [23], possibility theory [24] and certainty factors [22]. Us-
ing the subjective Bayesian method, Prospector [25] has been designed to provide
advice on mineral exploration. Mycin is a rule-based expert system (using certainty
factors) which attempts to reconmiend appropriate therapies for patients with bac-
terial infections. It interacts with the physician to acquire the clinical data it needs.
The associated Teiresias system enables the doctors to interact with Mycin to cisk
it questions about its recisoning and to modify and augment its knowledge base.
The knowledge base of an expert system contains human knowledge, most of
which is imprecise and qualitative. To describe situations where the boundary
between competing hypotheses is vaguely defined, human experts use terms such
as very likely, likely, more or less likely, low, medium, high, etc. Encoding this
sort of expertise by probabilities causes this fuzziness or imprecision to be lost.
Fuzzy logic may be used to express these vague terms with precision and accuracy.
Besides, facts about the world are rarely known with certainty. Conventional rule-
based systems, with two-valued logic, usually evade this issue of partial matching.
theories are not applicable, as the latter often require idealized assumptions such
as the independence of evidence and the mutual exclusiveness and exhaustiveness
of hypotheses.
Some approaches in fuzzy inferencing for traditional expert systems include the
approximate analogical reasoning based on similarity measures by Turksen and
Zhong [26], the problem-reduction method of Ishizuka et al [27], modelling of physi-
cian decision processes by Esogbue and Elder [28], inferencing in the framework
of inflammatory protein variations by Sanchez and Bairtolin [29] fusing weighting)
and the extension to Clips for manipulating fuzzy facts and rules by Sosnowski
[30].
in the traditional models) and allows dynamical adjustments to changes in the en-
vironment. Besides, the various characteristics of neural nets, viz., generalization,
tolerance to noise, graceful degradation at the border of the domain of expertise,
ability to discover new relations between vziriables, etc., are in-built and hence can
be utilised by the connectionist expert system model.
In this article we restrict our discussion to classification-type expert systems. A
set of environmental stimuli (features or symptoms) is clcissified into one or more of
many possible categories (output classes or diseases). The output is expressed as
continuous membership values to the various classes. This is suitable for handling
data in fuzzy input feature space where a pattern may have finite belongingness to
more than one class. In the medical domain this is analogous fo diagnosing a par-
ticular symptom set as affliction by more than one disease. Using fuzzy neural nets
for this purpose, helps one to incorporate the advantages of approximate reasoning
into the connectionist expert system. In this section we provide a discussion on the
various existing neurofuzzy approaches to inferencing and expert system design.
Sections IV and V present respectively a fuzzy logical version of the MLP and its
application to automatic rule generation.
Fuzzy inferencing has been used by Nakanishi and Takagi [36] for the recognition
of noisy patterns (English alphabets distorted by ink marks). An alphabet is
divided into n blocks X = {Xi,X^,...,X„), with each block being processed by a
separate neural network. Each block is divided into m smaller subblocks such that
the neural network NNj for block Xj has m inputs ^],^], • • •, ^T- Each NNj has
the number of outputs equal to the number of patterns to be classified and each
output is mapped into membership values to the three fuzzy sets zero, positive
medium, and positive big. The fuzzy outputs from the n neural nets are then used
to formulate the antecedents of fuzzy inference rules to recognise a noisy pattern
when its typical feature is still visible in a small block of the irik-blofched pattern.
An approach to fuzzy inferencing by Zhuang et al [37], for controlling a mathe-
matically intractable system, uses a multilayer network with each node standing for
a fuzzy subset of a linguistic variable. The truth values flow from the input nodes
to the output nodes according to the rule of propagation being modelled. The
interconnection weights are associated with ordered pairs of real numbers lying in
the range [-1,1] and are indicative of the certainty factors used to represent the re-
liability of the fuzzy control rules. The basic operations in fuzzy logic, viz., union,
intersection and negation, are implemented at the neuron levels. The linguistic
variables used are positive large, positive medium, positive small, zero, negative
small, negative medium and negative large.
The role of fuzzy logic in the control of the activation, training, reliability and
performance of neural networks has been investigated by Yen [38] using a hybrid
architecture. Fuzzy rules are used to detect situations under which certain ac-
tions need to be invoked for neural network modules based on their performance
measures. The networks process data obtained either from external sensor devices
or from the knowledge base of the symbolic system. The global knowledge base
consists of a fuzzy database and a neural network taxonomy (of neural net classes)
that describes meta-level knowledge about the neural nets themselves (using crisp
or fuzzy attribute values). The fuzzy database stores data and hypotheses that
can be uncertain, imprecise or vague. A production system that takes into account
the degree of partial matching of the fuzzy action rules thus enables the system to
respond in a robust way even in the face of incomplete or noisy data.
Masuoka et al [39] use knowledge in the form of membership functions and fuzzy
rules (in And-Or form), extracted from experts, to build and preweight the struc-
tured neural network which is then tuned using selected learning data. This neural
model consists of the input variable membership net, the rule net, and the output
variable net. Modified fuzzy rules, extracted from the trained neural network us-
ing pruning, can then be evaluated and unsuitable rules corrected using relearning.
Okada et al [40] use a similar approach to examine the bond rating of investors.
Changes in the weights and threshold levels of the neurons (after training) are
interpreted as adaptations in the membership functions and fuzzy rules. Yager
[41] employs membership neural modules for the antecedents, inverse membership
modules for the consequents and a rule neural module with a combiner (using min
or product functions) for modelling the rules of fuzzy logic controllers. The various
weights 2Lre learned and the importance of the antecedent clauses simulated.
The neural network based fuzzy reasoning scheme by Takagi and Hayashi [42]
is capable of learning the membership function of the If part and determining
the amoimt of control in the Then part of the inference rules. The input data
is clustered to find the best number of partitions corresponding to the number
of inference rules applicable to the reasoning problem, a single neural net block
128
modelling one rule. The optimum number of cycles required is determined to avoid
overlearmng and the minimal number of input variables selected for inferring the
control values.
A few of the other existing neurofuzzy approaches include the rule-based phoneme
recognition method by Amano and Aritsuka [43] for Japanese words, Takahashi
and Minami's technique [44] of modelling the subjective evaluation of humans,
the linguistic fuzzy modelling scheme of Horikawa et al [45], the knowledge-based
image interpretation system by Zahzah et al [46], and the self-learning rule-based
controller by Lee [47]. Among the applications in the medical domain are the
MLP Eind radial basis function network models of Bounds et al [48j that detect low
back disorders using crisp input-output values and considering networks of multiple
single-class nets in the process. Some image processing techniques are combined
by Silverman and Noetzel [49] to allow the MLP to recognize tumors (at various
scales, orientations and locations) from ultrasonic images. Pixel information is
used as input along with crisp output values, and the model acts as a classifier.
where yf is the state of the t"* neuron in the preceding /i"* layer and w^- is the
weight of the connection from the i"" neuron in layer h to the j " * neuron in layer
ha. For nodes in the input layer, y° corresponds to the j " " component of the input
vector.
The Least Mean Square (LMS) error in output vectors, for a given network weight
vector w, is defined ELS
where yyf(.(w) is the state obtained for output node j in layer H in input-output case
c and djc is its desired state specified by the teacher. The error E[w) is minimized
by the back propagation algorithm using gradient-descent. We start with any set
of weights and repeatedly update each weight by an ajnount
dE
^""^'^^^ " ''d^i ^ aA<(i-l) (3)
where the learning rate e controls the descent, 0 < a < 1 is the damping coefficient
or momentum and t denotes the number of the iteration currently in progress.
iF<}] (4)
where the /x values indicate the membership functions to the corresponding linguis-
tic TT-set [18] for eeich feature axis.
131
When Fj is numerical we use the ff-fuzzy sets [66] (in the one-dimensional form),
with range [0,1], given as
• 2 ( l - M ) ' , for|<|f,-c|<A
7r(F,-;c,A) = • 1 _ 2 ( M ) \ f o r O < | f , - c | <A| (5)
0, otherwise
where A > 0 is the radius of the 7r-function with c as the central point.
When the input feature is linguistic, its membership values for the 7r-sets low,
medium and high are quantified as
fO.95 0.6 0.021
where 2,^ is the weighted distance between the t"" pattern and the mean of the fc"*
class and the positive constants Fj and F^ are the denominational and exponential
fuzzy generators controlling the amount of fuzziness in this class-membership set.
Then for the x"' input pattern we define the desired output of the j ' " * output node
as
dy-My(^.) (7)
During testing, the output of the j " " output neuron denotes the inferred member-
ship value of a test pattern to the y"" class.
if;i = i
otherwise
where the T{S) operation at layer h is performed over all / S[T) operation out-
puts from the neurons in the preceding layer h — 1, provided / ^ i, for h = 1(2)
respectively; also let
S{yt-\w^n if/» = l
smf={ y \ ' V ,{ otherwise (15)
T{y^-\w^rA
133
r(5m*,fim^)j ifA = l
(16)
dwii 1 « ^fsm^jSmfj otherwise
where the t-norm T and t-conorm S aire given by eqn. (8) in order to model the
logical operators And cind Or. Expanding, we have
(17)
dwji 1 (1 - sm'')y^~^ otherwise
Analogously, we compute the sensitivity mecisure as
^ = (1 - . m ' - X - (18)
Substituting the values of ^ ^ and j ^ from eqns. (17-18) into eqns. (11,13) en-
ables one tp evaluate the error derivative j ~ _ of eqn. (3) and,thereby update the
connection weights during training. This constitutes the back propagation algo-
rithm for a network incorporating logical nodes.
Note that the e of eqn. (3) is gradually decreased in discrete steps, taking val-
ues from the chosen set {2, 1, 0.5, 0.3, 0.1, 0.05, 0.01, 0.005, 0.001}, while the
momentum factor a is also decreased. Let the various values of e be indicated by
Eo = 2,£i — l , . . . , f 5 — 0.001 such that EJ indicates the (t-f- l)"" value of e. Let
ao = 0.9 and ai - a2 = ... = a^ = 0.5. We use
if mse[nt — kn) — m&t[nC) < 6 f^^s
{:^^ otherwise ^ '
where i = 0 initially, |e| == g -I- 1 and Q < 5 < 0.0001. Note that nise{nt) is the
mean square error at the end of the nt*'^ sweep through the training set and kn is
a positive integer such that mse is sampled at intervals of kn sweeps. The process
is terminated when i > q.
A. Input form
The input for a test pattern can be in quantitative, linguistic or set forms or
a combination of these. It is represented as memberships to the three primary
134
If o o r t
r *
weights
•
known or When
I > unobVo noble output
neuron (s)
unknown
Figure 2 : Block diagram of the inferencing and rule generation phase of the proposed
fuzzy logical neural network.
linguistic properties low, medium and high as in eqn. (4), modelled as 7r-functions.
When the information is in exact numerical form like Fj is rj, say, we use eqn. (5).
However, when the input is given as Fj is prop (say), where prop stands for any
of the primary linguistic properties low, medium or high, eqn. (6) is used. The
model can also handle the linguistic hedges [67] very, more or less and not using
the Concentration (Con) and Dilation (Dil) operators [8]. In case of set form
usage, the input is a mixture of linguistic hedges and quantitative terms. Since
the linguistic term increases the impreciseness in the information, the membership
value of a quantitative term should be lower when modified by a hedge [67]. The
modifiers used are about, less than, greater than and between.
If any input feature Fj is not available or missing, we clamp the three corre-
sponding neurons x° = ij_i = i j ^ j = 0.5, such that k = (j — 1) * 3 + 1. We
use
rO.5 0.5 0.51 (20)
no information
as 0.5 represents the most ambiguous value in the fuzzy membership concept. We
also tag these input neurons with noinf^ — noinf°^^ — noinf°j^^ = 1. Note that
in all other cases the variable noinf^ is tagged with 0 for the corresponding input
neuron k, indicating absence of ambiguity in its input information.
B. Forward pass
Associated with each neuron j in layer h + 1 are its confidence estimation factor
conf^*^, a variable unknown'^'^^ providing a measure of the weighted information
from the preceding ambiguous neurons t in layer h (having noinf^ = l) and a
variable inoiyny"*"' giving a measure of the weighted information from the remaining
non-ambiguous preceding neurons (with noinf^ = 0). For neuron j in layer h > 0
135
we define
for A = 1
unknown^ — (21)
otherwbe
and
(22)
I
known j =
S[yt\w''-'
T{yi
h-l u fOT h=l
otherwise
(23)
for all i with noinf-^~^ = 0, where T and 5 stand for the conjugate pair of t-norm
Bind t-Conorm defined in eqn. (8). Here the T[S) operation at layer h is performed
over all t S{T) operation outputs from the neurons in the preceding layer /i — 1 for
/i = 1(2) respectively. Note that for /i > 0 we have
If there is no neuron j with noinfj = 1, then the system finalises the decision
inferred irrespective of whether the input information is complete or partial. In
case of partial inputs, this implies presence of all the necessary features required
for taking the decision.
A certainty measure (for each output neuron) is defined as
cert (26)
where 0 < cert^ < 1. The higher the value of certf, the lower is the difficulty in
deciding an output class j and hence the greater is the degree of certainty of the
output decision. Depending on the value of certf, the final inferred output may
be given in natural form.
C. Querying
If there is any neuron j in the output layer H with noinff = 1 by eqn. (24), we
begin the querying phase. We select the unknown output neuron j'l from among the
neurons with noinfj — 1 such that confj^^ by eqn. (25) (among them) is maximum.
136
Then we pursue the path from neuron j i in layer H, in a top-down manner, to find
the ambigupus neuron t'l in the preceding layer [h = H — l) with the greatest
absolute influence on neuron ji. This process is repeated down to the input layer
(A = 0). We select : = I'l such that with noinf^ = 1, for 0 < A < ff — 1, we have
For node t'l in the input layer (/i = 0), the model queries the user for the value of
the corresponding input feature ui.
Note that if a missing input variable by eqn. (20) is queried and found to be
missing once again, we now tag it as unobtainable. The inferencing mechanism
treats such variables as known with values i j ^ = a;°,+i = arj^^j = 0.5 but with
noinf^^ = noinf°^^y^ = noinf°^^2 ~ 0) s^<^^ ^^^^ '^i = (ui — l) * 3 + 1. The
response from an unobtainable input variable can allow the neuron activations in
the following layers to become non ambiguous with noinf^ = 0 such that an output
decision may finally be inferred.
D. Justification
In this phase the user can ask the system why it inferred a particular conclusion.
The system answers with an // - Tlien rule, applicable to the case at hand, in
terms of the salient input features. It is to be noted that these If - Then rules
are not represented explicitly in the encoded knowledge base; they are generated
by the inferencing system from the connection weights as and when needed for
explanations. As the model has already inferred a conclusion (in this stage), we
take a subset of the currently known input information to justify this decision.
Let the user ask for justification about a conclusion regarding class j . We choose
neuron t from the hidden layer fT — 1 if w^-~^ > 0, i.e., if it has a positive impact
on the conclusion at output neuron j . Let a set of nin-i neurons be so selected.
For the remaining layers we obtain the maximum weighted paths through these
neurons down to the input layer.
Let the set of the selected mo input neurons be given by la^, a ° , . . . , a^^ | and their
corresponding path weights to neuron j in layer H be denoted as Iwet^OjWet^o,...,
wet^^o^ \. We arrange these neurons in the decreasing order of their net impacts,
where we define the net impax;t for neuron : as
net impacti — y,° * wetio
Then we generate clauses for an If-Then rule from this ordered list until
J2wetio>2j2^etio (27)
where i, indicates the input neurons selected for the clauses and t„ denotes the input
neurons remaining from the set j a j , a°,..., a^\. This heuristic allows selection of
137
those currently active input neurons (current evidence of test patterns) contributing
the most to the final conclusion (ajnong those lying along the maximum weighted
paths of the embedded knowledge base to the output node j) as the clauses of the
antecedent part of a rule to justify the current inference.
Let u,, be the input feature corresponding to a neuron t',, in the input layer
[h — 0), selected for clause generation. The antecedent of the rule is given in
linguistic form with the linguistic property being determined cis
[ low if t „ - 3fu,. - 1) = 1
prop — < medium if t„ — 3(u,, — l) = 2 (28)
[ high otherwise
A linguistic hedge very, more or less or not may also be attached to the linguis-
tic property obtained for the antecedent part. We use the mean square distance
<f(u,i,prm) between the 3-dimensional input values (components of the pattern vec-
tor) at the neurons corresponding to feature u,^ and the linguistic property prop
obtained from u,, by eqn. (28) represented as pr^. The p r „ for which d[u,^,pr^
fs the minimum is selected as the antecedent clause corresponding to feature u,,
(or neuron t,j).
The procedure is repeated for all the input neurons selected by eqn. (27) to
generate a set of antecedent clauses for the rule justifying the inference regarding
output node j . The clauses are conjunctive when the corresponding generated
paths bifurcate at layer 1 nodes and are disjunctive when the said bifurcation occurs
at layer 2 nodes. This follows from the And-Or structure of the network given by
eqns. (9-10). Note that the selected clauses correspond to the salient input features
(determined both by the connection weight magnitudes learned during training as
well as the input feature components of the test pattern under consideration). All
input features (of the test pattern) need not necessarily be selected for antecedent
clause generation.
The consequent part of the rule can be stated in quantitative form as membership
value yf to class j . However a more natural form of decision can also be provided for
the class j , having significant membership value yf, considering the value of certf
by eqn. (26). For the linguistic output form we use : very likely for 0.8 < cert^ < 1,
likely for 0.6 < certf < 0.8, more or less likely for 0.4 < certf < 0.6, not unlikely
for 0.1 < certf < 0.4 and unable to recognize for certf < 0.1.
BOO
' 50C-
c
iOG
30C
2CC
600 SCO uQQ !500 1300 2100 2<C00 2700
Fj in Hz
Table 1 : Rule Generation for Vowel data with Fuzzy Logical Model
References
[1] R. P. Lippmann, "An introduction to computing with neural nets," IEEE Acoustics,
Speech and Signal Processing Magazine, vol. 61, pp. 4-22, 1987.
[2] D. E. Rumelhart and J. L. McClelland, eds., Parallel Distributed Processing. Vol. 1,
Cambridge, NLA.: MIT, 1986.
[3] T. Kohonen, Self-Organization and Associative Memory. Berlin: Springer-Verlag,
1989.
[4] E. Rich, Artficial Intelligence. Singapore: McGraw Hill, Inc., 1986.
[10] A. Kandel, Fuzzy Mathematical Techniques with Applications. Reading, MA: Addison-
Wesley, 1986.
11] H. -J. Zimmermann, Fuzzy Set Theory • and its Applications. Boston: Kluwer, 1991.
121 L. A. Zadeh, "The role of fuzzy logic in the management of uncertainty in expert
systems," Fuzzy Sets and Systems, vol. 11, pp. 199-227, 1983.
13] Y. H. Pao, Adaptive Pattern Recognition and Neural Networks. Reading, MA:
Addison-Wesley, 1989.
14) B. Kosko, Neural Networks and Fuzzy Systems. New Jersey; Prentice Hall, 1991.
15] Proc. of Second International Conference on Fuzzy Logic and Neural Networks
(UZUKA92), (lizuka, Fukuoka, Japan), July 1992.
16] Proc. of First IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), (San
Diego, USA), March 1992.
171 M. E. Cohen and D. L. Hudson, "Approaches to the handling of fuzzy input data
in neural networks," in Proceedings of 1st IEEE Conference on Fuzzy Systems, (San
Diego), pp. 93-100, 1992.
18] S. K. Pal and S. Mitra, "Multi-layer perceptron, fuzzy sets and classification," IEEE
Transactions on Neural Networks, vol. 3, no. 5, 1992,
19] A. Di Nola, S. Sessa, W. Pedrycz, and E. Sanchez, Fuzzy Relation Equations and
their Applications to Knowledge Engineering. Dordrecht, The Netherlands: Kluwer
Academic Publishers, 1989.
20] L. Shastri, "A connectionist approach to knowledge representation and limited infer-
ence," Cognitive Science, vol. 12, no. 3, 1988.
21] J. Pearl, "Fusion, propagation sind structuring in belief networks," Artificial Intelli-
gence, vol. 29, pp. 241-288, 1986.
22] B. G. Buchanan and E. H. Shortliffe, eds., Rule-based Expert Systems : The MYCIN
Experiments of the Stanford Heuristic Programming Project. Reading, MA: Addison-
Wesley, 1984.
23] G. Shafer, A Mathematical Theory of Evidence. Princeton, NJ: Princeton University
Press, 1976.
24] L. A. Zadeh, "Fuzzy sets as a basis for a theory of possibility," Fuzzy Sets and Systems,
vol. 1, pp. 3-28, 1978.
25] R. O. Duda, P. E. Hart, and N. J. Nilsson, "Subjective Bayesian methods for a rule-
based inference system," in Proceedings of the National Computer Conference, (USA),
pp. 1075-1082, 1976.
261 I. B. Turksen and Z. Zhong, "An approximate analogical reasoning schema based on
similarity measures and interval-valued fuzzy sets," Fuzzy Sets and Systems, vol. 34,
pp. 323-346, 1990.
271 M. Ishizuka, K. S. Fu, and J. T. P. Yao, "Inference procedures under uncertainty for
the problem-reduction method," Information Sciences, vol. 28, pp. 179-206, 1982.
28] A. O. Esogbue and R. C. Elder, "Fuzzy sets and the modelling of physician decision
processes, Part I : The initial interview - information gathering session," Fuzzy Sets
and Systems, vol. 2, pp. 279-291, 1979.
141
[29] E. Sanchez and R. Bartolin, "Jhizzy inference and medical diagnosis, a case study,"
Biomedical Fuzzy Systems Bulletin, vol. 1, pp. 4-21, 1990.
[30] Z. A. Sosnowski, "An extension of Clips for processing fuzzy data," Tech. Rep. NRCC
Pub.No. 31506, Laboratory for Intelligent Systenis, Division of Electrical Engineering,
National Research Council, Ottawa, Canada, 1990.
[31] D. C. Kuncicky and A. Kandel, "A fuzzy interpretation of neural networks," in Pro-
ceedings of Srd International Fuzzy Systems Association Congress, University of Wash-
ington, (Seattle, Washington), pp. 113-116, 1989.
[32] H. Takagi, "Fusion technology of fuzzy theory and neural network - Survey and future
directions," in Proceedings of the 1990 International Conference on Fuzzy Logic and
Neural Networks, lizuka, (Japan), pp. 13-26, 1990.
[33] J. M. Keller and H. Tahani, "Implementation of conjunctive and disjunctive fuzzy
logic rules with neural networks," International Journal of Approximate Reasoning,
vol. 6, pp. 221-240, 1992.
[34] J. M. Keller, R. R. Yager, and H. Tahani, "Neural network implementation of fuzzy
logic," Fuzzy Sets and Systems, vol. 45, pp. 1-12, 1992.
[35] H. Ishibuchi, H. Okada, and H. Tanaka, "Interpolation of fuzzy if-then rules by neural
networks," in Proceedings of Snd International Conference on Fuzzy Logic and Neural
Networks, lizuka, (Japan), pp. 337-340, 1992.
[36] S. Nakanishi and T. Takagi, "Pattern recognition by neural networks and fuzzy infer-
ence," in Proceedings of the 1990 International Conference on Fuzzy Logic and Neural
Networks, lizuka, (Japan), pp. 183-186, 1990.
[37] W. P. Zhuang, W. Z. Qiao, and T. H. Hang, "The truth-valued flow inference net-
work," in Proceedings of the 1990 International Conference on Fuzzy Logic and Neural
Networks, lizuka, (Japan), pp. 267-281, 1990.
[38] J. Yen, "The role of fuzzy logic in the control of neural networks," in Proceedings of the
1990 International Conference on Fuzzy Logic and Neural Networks, lizuka, (Japan),
pp. 771-774, 1990.
[39] R. Masuoka, N. Watanabe, A. Kawamura, Y. Owada, and K. Asakawa, "Neurofuzzy
system - fuzzy inference using a structured neural network," in Proceedings of the
1990 International Conference on Fuzzy Logic and Neural Networks, lizuka, (Japan),
pp. 173-177, 1990.
[40] H. Okada, N. Watanabe, A. Kawamura, K. Asakawa, T. Taira, K. IshiJa, T. Kaji, and
M. Narita, "Knowledge implementation multilayer neural networks with fuzzy logic,"
in Proceedings of Snd International Conference on Fuzzy Logic and Neural Networks,
lizuka, (Japan), pp. 99-102, 1992.
[41] R. R. Yager, "Implementing fuzzy logic controllers using a neural network framework,"
Fuzzy Sets and Systems, vol. 48, pp. 53-64, 1992.
[42] H. Takagi and I. Hayashi, "Artificial neural network driven fuzzy reasoning," Inter-
national Journal of Approximate Reasoning, vol. 5, pp. 191-212, 1991.
[43] A. Amano and T. Aritsuka, "On the use of neural networks and fuzzy logic in speech
recognition," in Proceedings of International Joint Conference on Neural Networks,
(Washington D.C.), pp. 301-305, 1989.
142
[44] H. Takahashi and H. Minami, "Subjective evaluation modelling using fuzzy logic and
a neural network," in Proceedings of Srd International Fuzzy Systems Association
Congress, University of Washington, (Seattle, Washington), pp. 520-523, 1989.
[45] S. Horikawa, T. Furuhashi, and Y. Uchikawa, "A new type of fuzzy neural network for
linguistic fuzzy modelling," in Proceedings of 2nd International Conference on Fuzzy
Logic and Neural Networks, lizuka, (Japan), pp. 1053-1056, 1992.
[46] E. h. Zahzah, J. Desachy, and M. Zehana, "A fuzzy connectionist approach for a
knowledge based image interpretation system," in Proceedings of 2nd International
Conference on Fuzzy Logic and Neural Networks, lizuka, (Japan), pp. 1135-1138,
1992.
[47] C. C Lee, "A self-learning rule-based controller employing approximate reasoning
and neureil network concepts," International Journal of Intelligent Systems, vol. 6,
pp. 71-93, 1991.
[48] D. G. Bounds, P. J. Lloyd, and B. G. Mathew, "A comparison of neural network and
other pattern recognition approaches to the diagnosis of .low back disorders," Neural
Networks, vol 3, pp. 583-591, 1990.
[49] R. H. Silverman and A. S. Noetzel, "Image processing and pattern recognition in
ultrasonograms by backpropagation, Neural Networks, vol. 3, pp. 593-603, 1990.
[50] H. Endo and H. Isshiki, "Application of "Associatron" to expert system," in Pro-
ceedings of the 1990 International Conference on Fuzzy Logic and Neural Networks,
lizuka, (Japan), pp. 751-754, 1990.
[51] A. F. Rocha, "K-Neural nets and expert reasoning," in Proceedings of the 1990 Inter-
national Conference on Fuzzy Logic and Neural Networks, lizuka, (Japan), pp. 143-
146, 1990.
[52] A. F. Rocha, M. Theoto, and P. Torasso, "Heuristic learning expert systems - gen-
eral principles," in Fuzzy Logic in Knowledge-Based Systems, Decision and Control,
(M. M. Gupta and T. Yamakawa, eds.), pp. 289-306, North-Holland: Elsevier Science
Publishers B. V., 1988.
[53] R. J. Machado and A. F. Rocha, "A hybrid architecture for connectionist expert
systems," in Intelligent Hybrid Systems, (A. Kandel and G. Langholz, eds.), CRC
Press, 1992.
[54] Y. Hayashi, J. J. Buckley, and E. Czogala, "Approximation between fuzzy expert
systems and neurcJ networks," in Proceedings of 2nd International Conference on
Fuzzy Logic and Neural Networks, lizuka, (Japan), pp. 135-139, 1992.
[55] H. F. Yin and P. Liang, "A connectionist incremental expert system combining produc-
tion systems and associative memory," International Journal of Pattern Recognition
and Artificial Intelligence, vol. 5, pp. 523-544, 1991.
[56] K. Yoshida, Y. Hayashi, A. Imura, and N. Shimada, "Fuzzy neural expert system for
diagnosmg hepatobiliary disorders," in Proceedings of the 1990 International Confer-
ence on Fuzzy Logic and Neural Networks, lizuka, (Japan), pp. 539-543, 1990.
[57] Y. Hayashi, "A neural expert system with automated extraction of fuzzy if-then rules
and its application to medical diagnosis," in Advances in Neural Information Process-
ing Systems, (R. P. Lippmann, J. E. Moody, and D. S. Touretzky, eds.), pp. 578-584,
Los Altos: Morgan Kaufmann, 1991.
[58] Y. Hayashi, "A neural expert system using fuzzy teaching input," in Proceedings of
Ist IEEE Conference on Fuzzy Systems, (San Diego), pp. 485-491, 1992.
This page intentionally blank
143
[63] E. Sanchez, "Fuzzy connectionist expert systems," in Proceedings of the 1990 Inter-
national Conference on Fuzzy Logic and Neural Networks, lizuka, (Japan), pp. 31-35,
1990.
[64] K. Saito and R. Nakano, "Medical diagnostic expert system based on PDF model," in
Proceedings of IEEE International Conference on Neural Networks, (SanDiego, Cali-
fornia), pp. I.255-I.262, 1988.
We start with a brief description of the system operation and then establish its key features
in the form of four theorems which are fully proven. The capabilities of the system are
illustrated with the aid of a rich set of simulation results derived for the case of trapezoidal
fiizzy sets. These results are shown to be better than the ones obtained with the original K-
Y-T systems.
Knowledge
Structured Unstructured
Symbolic AI Expert
Systems
Framework
Fuzzy Neural
Numerical Systems Systems
Each neural net is known to consist of a set of elementary units with n synapses and
weights wjj which are connected to a node f described by a strongly nonlinear operation.
These units are refered to as "neurons" (Fig.2).
Wii
OUTPUT-
x } INPUTS
The purpose of any learning procedure is to update weights wjj such that for each input for
the neural network, the output to be the desired one. Actually, a plethora of neural network
models and learning algorithms have been proposed in the technical literature (see e.g. [6-
11 ]). Of course there exist many situations where the weights are selected right from the
beginning by the designer without the application of any learning algorithms. The system
presented here belongs to this category.
147
The fijzzy systetns were intoduced by Zadeh in 1965 [12] and have shown a
significant development due to their ability for solving complex reasoning problems. They
are based on fuzzy logic and fuzzy sets. The fundamental operations of fuzzy sets are:
fuzzy complement, fuzzy union and fijzzy intersection [13]. With the aid of these
operations, the extension of logical tautologies becomes possible.
A particular tautology which is directly useful in fuzzy controUers is the modxis
ponens (MP) rule:
[{A -^B)AA)^B
which indicates that in a cause-effect relationship, when the cause is satisfied the
consequence is obtained. The question is "what happens when A and B are fuzzy sets?". In
this case the above MP rule is extended to the following generalized modus ponens (GMP)
rule:
{{A ->B)A A') -> B'
which shows that when the fiizzy cause is ahnost satisfied we nearly get the consequence.
One way to implement the GMP rule is through the use of Zadeh's max-min composition
operator [12,18].
A fuzzy controller consists of the following parts (Fig.3) [6,14,18]:
- fuzzification
- fuzzy inference mechanism
- defiizzification
-0-- Process
n
Defuzzificatlon
J Fuzzy Inference L-' Fuzzlftcatton
n
Mechanism J
Fuzzy Controller
Fig.3: General structure of a fiizzy coBtroller.
The fuzzy inference mechanism is implemented with the aid of a fiizzv knowledge base, i.e.
a knowledge base that contains fuzzy rules (MIMO: Multi-fnput/Multi-Output rules), ft is
148
From the discussion of the previous section it is clear that there are some basic
similarities between neural and fuzzy systems. Our purpose in the present section is to try
to use a neural net for drawing fuzzy conclusions. To this end we shall extend and improve
the solution proposed by Keller, Yager and Taham [15]. This sohjtion has the drawback
that, although it possesses a correct behaviour, it is away from the desired solution,
especially when the fact (event) occurs. Thus an improvement will be proposed here
together with a discussion on the application domain of the network.
Actually, one of the main consequences of the use of fuzzy reasoning in a rule based
(IF-THEN) system is the dramatic reduction of the number of rules which leads, among
others, to faster conclusions. Some of the available solutions for the organization of fuzzy
controllers are those proposed by Procy and Mamdani [20], Braae and Rutherford [21] and
Macvicar-Whelan [22], where the fiizzy matrix concept is utilized for the representation of
the fuzzy control relationships of concern. These solutions give very good results but
possess the following drawbacks :
i) They require a large memory space for the storage of the elements of the fuzzy
matrices (but they gain in comutation time).
ii) The fuzzification unit must produce results that are compatible with the
representation of the variables in the fuzzy matrices. For example if the input is
quntized as "LOW, MORL LOW, MEDIUM, MORL HIGH, HIGH", the output of
the fuzzifier must give one of the values "LOW, MORL LOW, MEDIUM, MORL
HIGH, HIGH". This drawback is weakened when the quntization step is reduced,
but then again the required memory increases.
149
In our system, no ftizzy matrix is used, but instead a sequence of fuzzy operations
(fuzzy algorithm). In this way the extra memory for the fuzzy matrix is no more required,
and resulting increase of the computational time is faced by employing a neural network
structure that has a high degree of parallelism. An additional benefit of the method
presented here is the "continuity" in the sense that the system is able to respond without any
particular care, to any input value that may not belong to the value set {LOW, MORL
LOW, MEDIUM, MORL HIGH, HIGH}.
The proposed neuro-fuzzy network must satisfy the following requirements:
a. The fiizzy sets must be given in sampled form, since every neural network has a finite
number of inputs, i.e. Aj and B in the rule (1) must be of the form:
b . In case the sets Aj are crisp sets the network must produce results that agree with the
rules of classical logic.
In the followmg, it is proved that every neural network of the type shown in Fig.4 with
describing equations
c/fr =u{l{a'ij,w,j)]
)
S = C(u(</fc))
k
b; = c{nu,,s)}
150
where
^1 ^m, ^1 Kn\,
Theorem 1
UA; - A, for all / = 1,2,.... n, then the result is B.
Proof
From Aj = A,' it follows that a^ = a]j for all / = 1,2,..., n and so for all v,j in the domain
of discourse it is true that tiW^) = a^ = 1 or 0 and //(c(i/,^)) = 0 or 1, respectively, where
151
p{x) is the membership function of concern. That is V/ = l,2,...,n the following relation
holds:
d^ = u{na,j,c{aij))} = 0
j
s = c{u(rfk)} = c(0) = l
k
and so b- = c{i(Uj,s)} = c(u,) = c(c(fa,)) = 6, for / = 1,2 n. This means that the result
is B as required, q.e.d.
Theorem 2
If for some i from 1 to n the condition A^ nAI^O holds, then the neural network
produces the result "y is UNKNOWN" (which means that £>/ = 1 V/ = 1,2,..., n).
Proof
Since fJ cWjj) = f^A'^^ij) = 1 ^or some vij in the domain of discourse and /(a,y,M',y) = 1 for
this Vfj,
du=u{na;j,wy)} =\
Theorem 3
If Al Q Af (i.e. Al is more specific than A/- ft A. ^ PA) for some i from 1 to n, then the
result is B.
Proof
From Al c A, for some i, it follows that ^ nA^ = 0 . Hence d/, = 0 which gives the
result "Y is B". q.e.b.
This network is used to solve problems similar to the following example [15]. Consider the
rule
IF a person is OLD THEN the person in UNHEALTHY
152
If some time we learn that the person is VERY OLD, then this information is sufficient to
say that the person in UNHEALTHY. To confirm this fact the following theorem will be
shown.
Theorem 4
Consider two inputs Aj and /I" / = 1,2 n of the neural network and suppose that
A " £ A' £ A ^°^ all / = l,2,...,n. Then the corresponding results B' and B" will satisfy
the condition
B c B" c B'
Proof
From the relation /V'e A G A . V/ = 1,2,...,n, it follows that
0 < /i^.(-) < p^.(-) < fiA, (-). V/ = 1,2,...,n
for any point in the domain of discourse.
Thus
and so if c/j^ < dji then u((^it) - U(<^A^) which implies that
k k
c{uK)}<c{u(4')}
k k
or
s' < s"
Thus we have 6/ > 6," i.e. B c B" ^ B' because 6/,6;'> 6,, V/ = 1,2 n where
6/ = c{/(y,,:j)} = c{/(c(6,-),s)} > c{c(6,)} = ft,
Remark
For any input Aj it will be true that B c B'. It is therefore clear that for any triple of
functions (c, u, i) the network will be working for examples of Che "unhealthy old person"
type mentioned above.
The problem now is to select the functions u(-) and i(-,-) that satisfy the above
properties. Actually, many functions are appropriate for this purpose [13]. A first selection
is the one of Keller, Yager and Tahani [15]. Looking carefully at the neural net of Fig.4,
one concludes that the dominant role is played by the second player where the d^'s are
153
computed. The core of this operation is the comparison of the input with the prototype that
is represented by the weights Wjj, i.e. the event i of the GMP rule. To achieve this, the
complement of this event is stored in the Wy's and the maximum element of each
intersection with the input is obtained. The function of this intersection must be as "hard" as
possible, so as to give a value as near to zero as possible if there exists a relation of the
input and the event. Otherwise it must give the value 1, which means that a decision cannot
be obtained and is thus left to another GMP rule.
Most of the fuzzy function classes for the intersection, give the possibility for such a
"hard" intersection under a suitable choise of the values of their parameters. The final
selection of the intersection fimction among them, is made on the bases of the validity of
the conditions mentioned above and used in the proofs of theorems I through 4, and the
computation time required to execute the mathematical operations that support it.
In our system the following neuron model is selected in order to possess all desired
features:
d^ - max.[iH(3ij,Wij))
J
s = 1- max {dif}
k
u, = l-b,
bj = 1 - s • Uj
where //y(-) is Hamacher's ftmction [13]:
dff = Ti\av.[a'fj,w^)
s = Ci(max(</t))
u, = C2(bi)
bi = c-^(sui)
where Cj and C2 are complementation functions of the Sugeno type [13]:
C/i(a) = — , ?i e ( - l , Q o )
\ + Aa
In practice one must select suitable values for \ and /Ig - This choise can be made through
simulation as described below.
155
Table 1: The weaning of linguistic terms defined on the domain [1,11] and sampled at
integer points.
3. Simulation example
The proposed neural networks were simulated for trapezoidal fuzzy sets [17] according to
the rule:
IF X IS LOW THEN Y IS HIGH
and the rule:
IF X, IS LOW AND Xj IS LOW THEN Y IS HIGH.
The results obtained are shown in Tables 2 through 8. We observe that they approach the
desired result, the first without deviation and the second with some tolerable deviation.
Table 2: The results oftite simulation of the first model using Hamacber's function (y=50)
Table 3: The results of the simulation of the first model using Hamacher's function
(Y=100)
Label
LOW
Values of the membership function 1
0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.60 0.80 tool
MORL LOW 0.03 0.03 0.03 0.03 0.03 0.03 0.22 0.42 0.61 0.81 i.ool
VERY LOW 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 i.ool
MEDIUM 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 t.OO I.ool
NOT LOW 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 I.ool
Table 4: The results of the simulatioa of die first model using Hamacher's function
(r=200)
Label Va ues of the membership function |
LOW 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 I.ool
MORL LOW 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.61 0.80 I.ool
VERY LOW 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.40 0.60 0.80 1.001
MEDIUM 1.00 t.OO 1.00 l.OO l.OO l.OO l.OO 1.00 1.00 1.00 I.ool
NOT LOW 1.00 1.00 1.00 1.00 l.OO 1.00 1.00 l.OO l.OO l.OO I.ool
"able 5: The results of the simuladon of die second model using Sugeno's function
Label Values of the membership function |
LOW 0.05 0.05 0.05 0.05 0.05 1 0.05 0.24 0.42 0.61 0.80 I.ool
MORL LOW 0.09 0.09 0.09 0.09 0.09 0.09 0.30 0.49 0.67 0.84 I.ool
VERY LOW 0.02 0.02 0.02 0.02 0.02 0.02 0.18 0.35 0.54 0.76 I.ool
MEDIUM 1.00 1.00 l.OO 1.00 t.OO 1.00 t.OO 1.00 1.00 too I.ool
NOT LOW l.OO 1.00 l.OO 1.00 1.00 1 1.00 t.OO 1.00 t.OO 1.00 tool
Table 6: The results of the simulation of the first model using Hamacher's function (y=50.
two inputs/one output)
Label Values of the membership function
LOW. 0.02 0.02 0.02 0.02 0.02 0.02 0.21 0.41 0.61 0.80 1.00
LOW
MORL LOW, 0.05 0.05 0.05 0.05 0.05 0.05 0.24 0.43 0.62 0.81 1.00
LOW
MORL LOW, 0.05 0.05 0.05 0.05 0.05 0.05 0.24 0.43 0.62 0.81 1.00
MORL LOW
VERY LOW, 0.02 0.02 0.02 0.02 0.02 0.02 0.21 0.41 0.61 0.80 1.00
LOW
VERY LOW. 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.40 0.60 0.80 1.00
VERY LOW
VERY LOW, 0.05 0.05 0.05 0.05 0.05 0.05 0.24 0.43 0.62 0.81 1.00
MORL LOW
VERY LOW, 1.00 l.OO 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
MEDIUM
157
Table 7: The results of the simulation of the first model using Hamacher's function (y=IOO.
two inputs/one output)
Label Values of the membership function
LOW. 0.01 0.01 0.01 O.OI 0.01 0.01 0.21 0.41 0.60 0.80 1.00
LOW
MORL LOW. 0.03 0.03 0.03 0.03 0.03 0.03 0.22 0.42 0.61 0.81 1.00
LOW
MORL LOW, 0.03 0.03 0.03 0.03 0.03 0.03 0.22 0.42 0.61 0.81 1.00
MORL LOW
VERY LOW. 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.60 0.80 1.00
LOW
VERY LOW, 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00
VERY LOW
VERY LOW, 0.03 0.03 0.03 0.03 0.03 0.03 0.22 0.42 0.61 0.81 1.00
MORL LOW
VERY LOW. 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1,00 1.00 1.00
MEDIUM
Table 8: The results of the simulation of the first model using Hamacher's function (y=200,
two inputs/one output)
Label Va ues of the members!lip function
LOW, 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00
LOW
MORL LOW, 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.61 0.80 1.00
LOW
MORL LOW, 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.61 0.80 1.00
MORL LOW
VERY LOW, 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00
LOW
VERY LOW. 0.00 0.00 0.00 0.00 0.00 0.00 0.20 0.40 0.60 0.80 1.00
VERY LOW
VERY LOW, 0.01 0.01 0.01 0.01 0.01 0.01 0.21 0.41 0.61 0.80 1.00
MORL LOW
VERY LOW, 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
MEDIUM
In the first case (Tables 2-4) we present the results of the network for the Hamacher
ftmction with y=50, 100 and 200, respectively (the plots are omitted since they almost
coincide with the desired results). In the second case (Table 5) we use the complement
function of Sugeno with /Ij = 10 and /Ig = 6 and give the corresponding graphical
representations. In the third case (Tables 6-8), we present the results of the two inputs/one
output network for the Hamacher function with Y=50, 100 and 200 respectively.
158
1
0,9 Output of low (K-
0J8 Y-T)
0,7 • - — • Output of very
0,6 tow (K-Y-T)
0,5 • Output of morl
low (K-Y-T)
0.4
Higfi (K-Y-T)
0,3
Output of morl
0,2 low (1st model
0,1 V=200)
0 1 I
2 3 S 6 7 8 9 10 U
Fig.5: Comparison between the first model (y=200) and the K-Y-T
model (the outputs of hw and very low of the first model are almost
identical to High)
4. Conclusions
The neural network proposed here has the same structure as the K-Y-T model but it is
generalized such that one can use any meaningful fitzzy intersection, union or complement
ftmction. This is reflected in the more generality of Theorems 1 through 4. Here we
selected and used the Hamacher intersection function which leads to moie accurate
inferences than those obtained by the K-Y-T network. Also we used the Sugeno
complement function which gave us an other improvement A further extension would be
obtained if one uses rules of the type: "IF water is WARM THEN open COOL tap", "IF
water is VERY WARM THEN open VERY COOL tap", and so on (as mentioned at the
end of section 2).
The authors are currently working to design a system that can accomodate such rules
and also in the area of tuning PID and otilier controllers with die aid of the present inference
system in place of the so called "fuzzy matrices" [19]. Hopefully the use of the present
system will give equivalent tuning results as in [19] but more rapidly and with increased
robustness. Work is also in progress for enriching the structure with a learning capability.
Acknowledgement
This chapter is based on a paper presented at the AIENG 93: Applications of Artificial
Intelligence in Engineering (Toulouse, June-July 1993).
159
1 y
0.9
0.8 .' /
• ' /
0.7 +
0.6 Mori high
03 Output of morl
0,4 • bw (2nd mod«Q
0.3..
0.2-
oa^. . J-
0
-i- 8
-*-
9 10 11
H
0.9-
03
0.7
0,6 Very high
0,5 Output of very
0,4 tow (2nd modeO
0^1
0,2
oa
0 - - t - T - r - I — I - if • ' i -+- -1- -H
2 3 5 6 7 8 9 10 U
1
0,9
0,8
0,7-
0,6
High
0^..
• — — Output of low
0,4- (2nd model)
0,3
o;?
0 -i 1 1 »- ~r' I -i H
2 3 4 5 6 7 8 9 10 H
F>fl6: Comparison between the output of tt)e second model and the desired one.
160
References
[14] Fllor Van Der Rhee, Hans R. Van Nauta Lemke, Jaap G. Dijkman, 'Knowledge Based
Fuzzy Control of Systems', IEEE Trans, on Automatic Control, Vol. 35, NO. 2, February
1990.
[15] J. M. Keller, R. R, Yager, H. Tahani, 'Neural Network Implementation of Fuzzy
Logic', Fuzzy Sets and Systems 45 (1992), pp. 1-12, North-Holland.
[16] C. C. Lee, "Fuzzy Logic in Control Systems: Fuzzy Logic Controllers, Part I, 11', IEEE
Trans, on Systems, Man and Cybernetics, Vol. SMC-20, No. 2, March/April 1990.
[17] M. Mizumoto, S. Fukami and K. Tanaka, 'Some methods of Fuzzy Reasoning', in: M.
M. Gupta, R. K. Ragade and R. R. Yager, Eds., Advances in Fuzzy Set Theory and
/4/ip7/ca/«7i2s (North Holland, Amsterdam, 1979) pp. 117-126,
[18] S. G. Tzafestas, 'Fuzzy Expert Control: Recent Results with Applications to Robotic
Systems', Advanced Study Institute oa knowledge Based Robot Cootrol, France, 1988.
[19] S- Tzafestas and N. P. Papanikolopoulos, 'Incremental Fuzzy Expert PID Control",
IEEE Trans, on Industrial Electronics, Vol. 37, NO. 5, October 1990.
[20] T. J. Procyk and E. H. Mamdani, 'A linguistic self-organizing process controller,'
Automatica, vol. 15, pp. 15-30, 1979.
[21] M. Braae and D. A. Rutherford, 'Theoretical and linguistic aspects of the fuzzy logic
controller,' Automatica, vol. 15, pp. 553-577, 1979.
[22] P. J. Macvicar-Whelan, "Fuzzy sets for man-machine interaction,' Int. J. Man-Mach.
Studies, vol. 8, pp. 687-697, 1976.
This page intentionally blank
PART 3
FUZZY CONTROLLERS
CHAPTER 8
Control Laboratory
Department of Electrical Engineering
Delft University of Technology
P.O. Box 5031, 2600 GA Delft, The Netherlands
1 Introduction
Fuzzy control has become a popular topic in t h e control engineering community.
T h e proliferation of fuzzy control m e t h o d s get really underway by t h e introduction
of fuzzy control in consumer products by t h e Japanese electronic industry. Also,
applications are reported in the process industry (Aoki et al, 1990; Yu et al, 1990)
of which the best known are in the cement industry (Haspel, 1991; Ostergaard,
1990; Flintham, 1991) Many instrumentation firms offer fuzzy hard- and software
as implementation tools for fuzzy control, like for example, Siemens, Yokogawa,
Inform, Togai, etc.
Most control algorithms, however, axe based on Hnear control theory and too much
relying on this approach leads to a dead end for their use in modern production
methods. Moreover, it is clear that the present state of t h e analysis and synthesis of
non-lineax controllers based on non-Iineax system theory is still i m m a t u r e and does
not provide the solutions needed in industry.
So, t h e control engineering community was waiting for new methods which could
alleviate the problems which have to be solved in process industry.
Fuzzy control algorithms could be very helpful, because of t h e following two reasons.
• They allow the use of linguistic expressions to describe the behaviour of the
system or the control actions. Using this property it is possible to i m i t a t e the
actions of operators.
• They are inherently non-linear and are thus able to perform control actions,
which are not possible with a purely linear approach.
These properties alow fuzzy control to be brought into action in many control prob-
lems. However, some people think t h a t fuzzy control is t h e panacea for everything
t h a t is difficult to control. It is true t h a t some authors and advocates of fuzzy con-
trol have stimulated this feeling by only indicating the advantages of fuzzy control
and mystifying the real problems arising when applying fuzzy control.
In this contribution we want to show the real face of fuzzy control, showing the
strength of this m e t h o d in comparison to other methods, but also the weak points
and unsolved problems.
The first three sections of our contribution are concerned with the basic parts of a
fuzzy controller as shown in figure 2. First the fuzzification of the inputs of the fuzzy
167
FS
r u 1 1 y y
FC process
(a) (b)
Fuzzy controller (FC). Fuzzy supervisor (FS).
controller will be described, dealing with the choice, number and (relative) position,
of membership functions. T h e second section will discuss the inference part of fuzzy
controllers. T h e a i m of section 4 is t h e comparison a n d evailuation of defuzzification
methods. Next the relation between fuzzy control and lineaj control is described.
T h e last section will be used for conclusions and discussion.
rule
base
n
X rule of u^ 'S
inference
-a
2 Fuzzification
n u m b e r and shapes of those membership functions a n d how they cire related to Ccich
other, as well as the number of overlapping membership functions, determine the
resulting controller o u t p u t as function of the controller inputs of the fuzzy controller.
In this section a centre-of-gravity oriented defuzzification (see section 4) method is
assumed.
T h e first three subsections will deal with t h e fuzzification of the inputs. T h e last
subsection will address "fuzzification" of the universe of discourse of the o u t p u t .
W h e n designing a fuzzy controller one of the first questions which arise, after having
chosen the inputs, is how many membership functions are needed eind how the
membership functions should be divided on the universe of discourse of the inputs.
(a)
Linear division of membership functions on universe of discourse.
(b)
Non-linear division of membership functions on universe of discourse.
cire shown. Whether to use a linear (3a), a non-linear (3b) or an other set of mem-
bership functions on the universe of discourse of an input depends on the problem
to deal with.
It can be shown that having more t h a n two overlapping membership functions will
lead to a sort of "smoothening" of the control hypersurface, t h a t would result from
having only two overlapping membership functions.
(a) (b)
Smoothing of controller output u{x) Influence of changing conclusion of
as function of controller input u. one fuzzy mle controller output.
Because of this and the interpolative character of a fuzzy controller, a chosen set of
membership functions do not limit the resolution of t h e controller o u t p u t , assuming
a centre-of-gravity oriented defuzzification method (see also section 4). T h e number
and positions of the membership functions on the universe of discourse of the output,
in combination with t h e fuzzy rules and the universes of discourse of t h e controller
inputs, should be chosen according to the control hypersurface desired.
3.1 Connectives
For implementation of the and connective a T-norm (Schweizer & Sklar, 1963)
is used. T h e conditions which a T-norm, for applying intersection or conjunction,
must satisfy are:
T-1 T(a,l) = a
T-2 T{a, b) < T ( c , d), whenever a < c, b < d
T-3 T{a,b) = T{b,a)
T-4 T{T(a, b), c) = T(a, T(b, c))
ab
fd&p,e(a,<> = , , .. (la
max(a,b,(fj
with 6(z [0,1]. When 6 equals 0, <<i&p,e acts as 7Tim(a,6). In case 6 equals 1, tdicp,e acts
as ab from probability theory. T h e or, else and also connectives are implemented
using a T-conorm or S-norm. T h e conditions defining a T-conorm, for applying a
union or disjunction, are, besides T-2, T-3 and T-4;
S-I S(a,0) = a
A S-norm will be denoted in this paper by \l. Examples of S-norms are: max(a,b),
min[a-\-b, 1), o-|-6 — ab, etc. As with the T-norm there exist a lot more possibilities.
173
often combined in a family of T-norms and -conorms. For example, the S-norm
related to ( l a ) is defined as:
a + 6 — a6 — m w ( a , 6 , 1 — ^) , ,
max{\ — a,l — 0,6)
This S-norm td&p,e acts as max(a, b) when 6 equals 0 and acts as a + b — ab from
probability theory, when 6 equals 1. For more detailed information on T-norms emd
-conorms we refer to Klir and Folger (1988), Buckley and Siller (1987) and Turksen
(1986).
Although not further used in this paper, but for completeness, t h e minimal con-
ditions for the complement, used for implementation of the connective not, are
(Hellendoorn, 1990):
C-1 C(0) = 1
C-2 C{a) > C(b), whenever a <b
C-3 C(C{a)) = a
For t h e complement function normally 1 — a is used. The next section will use the
T-norm and -conorm in the compositional rule of inference.
In fuzzy control rules are used to represent the knowledge implemented in the con-
troller. In this section we will not discuss the type of knowledge the fuzzy rules
represent, like state evaluation or object evaluation fuzzy rules (Lee, 1990b), be-
cause this reflects more the application of fuzzy rules than t h e form. Here we will
address only the form of fuzzy rules and one can mainly distinguish two types:
rules with "symbolic" consequents and rules with "numerical" consequents, from
now on referred to as Mamdani- and Sugeno-rules, respectively, eind discussed in
t h e following subsections.
This type of rules was used in the first applications of fuzzy control (Assilian, 1974;
Mamdani, 1974; Mamdani & Assilian, 1975) and has the following general form:
174
with:
R^ = {Xix...xXi,JxU^ (2a)
Using the Mamdani-types of rules in a fuzzy controller will result in a fuzzy controller
o u t p u t as result of the fuzzy inference, which has to be defuzzified to obtain a
numerical controller output. Defuzzification methods are addressed in section 4.
3.2.2 S u g e n o - t y p e o f f u z z y rules
T h e general form of the other rule type, referred t o as Sugeno-rules, because of the
introduction of those rules by TaJcagi and Sugeno (1983) and further exploited by
Sugeno and co-workers, is as follows;
with constant parameters af and ¥ and which can b e even more simplified when
t h e parameters af are 0, resulting in a constant numerical consequent 6^'. Successful
use of this t y p e of fuzzy rule in t h e control of a model car was reported by Siigeno
and Nishida (1985) and Sugeno and Murakami (1985). T h e "fuzzy" o u t p u t of a
fuzzy controller using this t y p e of fuzzy rules is in fact a set of weighted numerical
o u t p u t s . A simple defuzzification results in the final controller o u t p u t .
Although one can distinguish the Mamdani-rules from the Sugeno-rules, in practical
set-ups the Mamdani-type of rules are quite often similar to Sngeno-mles with con-
stant consequent, for example, see Batur and Kasparian (1991), Harris and Moore
(1989) and Matsuoka (1991). This is due to t h e defuzzifi cation m e t h o d used to
defuzzify the fuzzy output of t h e controller: fuzzy-mean or centroid. See section 4
for details on defuzzification methods.
176
R=\:)Ri (3a)
Nr
fJ-R{xi,••• ,XN,,u) = ^ fiRjixi,--- ,XN,,u) (3b)
3=1
U^X,o{X2o{---iXr,,oR)...)) (4a)
U = Xo R (4b)
where X is the cartesiaji product of the fuzzy inputs Xi and using fuzzy sets:
with:
AT.
where one should note t h a t here Xi(t) is used instead of Xi as in (5). This means
t h a t the left T-norm in (6) operates on membership values and not on membership
functions as t h e right T-norm does.
In figure 6 this practical way of fuzzy reasoning is shown schematically, where the
rnin-operator is used as T-norm. Combination of those individual fuzzy outputs by
applying a union results in the complete fuzzy o u t p u t of the system:
T h e most commonly used inference rule is the maa;-mm-rule (the one used in the
examples in figure 6 and 7). Also frequently used is the max-prod type of inference.
Fuzzy inference using non-lineax operators (others thaji the product and summation
operator for T-norm and -conorm, respectivily) results in a non-linear control hy-
persurface. Even when the fuzzy rules form a consistent, complete and "linear" rule
base, the control hypersurface will b e non-linear. This would not be the case using
t h e sum-prod-rule of inference, consisting of "linear" operators. T h e non-lineaxity
of the control hypersurface is completely determined by the logical operators and
not by the fuzzy rules, which, in our opinion, should define t h e non-linearity of the
control hypersurface. Section 5.2 will show this in more detail in case of a fuzzy
PID-controller.
Several defuzzification methods are available to translate the fuzzy output of the
controller into a crisp, numerical output value. Defuzzification will be discussed in
t h e next section.
178
(a)
If xi is Ai and 12 is B2 then u is Ci.
1.0
J''
0.5 ^1 y G2
4 Defuzziflcation
The defuzzifier of a fuzzy controller can deteriorate a great deal the results achieved
by the inference of fuzzy rules. It appears that the application of a specific defuzzi-
fication method can reduce completely the fuzzy character of the controller, as will
be shown in subsection 4.2.
This section will describe the basic defuzzification methods and related methods.
Probably the best known method is the centre-of-area defuzzification method (coa),
180
coa{U) TOO
/
Hu(u)du^ I iJ,u{u)du (8a)
resulting in a numerical value u = u^oa which divides the integral of the fuzzy output
in two equal parts. This way of calculation is normally not applicable in practical
situations. T h e centre line u — Ucoa, which divides the area under the fuzzy output
function caji be approximated by calculating the centre of "gravity", which explains
the alternative name, centre-of-gravity (cog) method:
JV,
cog(C/) = i ^ i (8b)
where iV, is the number of quantisations used for the discretisation of the fuzzy
output ^i/(u).
k=\
where N^ is the number of fuzzy sets defined on the universe of discourse of the
controller o u t p u t , 7^ denote the support values for membership function Uu as result
of the inference within the controller and u^ represents the defuzzification of U^. The
support values 7fc are determined by the conjunction of t h e condition part of the
fuzzy rules and therefore the fuzzy-mean defuzzification method is very suitable
in case the calculation load of the fuzzy controller should be small (see subsection
3.3). In figure 6 the support values are t h e values by which the fuzzy outputs Ci
and C2 are limited (in case of m i n i m u m as T-norm for the implication). These
values are calculated in figure 6 by taking t h e m i n i m u m of the membership values
in the premise of the fuzzy rules. Because only crisp representations are used of
181
In this subsection we will take a closer look at the several defuzzification methods
as described in the previous subsection. Comparisons will be made, advantages
and disadvantages will be discussed. First t h e mean-of-maxima method is evalu-
ated; after t h a t a comparison is made with t h e centre-of-area related defuzzification
methods.
182
£ind it is assumed t h a t the membership values do not have values equal to ~, then
there exists only one dominant fuzzy rule in the rule base. This domincint rule
will completely determine the numerical output of the fuzzy controller in case the
chosen S-norm for combining t h e individual fuzzy conclusions of the fuzzy rules
(union) satisfies besides the conditions S-1, S-3 and S-4 a stronger condition for S-2,
namely:
In case we consider also the cases in which more t h a n one fuzzy rule are equally
dominant (two or more membership values are equal to i ) , t h a n the (finite) number
of possible numerical values for the o u t p u t is still restricted by:
zJ r,
n I ""^
„fj ;n\{N^
:T?^ _
- ^n)\
M (^^
r
1
/ \ / \:
0 y V 0 \ *• 0 > 0/ \
(a) (b)
A-function. U-function. Triangular,
m
Trapezoidal.
/*£/(")
(«)
/H/(ti)
(b)
it
0 200 400 600 800 lOOO 1200 1400 1600 1800 2000
tune(s)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
time (s)
(a)
Mean-of-maxima defuzzification.
a I
^ Oh
0 200 400 600 800 1000 1200 1400 1600 1800 2000
time (s)
0 200 400 600 800 lOOO 1200 1400 1600 1800 20O0
time (s)
(b)
Centre-of-gravity defuzzification.
method *,/ + ,-
cog Ng + l 2 K - 1)
fm Nu + 1 2(^„ - 1)
wfm SAT,+ 1 2(iV„ - 1)
Comparing the defuzzification methods, it appears that the choice of the defuzzifi-
cation m e t h o d determines to a W g e extend the "quality" of control as well as t h e
computational cost of the controller. Comparing the results based on control and
computational performance, one can choose the appropriate defuzzification m e t h o d
for a specific application.
if e is AZ and Ae is AZ then u is AZ
Knowing t h e linear relation between the number of quantisations Ng used for the
centre-of-gravity related defuzzification methods (see table 1) and the neglectable
difference in the resulting control performance (Jager et al., 1992), one can conclude
that t h e (weighted-)fuzzy-mean method is more preferable t h a n the centre-of-gravity
m e t h o d , which uses more quantisations in its calculation. Using this (weighted-
)fuzzy-mean defuzzification method in fact transforms a fuzzy controller according
to Mamdani (Mamdani & Assilian, 1975) into a Sugeno-type (Takagi & Sugeno,
1983) of fuzzy controller (see section 3.2 for description of Mamdani- and Sugeno-
type of fuzzy controllers), some examples can b e found in B a t u r and Kasparian
(1991), Harris and Moore (1989) and Matsuoka (1991).
186
In this section we will discuss the relation between fuzzy and linear control. It
is shown that any linear controller can be described as a fuzzy controller: fuzzy
control can be seen as a superset of linear control or linear control as a subset
of fuzzy control. The most widely used controller is the PID-controller £ind most
fuzzy controllers described in literature imitate a conventional PID-controller. The
second subsection will address these fuzzy PID-controllers. Finally, fuzzy control of
non-linear processes is discussed.
When designing a fuzzy controller and applying specific choices for membership
functions, logical operators and scaling of in- and outputs, the fuzzy controller can
emulate a linear controller. From this point of view linear control can be seen as
a subset of fuzzy control. First let us start with looking at a fuzzy controller as a
controller which represents a mapping / ( ) from the N^ inputs x to the output u:
The fuzzy controller function (10) can emulate the linear controller (H) when meet-
ing the following criteria:
1. the membership functions on the universe of discourse of the inputs are trian-
gular shaped;
3. the membership functions form a complete set: the sum of the membership
functions on a universe of discourse equals 1;
5. the intersection operator for combining the premises of the fuzzy rules is the
product operator (probabilistic intersection);
In a lot of fuzzy controllers found in literature criteria 1 ^ axe met. Although not
explicitly stated they normally also meet criterium 8. This leaves criteria 5-7 to
be the main differences between "standard" fuzzy controllers and linear controllers.
Using t h e product a n d summation operators instead of the m a x and m i n operators
respectively, is necessary because t h e emulation of linear controllers requires linear
operators.
T h e most important criterium to be m e t is 6, because due to criteria 1-5 and 7-8 the
o u t p u t of a fuzzy controller, with N^ inputs results in a linear interpolation between
2 ' points in a iVj.-diniensional space and the numerical conclusions of the at most
2 ' fuzzy rules, which contribute to the output, determine wether or not a linear
relation, by connecting these points, (hyperplane) exists. When this hyperplane
exists it equation (11) holds.
We can write this more formally by stating that the controller o u t p u t u is described
by applying the fuzzy-mean defuzzification, as defined in (8c)) (subsection 4.1), and
criterium 2, as listed above, is met:
^='-% (12a)
k=\
2Nx
(12b)
(13a)
3=1 t=l
This is also the reason why in (12a) we use 2^" as the number of active rules. T h e
number of active rules is in fact greater t h a n 0 and equal or less than 2^"^. Assuming
some non-active rules to be active but not contributing to the output, (12a) is correct
when criteria 2-5 and 7 are met.
When (11) describes the control hyperplane: the controller o u t p u t as linear function
of the controller inputs and the numerical result of each individual fuzzy rules lies
on this hyperplcine, the controller output from (12a) can be described by:
(14a)
j=i
(14b)
J=l
189
(14c)
u = a x[t) + b (15)
N^
( M I I ^ I I + IJ-12X12) Ylifikl + fJ-k2)
k=2
JVi-1
[flN^lXN^l + fJ.N^2XN^2) J J (MW +/^fc2)
fc=l
t^N^l^Nzl + f^N^2XNz2
In case of Sugeno-type of rules there is a very simple proof: all rules should have
(11) as consequent and the fuzzy-mean defuzzification methods should be used. This
result in a controller o u t p u t , which equals (11):
Nr
(17a)
E7.
3=1
190
A',
= '^^^. (17b)
= a^x + 6 (17c)
Although PID-controUers are known to be linear controllers, they are not in practical
implementations. First of all t h e control signal is limited. Secondly anti-windup
mechanisms axe used in practical set-ups.
if e is P M and Ae is NS then u is P B
In table 2 several families of fuzzy rules and their corresponding linear controller
type are given. Using derivatives of the error instead of previous values resembles
more closely the way humans use error signals to clcissify the deviation between the
process state and the reference signal. Applications of fuzzy controllers which use
the error and its derivatives as inputs emd satisfy conditions 1-4, 6 and 8 as stated
in the previous subsection (5.1), are in fact unnecessary approximations of linear
controllers. T h e only difference between such an application and the pure emula-
tion of the linear equivalent is t h a t the moas-min inference rule is applied. Due to
the use of these non-linear operators t h e control hypersurface will be a non-linear in-
terpolation between the numerical results of t h e individual fuzzy rules, defining the
characteristics of the control hypersurface. This non-linecir interpolation, however,
is a non-trivial and it is more trivial to perform the most objective interpolation be-
tween these chju-acteristic points on the control hypersurface: a linear interpolation.
Figure 12 shows the comparison of a linear fuzzy PD-controUer and a "full" fuzzy
version of this controller. As can be seen in figure 12b the use of linear operators
will lead to a linear control hypersurface (not taking in account the limiting of sig-
nals). Figure 12a shows the case where the non-linear operators were used. This will
191
\ f
(a) (b)
Non-linear opermtors (miuimum and Linear operators (product and sum-
maximum,). m,atioii).
result in "waved" version of the control hypersurf£ice defined by linear case. Figure
12 shows also another interesting detail of the control surface. Because t h e limit-
ing of the controller o u t p u t , to prevent wind-up, is explicitly embedded in the rule
base, there exist non-linear characteristics due to t h e transition from the "linear"
paxt of the rule base to the "limiting" part. These non-linear characteristics can be
recognized in figure 12 by the b u m p s marked by arrows.
Using fuzzy controllers looks quit sophisticated, to use a fuzzy controller t o mimic
an operator. However, applying fuzzy controllers in this way ignores the possible
strength of fuzzy controllers: being able to define a non-linear controller using a fairly
stretightforward and easy to use concept of design. One should note t h a t a fuzzy
P ID-controller is more understandable to non-control engineers than a conventional
PID-controller. This is probably also a reason of the popularity of fuzzy control:
you do not have to be a control engineer to understand how it works or how to build
In the previous subsection it was stated t h a t with a fuzzy controller a more "optimal"
control can be obtained. This holds cdso in case of non-linear processes. Often fuzzy
193
T h e fuzzy controller is not an "optimal" average for the whole input space like a
PID-controller is, b u t an interpolation between more local "optimal" averages. This,
however, has nothing t o do with t h e non-linearity of the process: a fuzzy controller
can control a non-linear process as least as good as its corresponding linear controller,
just because a fuzzy controller cem control a linear process as least as good as its
corresponding linear controller.
more rules. W h e n we extend the example, as show above, to more "local" PI(D)-
controllers for specific situations a^s described in the condition p a r t of t h e fuzzy-
rules, one obtains the same effect as is obtained when using gain-scheduling. T h e
difference between the classic2d gain-scheduling and these type of fuzzy controllers is
t h a t the fuzzy aspect of the fuzzy controllers causes smooth transitions between the
control hyperplanes, each determined by a parameter set (and thus a fuzzy rule).
In the introduction it was stated t h a t the popularity of fuzzy control could be the
need for a control concept to build controllers t h a t cire able to deal with changes in
production rate, product mix and production procedures and recipes. Looking at
applications of fuzzy controllers nowadays, it seems t h a t very few applications for
solving these kind of problems have been reported in literature.
A major problem in the application of fuzzy control is the fact t h a t using fuzzy con-
trol to implement operator knowledge and experience will bring u p the knowledge-
engineering-bottleneck as major obstacle to the use of knowledge based systems
(Feigenbaum & McCorduck, 1983), and thus to the use of fuzzy expert systems.
It is indeed very difficult to model h u m a n knowledge, reasoning and experience in
a way t h a t can easily handled by computers.
On the other hand, control problems which need non-linear controllers can be solved
by using fuzzy controllers, but the main problem in this case is t h a t to design t h e
proper fuzzy controller (read: non-linear controller) requires a non-linear control
theory or a good description of the process to control. A good applicable non-linear
control theory has not yet been proposed and in case the last d e m a n d could be
fulfilled, there are also other ways to achieve similar results. In our opinion fuzzy
control is very useful for prototyping and designing non-linear controllers and offers
a great tool for the translation from high-level descriptions to low-level algorithms
and the other way around.
Fuzzy control can be used to provide a high-level design concept for the design
of non-linear controllers. T h e membership functions on universes of discourse in
combination with fuzzy rules can be "trcinslated" into a straight-forward non-linear
controller definition, which can be used for real-time control. This non-linear con-
troller definition exists of a simple look-up table, which contains the characteristic
points of the control hypersurface. Interpolation between these characteristic points
can be done using standard interpolation techniques.
195
Applications of fuzzy control in consumer electronics quite often axe implemented us-
ing look-up tables. Whether it was recognized or not, implementing fuzzy controllers
with look-up tables result in similar behaviour as implementing fuzzy controllers ac-
cording to theory. The main difference between the two is the interpolation between
the numerical results of the individual fuzzy rules. In case if look-up tables the in-
terpolation ccui be done explicitly using a lineaj interpolation (if there is a need for
interpolation with respect to the quantisation of the controller inputs and output).
Using fuzzy controllers according to the theory, this interpolation is done implicitly
by the logical operators in combination with the membership functions.
X,
i*'' input of fuzzy controller
A T-norm operator fuzzification of Xi
V T-conorm (S-norm) operator Xi fuzzy set for Xi in rule Vj
7j support value for rule TJ u numerical controller o u t p u t
Ae derivative of e Uj numerical o u t p u t due to rule TJ
e error signal {r — y) U-' numerical consequent of rule rj
Mfl, fuzzy set of Rj U fuzzy controller output
MR fuzzy set of fuzzy controller Uj fuzzy output due to rule rj
Aty fuzzy set of controller o u t p u t W fuzzy consequent of rule rj
l^u, fuzzy set of Uj y process output
f^Ui fuzzy set of U^
N, number of quantisations coa centre-of-area
K number of rules cog centre-of-gravity
Nu number of fuzzy sets for w ctrd centroid
N, number of inputs fm fuzzy-mean
r reference signal wfm weighted-fuzzy-mean
To j rule of fuzzy controller mom mean-of-maxima
R. fuzzy relation of rule Vj idfz indexed- defuzzification
R fuzzy relation of controller hgt height
196
References
AOKI, S., S. K A W A C H I A N D M . S U G E N O ( 1 9 9 0 ) . Application of fuzzy control logic
for dead-time processes in a glass melting furnace. Fuzzy Sets and Systems,
vol. 38, pp. 251-265.
ASSILIAN, S. (1974). AHificial intelligence techniques in the control of real dy-
namic systems. P h D - t h e s i s , Queen Mary College, University of London.
B A T U R , C . A N D V . K A S P A R I A N (1991). Predictive fuzzy expert controllers. Com-
puters in Industrial Engineering, vol. 20, no. 2.
B U C K L E Y , J . A N D W . S I L L E R (1987). Fuzzy operators for possibility interval
sets. Fuzzy Sets cind Systems, vol. 22, p p . 215-227.
D U B O I S , D . A N D H . P R A D E (1980). Fuzzy sets and systems: theory and applica-
tions, vol. 144 in Mathematics in science and engineering. Academic Press.
F E I G E N B A U M , E . A N D P . M C C O R D U C K (1983). The fifth generation: artificial
intelligence and Japan's computer challenge to the world. Addison-Wesley.
F L I N T H A M , T . J . M . (1991). Expert systems in control, why so few? Proceedings
l E E Control 91. Edinburgh, Scotland, U.K., March 1991.
H A R R I S , C . J . A N D C . G . M O O R E (1989). Intelligent identification and control for
autonomous guided vehicles using adaptive fuzzy-based algorithms. Engineering
Applications of Artificial Intelligence, vol. 2, p p . 267-285. December 1989.
H A S P E L , D . (1991). A new approach on the cement making process through
Linkman. Proceedings of 3^ International N C B Seminar on Cement and Build-
ing Materials. Dehli, India.
H E L L E N D O O R N , H . (1993). Design and development of fuzzy systems at Siemens
R&D. Proceedings of 2 I E E E International Conference on Fuzzy Systems
(FUZZ-IEEE 1993), p p . 1365-1370. San Fransisco ( C a ) , U.S.A., March 1993.
H E L L E N D O O R N , J . (1990). Reasoning with fuzzy logic. P h D dissertation. Delft
University of Technology. Delft, T h e Netherlcinds.
J A G E R , R . , H . B . V E R B R U G G E N A N D P . M . B R U I J N (1992). The role of defuzzi-
fication methods in the application of fuzzy control. Proceedings IFAC Sym-
posium on Intelligent Components and Instruments for Control Applications
1992, A. Ollero and E . F . Camacho (editors), p p . 111-116. Malaga, Spain, May
1992.
K L I R , G . J . A N D T . A . F O L G E R (1988). Uncertainty and information. Prentice
Hall.
L A R K I N , L . I . (1985). A fuzzy logic controller for aircraft flight control. Industrial
Applications of Fuzzy Control, M. Sugeno (editor), p p . 87-103.
This page intentionally blank
197
Kumar S. Ray
1. Introduction
Immediate after World War II people were very keen to develop sophisticated
tool for communication and control. Despite the landmark achievement of
the classical control theory through the launching of the first sputnik in 1957
and the subsequent developments of the clcissical control theory to modern
control theory which has been tested through a number of important high-
technology projects (viz the U.S. Apollo project), there are still serious problems
in the control of complex system. In manufacturing technology such as in
chemical processes or the steel industry, in power generations industry etc. the
conventional control algorithms are unable to manage the huge uncertainties
involved in the entire process and thus often require human interventions for
readjustments of the designed scheme.
Zadeh first realized that people can base decisions on imprecise, nonnumerical
information. In 1965, he was implicitly advancing a thesis which indicates that
199
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 199-275.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
200
under uncertain complex situations people are better at control than Machine.
In this connection Zadeh's significant achievements are the seminal papers on
the linguistic approach and system analysis based on the theory of fuzzy sets
[17, 18, 19, 20, 21, 22).
Being motivated by the above said contributions of zadeh, in mid 70's Mamdani
and his colleagues first demonstrated the successful applications of the fuzzy
logic controller (FLC). About the same time the first significant industrial
application of the FLC came up in Denmark at F.L. Smidth corp's cement
kiln.
During the past several years, fuzzy control has emerged as one of the most
potential areas for research in the application of fuzzy set theory. The con-
cepts of FLC is now an important adjunct to conventional control theory. The
tremendous applications of FLC indicate its effective utilization in the context
of complex ill-defined systems that can be controlled by a skilled human oper-
ator without the quantitative knowledge (in terms of deterministic algebra and
differential equations) of their dynamics.
The essential component of FLC is a set of linguistic control rules which are
generated by an experienced operator and which can be related by the dual
concepts of fuzzy implication and the compositional rule of inference.
Experience shows that the FLC sometimes yields results superior to those ob-
tained by conventional control algorithms. In fact the methodology of FLC
appears very useful because it can easily combine heterogeneous control laws
over a wide range of operating conditions within which almost every real sys-
201
2. M e t h o d of approximate Reasoning
Since the first formal description of fuzzy reasoning by Zadeh in [22j many
of t h e form
P: if X is A then Y \s B
q: X is A'
Where the variables X,Y take values in universes of discourse U,V respectively
of t h e form
202
P:if Xi is Ai and X2 is A2 and ... X„ is A„ then Y is B
q:Xi is i4j and Xj is ^'2 and ...X„ is A'^
Where XJ5 are variables taking values in universes of discourse U,; i = l ,2, .... n
and Y is a variable taking values in Y.A; and A'^ are fuzzy subsets of {/, : i = 1,
2, ...., n. B and B' are fuzzy subsets of V. The above fuzzy implication inference
rules are known as generalized modus ponens (GMP). The propositions above
the line are the premises; and the proposition below the line is the conclusion.
The criteria of inferences proposed by Zadeh [22] are summarized in Table 1.
We note that if a causal relation between " x is A" and "y is B" is not strong in
a fuzzy implication, the satisfaction of criterion 2-2 and criterion 3-2 is allowed.
Criterion 4-2 is interpreted as: if x is A then y is B, else y is not B. Although
this relation is not valid in formal logic, we often make such an interpretation
in everyday reasoning.
union x V y = max(x, y)
algebraic sum x^y = x + y — xy
bounded sum x ® y = min(l,x + y)
X y= 0
drastic sum xl±iy=:< y x = Q
1 x,y>0
disjoint sum xAy = max{77M'n(i, 1 — y),min(l — x,y)}
function and is denoted by A —*• B , where A and B are fuzzy sets in universes
Based on the above definitions many fuzzy implicatioii functions may be gen-
the drastic product are used respectively. T h e following fuzzy implications are
HR^ = A X B ^ Ji,^yfj.A{u)fJ.B{v)/{u,v).
where
205
Where
Although nn^, and fiRp do not have a well-defined logical structure the results
tabulated in Table II indicate that they are well suited for approximate reeison-
ing especially for the generalized modus ponens. fiR^ has a logical structure
which is similar to iiR^-tiR^ is based on the implication rule in Lukasiewicz's
logic. However, HR^ and ILR^ are not well suited for approximate reasoning
since the inferred consequences do not always fit our intuition. Furthermore,
for multiple-valued logical systems JIR^ and /x^^ have significant shortcomings.
Over-all ^IR^ yields reasonable results and thus constitutes an appropriate choice
for use in approximate reasoning.
206
If the membership functions of fuzzy sets A and B satisfy the following condi-
tions:
then the combination of the max-min compositional rule of inference and the
fuzzy relation /x/j, satisfies the following criteria which are the subset of the
criteria given in Table I. For the rule if X is A then Y is B:
The aim of this section is to introduce the basic ideas underlying the design
of FLC. In Figure-1 we have shown a the basic configuration of a closed-loop
system associated with a fuzzy logic controller.
207
Now, in the following we will describe the methods through which we can
realise the individual components of FLC and the functional involvements of
the components of FLC.
In most of the fuzzy control applications, the error inputs to FLC are nonfuzzy
in nature. But the data manipulation in a FLC is based on fuzzy set theory.
Hence, fuzzification of the error input to FLC is necessary. In the following we
208
r " -^
J Knowledge
^^. Base
^^
1
NonKJizy
input
, L
Fuzzy Inftrence
1
Nonhjzty
o j t p j f of F
; Fu22)fiC0»'0n of using fhe method Defuzzifico* on
the prror tnpu! of opproximote of Ihe t u n y
reosomng inference
1
f
1
Fu2 2y ioqic controMtr ( F L O .
(i) An error input, Co, is assumed to be a fuzzy set A with the membership
function /XA(*O) equal to zero except at the point e„ at which ^^(^0) equals
to one [60]
(ii) We may assign each error input to a fit vector of length L where the ith
fit or fuzzy unit [25] equals the value of the ith fuzzy set at the input
value. For instance, let us consider the Fig-3 which provides the library
of overlapping fuzzy set values defined on a universe of discourse.
let the error inputs be —6,2,3.9. Then from the library of fuzzy set we
obtain the fit vectors of length 5 as follows:
209
u.
1
/
0 /
10
-6 ^ (0,1,0,0,0)
2 => (0,0,0, .7, .7)
3.9 => (0,0,0,.1,1)
(iii) First the observed data are disturbed by random noise. Then the prob-
abilistic data are converted into fuzzy numbers, i.e. fuzzy (possibilistic)
data. In [26] [27] an isosceles triangle is chosen to be the fuzzification
function. The vertex of this triangle corresponds to the mean value of a
data set while the base is twice the standard deviation of the data set. In
[28] Dubois and Prade define a bijective transformation which transforms
a probability measure into a possibility measure using the concept of the
degree of necessity.
(iv) Sometimes observed data are precise as well as imprecise in some statisti-
cal sense. Such cases can be handled by hybrid number [27] which involve
210
3.2 Knovs^ledge B a s e
The knowledge base of a FLC has two components; namely a data base and a
fuzzy rule base. We shall consider some issues relating to the data base and
rule base in the following subsections.
3.2.1 D a t a base
In the following we discuss some of the important aspects relating to the con-
struction of the data base in a FLC.
where a = [0.4,0.6,1.0,0.8,0.4].
The most important feature of the FLC is to generate the fuzzy control rules
in an appropriate manner. There are, in general, four modes of generation of
fuzzy control rules as reported in [30]. Apart from mode of rule generation,
the choice of input (state) variables and output (control) variables, types of
fuzzy control rules and justification behind the control rules are are the salient
features of the rule base of a FLC.
213
*- trror input
10)
* - d«fuz7ifi*()
lb) outpirt of
FLC
The fuzzy control rules provide a convenient way to express the domain
specific knowledge through fuzzy if then rules.
The dimension and the dynamics of the real system are so huge and com-
plex that most of the time it becomes almost impossible to generate the
model (in terms of the state space equation or Transfer function). Thus
214
ural t o express the operator's rules as fuzzy if-then rules using linguistic
variables.
design based on the deep analysis of the system dynamics then we have
which consists of two rule bases. T h e first one is the general rule base of
215
the general rule bcised depending upon the desired overall performance
The choice of linguistic variables and their membership function play a signif-
icant role in the linguistic structures of a FLC. The linguistic variables in a
FLC are the state, state error, state error derivative, state error integral etc.
Generally, there are two types of fuzzy control rules, a) state evaluation type
b) object evaluation type.
Most FLC's have state evaluation type fuzzy control rules which can be
described as follows;
Ri : ifxisAi andyisBithenZisCi
Ri : ifxisAi andyisB2thenZisC2
Rn '• ifxisAn....andyisBnthenZisCn
Ri : ifxisAi...andyisBithenZ = fi{x,..y)
The state evaluation type fuzzy control rules evaluate the process state
(e.g. state, state error, state derivative, state integral) at time instant
t and compute a fuzzy control action at time instant < as a function of
(x, ..J/) and the control rules in the rule set.
The object evaluation type control rules satisfy the states and objectives.
A typical example of such a control rule is as follows;
There are two main ways to justify the generation of control rules ;
217
(a) A heuristic approach when control rules are generated in such a way that
the deviation from a desired state can be corrected and the control objec-
tive can be achieved. The generation of rules is purely on the quantitative
knowledge of the system behavior. Several approaches for adjustment of
rule selection have been reported in [43] [44] [45].
The decision making logic is mainly responsible for fuzzy inference using the
method of approximate reasoning as mentioned in Section 2.0. It has the
following components.
(iii) change of the insignificant part of a membership function from the exact
number zero to a very small positive value. This idea is discussed in [49].
Illustration:
Rule n : A„andBn ^ C„
^0 — m
(iii) Midpoint of M a x i m a M e t h o d :
Instead of taking all elements Wj which give the maximal grade, the
smallest element w' and the largest element w" among t h e m are
picked up a n d the midpoint of w' and w" is given eis the representa-
Wo are all derived from the calculation of the fuzzy set C which was
aggregated from C i , . . . , C ^ as in ( l ) .
(iv) Height M e t h o d :
(vi) Area M e t h o d :
_ ml-51+iD2-52+...+wnSn
^° Sl+S2+...+Sn
Proposition 1
Proof:
n n
C = {A',B') o\jRi = {A',B') o [J{AiandB, -^ C,).
1=1 1=1
The membership function ^lc' of the fuzzy set C is point wise defined for
a.\\weW by
221
Therefore
C = \{A',B')OR^]U\{A',B')OR^]LI...\J[(A',B')OR„]
= \J(A',B')oR,
t= l
n
= U (^', B') o (AtandBi -^ C.)
1=1
Preposition 2
{A\B')*\}URi^\^U{A',B')*IU.
Proposition S
i f / ^ A . x B , = /^>l, • A«fl,
222
Proof:
= supmin{[(/XA',Mfl')>min[(MA. -* M C , ) , ( M B . -* Mc,)]}
Hence we obtain
Proposition 4
5.1.1 C o m p l e t e n e s s of FLC
u eU 1 <i < n
In other words, the union of fuzzy relations, At be greater than zero for
all u e U;
the above relation holds because we deal with fuzzy labels which usually
overlap. The inequality condition (2) will be violated if, at the time
of construction of FLC, we miss some label or forget to describe the
relevant "if-then" control rule. Under such circumstances an additional
rule should be added to satisfy the inequality condition.
Ri : ifxisAithanyisBi,i = l,...,n
and if the input is A, then we would expect that the control action is B,.
But, in fact the derived control action may be a sub set or super set of the
225
Vi < i <„_Bi C Ai oR
Proposition 6 If for the fuzzy relations Ai, V^ are pairwise disjoint, i.e.,
Proof: Let supp {Aj) be a support of the fuzzy relation Aj and assume
the sujvmin composition of Aj and the relational matrix R.
226
5.1.4 C o n s i s t e n c y of f u z z y c o n t r o l r u l e s
The above phenomenon suggests to combine the FLC and the PID con-
troller. For distant control, the fuzzy controller is preferable while for
close by control (where the system is close to the setpoint) the PID con-
troller is preferable. Such ideas are discussed in [59][l6][l3] and some
encouraging simulation results are provided. But we need to determine
the appropriate switch criterion for such combined control or we can use
them sinaultaneously.
6. Implementations of FLC
The FLC has emerged as one of the most useful areas of research in the
application of fuzzy set theory. In many complex situations use of FLC-
based systems have proved to be superior in performance to conventional
systems.
228
[kg/cm']
1850
103 U ,
t t I [sec]
50
2.5 35
Throtlle pressure
[kg/tec]
188.8 I /f\ji'^^''-mm I I
ISO'-I-S 3 50
Sleom flow to HP turbine
[cm]
60
K Drum level
tisec)
50
T' ^"K
SOlsec]
ThroHl* prrssurv
[kg/s»c]
160 • T» IT • - J ^ ^ ^ S ^
'V2,b SO (see 1
1*0
SUam flow lo H P turbm*
[cm]
68
60 ^—^- '-«>r»r»
t so [s*c]
15
Orum (9V«t
P a r a m e t e r sensitivity:
0.5[s«?c]
15A.)5
I •<r<-,„ ^ 1 ^ .
US
JJ Throttle pressure
bO [sec]
[ kg/sec]
198
[cm]
1
6(.
50 [sec]
60
Drum level
Thrott>e pr«ssur»
L«ft Sighl
!PT) (PT)
govern i-f
D/Al
Mostfr <J«monc) j ^
signoL n T r »
Pulverizer's feeder-^
,'
© Pressure (ran s m i tier
[XI
jfference
@
Flow temperolure DJ Transfer switch
The function of this control loop is to maintain the drum water level
within the prescribed limits. This variable is controlled by modulating
the feed water control valves. Two 100% rated full load control valves
and two low load control valves are provided. Under low loads, one of the
low load valves is taken into service to control the drum water level, the
other is acting as standby. When the boiler load exceeds a preset value
(25-30% MCR), the control function is changed over through operator's
action, to one of the selected full load control valves the other one is
serving as standby. The schematic representation of the drum water level
control is shown in fig.9.
Both the control loops mentioned above are tested for 20% load change
at 60% MCR and the plant responses are shown in fig.10.
urve-fitted controller
Lkg/cm J
154 IS
U2.0
188.d
curve-tilted controller
Criteria 1 0 o X X 0 X X
Criteria 2-1 X X X X 0 X X
Criteria 2-2 o o X X X X X
Criteria 3-1 X X X X o X X
Criteria 3-2 o o X X X X X
Criteria 4-1 X X 0 o 0 0 o
Criteria 4-2 X X X X X X X
key: o = Satisfaction
X= Failure
238
MERICAL DEFINITION
Range NB NM NS ZE PS PM PB
-6.4< Co < -3.2 0.8 0.6 0.0 0.0 0.0 0.0 0.0
-3.2< £„ < -1.6 0.4 0.8 0.2 0.0 0.0 0.0 0.0
-1.6< Co < -0.8 0.2 1.0 0.6 0.2 0.0 0.0 0.0
-o.8< e„ < -0.4 0.0 0.8 0.8 0.4 0.0 0.0 0.0
-0.4< e„ < -0.2 0.0 0.6 1.0 0.8 0.0 0.0 0.0
-0.2< Co < 0.2 0.0 0.2 0.8 1.0 0.4 0.0 0.0
0.2< e„ < 0.4 0.0 0.0 0.6 0.8 0.8 0.0 0.0
0.4< e„ < 0.8 0.0 0.0 0.2 0.4 1.0 0.4 0.2
0.8< e„ < 1.6 0.0 0.0 0.0 0.2 0.8 0.8 0.4
1.6< e„ < 3.2 0.0 0.0 0.0 0.0 0.4 1.0 0.6
3.2< e„ < 6.4 0.0 0.0 0.0 0.0 0.0 0.8 0.8
ei fi ei «i ei "1
63 V3 63 W3 63 V3
Univ NULL very SMALL more or less MEDIUM more or BIG very
SMALL MEDIUM less BIG BIG
0.01 1.0 0 0 0 0 0 0 0
0.1 .75 .2 0 0 0 0 0 0
0.2 .5 .5 .3 0 0 0 0 0
0.35 .3 1.0 .5 0 0 0 0 0
0.5 .1 .5 1.0 .2 0 0 0 0
0.75 0 .3 .5 .5 0 0 0 0
1.0 0 0 .3 1.0 0 0 0 0
1.25 0 0 0 .75 0 0 0 0
1.5 0 0 0 .6 0 0 0 0
1.75 0 0 0 .5 .3 0 0 0
2.0 0 0 0 .3 .5 0 0 0
3.0 0 0 0 .2 .75 0 0 0
3.5 0 0 0 0 1.0 0 0 0
5.0 0 0 0 0 .75 0 0 0
7.0 0 0 0 0 .45 0 0 0
8.0 0 0 0 0 .3 .4 0 0
10. 0 0 0 0 .25 .5 0 0
12. 0 0 0 0 .1 1.0 0 0
13 0 0 0 0 0 .75 .2 0
15 0 0 0 0 0 .4 .3 0
17 0 0 0 0 0 .3 .5 .1
20 0 0 0 0 0 .25 1.0 .2
22 0 0 0 0 0 .1 .75 .25
25. 0 0 0 0 0 0 .5 .3
28. 0 0 0 0 0 0 .25 .5
30. 0 0 0 0 0 0 .1 .65
33 0 0 0 0 0 0 0 .8
35 0 0 0 0 0 0 0 1.0
244
Univ very SMALL more or less MEDIUM very more or BIG very
SMALL MEDIUM MEDIUM less BIG BIG
0.1 1.0 .1 0 0 0 0 0 0
0.15 .5 .5 0 0 0 0 0 0
0.2 .1 1.0 .1 0 0 0 0 0
0.35 0 .4 .5 0 0 0 0 0
0.45 0 .1 .7 .1 0 0 0 0
0.5 0 0 1.0 .3 0 0 0 0
0.75 0 0 .6 .5 0 0 0 0
1.0 0 0 .4 .7 0 0 0 0
1.5 0 0 .1 1.0 0 0 0 0
1.75 0 0 0 .8 0 0 0 0
2.0 0 0 0 .7 .1 0 0 0
3.0 0 0 0 .4 .2 0 0 0
3.5 0 0 0 .3 .4 0 0 0
5.0 0 0 0 .1 .5 0 0 0
7.0 0 0 0 0 .7 0 0 0
8.0 0 0 0 0 1.0 0 0 0
10. 0 0 0 0 .6 .1 0 0
12. 0 0 0 0 .5 .25 0 0
13 0 0 0 0 .3 .3 .1 0
15 0 0 0 0 .25 .5 .3 0
18 0 0 0 0 .1 1.0 .45 0
20 0 0 0 0 0 .6 .6 0
22 0 0 0 0 0 .4 .7 .2
24. 0 0 0 0 0 .3 1.0 .4
26. 0 0 0 0 0 .1 .7 .6
28. 0 0 0 0 0 0 .5 .8
29 0 0 0 0 0 0 .3 .9
30 0 0 0 0 0 0 .2 1.0
245
Since the successive input to a fuzzy logic controller is the error between
Now the fuzzy control algorithms art based on fuzzy implication and com-
positional rule of inference which has many drawbacks [18] and which
In this section, our basic aim is to present a design technique which does
not suffer from the existing drawback of the compositional rule of inference
with the technique of curve-fitting through some points obtained from the
248
p «* X ts A'
q •^ if X is AthenYis B
Now we first defuzzify the relations in the premise q. This can be done
in two ways.
The relations Ri, R2.,...,Rn can be formed using many available translat-
ing rules [l)[60]. We use two widely used rules for the translation of such
compound propositions. They are Zadeh's arithmetic rule, Mamdani's
min rule, viz.,
Now if we use Zadeh's rule for the translation then we construct the fuzzy
relational matrix R according to R = Ry (~\ R2 f]... f\ i2„ otherwise, we use
R = RiDR2D...[JRn for the same.
Now every relational matrix R of the form discussed so far, when defuzzi-
250
very beginning. Now a fuzzy set A can be defuzzified using many differ-
ent rules [11). We use a simple one : T h e defuzzy value of the fuzzy set
Let us consider the following problem in which we are given three fuzzy
if X is A then Y is B
251
where X a n d Y are two linguistic variables defined over U a n d V respec-
tively a n d a n o t h e r fuzzy premise of t h e form
Xis A'
fJ-R{x,y) = mm{niti{x,y),fj.K2(x,y),lJ.R3(x,y)}
we find
1 1 .6 .3
R = 3 .6 1 .6
5 .3 .6 1
X is A'
iii ^^ if X is 1 then y is 6
^2 ^^ if X is 3 then y is 7
iia ^ if X is 5 then y is 8
y = 5.5 + 0.51 3
from t h e relation (3) for a: = 4 we find y — 7.5, which is the same as the
defuzzy value of "Y is (.6/6 + .85/7 + .85/8) ". But these two values of
plant using the d a t a in tables X, XI, XII, XIII. Curve-fitted control action
follows.
From the given tables, it is obvious that the proposed curve-fitted con-
troller is as good as a fuzzy logic controller for an easy implementation of
control action for nonlinear plant. The switching line for the bang-bang
control or variable structured control of a nonlinear plant is difficult to
calculate. Hence the simplicity of the curve-fitted controller and the fuzzy
logic controller is an added advantage over the classical approach to con-
troller design.
Finally, in Figure 11 we have shown that the steam flow and throttle
255
Inference obtained
by conventionoi
fu72y logic controller
135Kg/S»c
ISlKg/Cm' St»om pressure 175Ks/Cm'
8. conclusion
need a very sophisticated tool for knowledge acquisition for the design of
the F L C .
"negative m e d i u m " .
&: Tanaka
D u t t a Majumder
(14][17l[83l [841(85].
References
16. K. S. Ray and D. Dutta Majumder " Fuzzy logic control of a non-
linear multivariabje stem generating unit (200 |MW]) using decu-
oupling theory" in IEEE Trans. Systems, Man and Cybernetics,
(1985), SMC 15, 4, pp. 539 - 558.
19. L.A. Zadeh, "A rationale for fuzzy control". Trans ASME. J. Dy-
nam. Syst, Measur, Control. (1972) Vol 94, pp, 3-4.
261
21. L.A. Zadeh, "The concept of a linguistic variable and its application
to approximate reasoning 1, 11, 111," In formal Sci., (1975), vol. 8,,
pp.199-251, pp. 301-357, vol. 9 pp. 43-80.
28. D. Dubois and H.Prade, "Unfair coins and necessity measures: To-
ward a possibilistic interpretation of histograms". Fuzzy sets and
system (1985) vol- 10, No 1, pp. 15-20.
262
29. L. larkin, "A fuzzy logic controller for aircraft flight control", In
industrial application of fuzzy control, M. Sugen, Ed. Amsterdam:
North Holland (1985) pp. 87-104.
35. S. Shao, "Fuzzy self-organizing controller and its application for dy-
namic processes." Fuzzy Sets Syst., (1988), vol. 26, pp. 151-164.
36. R. Tanscheit and E.M. Scharf, "Experiments with the use of a rule-
bcLsed self-organising controller for robotics applications," Fuzzy Sets
Syst., (1988) vol. 26, pp. 195-214.
125-138.
2. no. 3, p p . 419-432.
(1987) p p . 835-838.
11 p p . 625-626.
199.
553-577.
49. T. C. Chang, K. Hasegawa and C.W. Ibbs, " The effects of mem-
bership function on fuzzy reasoning." Fuzzy Set. System (1991) vol.
44, pp. 169-186.
52. S. Gottwald and W. Pedrycz, " Problems of the design of fuzzy con-
trollers." inApproximate Reasoning in Expert Systems, M.M. Gupta
A. Kandel, W. Bandler, and J. B. Kiszka, Ed. Amsterdam: North-
Holland ,(1985), pp. 393-405.
53. B.R.Gaines and L.j. Kohout, "The fuzzy decade: "A bibliography
of fuzzy systems and closely related topics," Int. J. Man. Mach.
Studies, (1977) vol. 9, pp. 1-68.
56. D. Dubois and H. Prade, "Fuzzy logic and the generalized modus
ponens revisited," Cybern. Syst., (1984) vol. 15, pp. 3-4.
60. C.C. Lee, "Fuzzy logic in control systems: Fuzzy logic controller -
Part I & I F , IEEE Trans. Syst. Man and Cybern. (1990) vol.- 20,
no.2, pp.404-435.
62. W. Pedrycz, "Fuzzy control and Fuzzy system," John Willey and
Sons INC. New York (1989).
64. R.M. Tong, M.B. Beck and A. Latten, "Fuzzy control of the acti-
vated sludge wastewater treatment process," Automatica (1980) vol.
16, no. 6, pp. 695-701.
66. C.P. P a p p i s and E.H. M a m d a n i , " A fuzzy logic controller for a traffic
10 p p . 707-717.
68. I.G. Umbers and P.J. King, "An analysis of h u m a n decision making
(1980)
105-124.
(1983) p p . 43-48.
267
77. J.A. Bernard, "Use of rule-based system for process control," IEEE
Contr. Syst. Mag., (1988) vol. 8, no. 5, pp. 3-13.
82. G.J. Klir and T.A. Folger, Fuzzy Sets, Uncertainty, and Information.
Englewood Cliffs, NJ: Prentice Hall, (1988).
268
84. J.B. Kiszka, M.M. Gupta, and P.N. Nikiforuk, "Energetistic stability
of fuzzy dynamic systems," IEEE Trans. Syst. Man Cybem., (1985)
vol. SMC-15, no. 5, pp. 783-792.
Appendix
Physical description of the plant [13]
The 210 (MWJ thermal power station was designed on the concept of
the unit system, where single steam generator supplies steam to a single
turbine coupled with an AC Generator.
S t e a m generator plant:
The steam generator made in the BHEL is of single drum water tube
natural circulation type with superheaters and reheaters. The maxi-
mum continuous evaporation of the steam generator is 680[Te/hr.], at
I54.15[kg/cm^\ and 540[°C]. The final feed water temperature is 243[°C].
temperature \"c\
(1) (2) (3) (4)
1 0.20(37 0.2122 0.1710 0.2178
S c h e m a t i c r e p r e s e n t a t i o n a n d m a t h e m a t i c a l d e s c r i p t i o n of t h e
plant :
Control valve
of H, R torbine
variables are throttle pressure P, steam flow to H.P. turbine S and drum
water level L. Thus, a nonlinear multivariable plant having three inputs
and three outputs is considered. The validity of the model is tested
with real industrial data of 210 [MW] unit (BHEL, INDIA). The plant is
assumed to be operated at 60% maximum continuous rating (MCR). A
step change of ± 10 - 20% around the operating conditions of the plant is
injected into the open-loop plant. It is observed that the open-loop plant
dynamics shown in fig.A2 agree quite well with real plant responses. The
details of the system model is given in [13].
[cm]
68 2 6 A 5 -
64 0267
Drum 'evel
• i~ 2'.5 T = 3000[sec]
62 8 2 5 3 - 5 0
SO
175.8
TJ 3000
Drum pressure
151.30
i l<9 / s e c j
U 8 36
SteOm flow
135,0
8 T = 300n(sec]
73.96r
H«9/s»cJ
135-00 ^ ' ^ - - - _ _ _ _ S < . o m fto* T = 3000ls»c)
133-85
[kg/cm']
175.7968,
175.8 k
\\..___D™m pressuro 1-3000[s«)
172,8231
3
51 T= 3000 [sec]
S6.5*
kg/cm ]
209.31-
17 6 8
T-SOOOlsecj
[ kg/sec]
11.7.30
Sfeom flow
135.0
T = 3000 [sec]
The set of first order nonlinear differential equations which represent the
system's dynamics in time-domain is given by
dPIdt = {-0.001935.P. + 0.014525Pf - 0.000736Fw
+0.00121i + 0.000176r£')
dS/dt = (lOCv.pi/2-0.7857165) {A\)
In the above description of the model, the variable 'TE', which repre-
sent the feed water input temperature, may be reasonably assumed to
be constant (no heater [high pressure or low pressure] is tripped during
operation). However, if any fluctuation in the variable 'TE' occurs due
to either improper heat transfer in the economizer section of the boiler,
through which the input feedwater finally passes, or sudden tripping of
any heater etc., that fluctuation of 'TE' becomes very unpredictable in na-
ture. Hence, for simplicity of the decoupling controllers structure, which
is based on the mathematical description of the nonlinear multivariable
plant/system, the variable 'TE' of the model (Al) is set to zero. Thus
we get the following simplified mathematical description :
ii -0.00193i2i?^^^ + 0.0012113
i2 = -0.785716i2
X3 - 0 . 0 0 0 0 0 6 1 ? - 0.0073283;^ - 0 . 0 0 9 1 4 i 3 - 0 . 0 0 0 0 0 2 i §
(A2)
1 0 0 Xi
0 1 0 X2
Yz 0 0 1 X3
The state variables (P,S,L) = (xi, 12)2:3) and control inputs (Pp, Cv,Fw)
274
(UIJUJJWS), respectively.
(i) Deriving the nonlinear state variable feedback law for the block de-
coupling of the plant [13]
(ii) Generation of the FLC/CFC for the block- decoupled subsystems.
The three outputs of the plant are steam flow to H.P. turbine, throttle
pressure and drum water level. Among these outputs, two of them (steam
flow to H.P. turbine and throttle pressure) will be controlled together and
the remaining one (drum water level) will be controlled separately. A
careful inspection reveals that the second output (i.e., the steam flow X2)
of the plant represented by (A2) does not depend directly on drum water
level (xs). Hence, in the present problem, for the sake of block-decoupling,
the first output will be straightway decoupled from the third. Thus, the
first and the second outputs will be decoupled from the third, whereas the
first and the second outputs will remain interacting within themselves.
Here, the advantages of a particular mathematical description (A2) of the
given steam generating unit will be exploited for block-decoupling [13].
'-^ ^ y^ eqOQtions ^ L
Blok
11 decoupled
I r i x ) * <J
outputs
I
I I
Outer-loop feedbock
Acknowledgment
Abstract
A fuzzy-sliding mode controller, which is designed by the techniques of the fuzzy logic
controller and the sliding mode controller (or called variable structure control), is pro-
posed in this work. Like the sliding mode of the sliding mode control system, the fuzzy-
sliding mode control system has a fuzzy-sliding mode. The reason for calling "fuzzy-
sliding mode" is that the sliding surface in the proposed scheme is a fuzzy set rather than a
crisp set found in the conventional sliding mode control system. In the design of the fuzzy-
sliding mode controller, one can easily determine the membership function, observe the
fuzzy rules and predict the controlled system behavior. Furthermore, the number of infer-
ence rules, which is an exponential function of the number of system state variables in a
conventional fuzzy logic controller, is reduced to a linear function of the number of system
state variables in the fuzzy-sliding mode control system. Simulation results show that the
proposed scheme has the following advantages:
277
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 277-306.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
278
1. Introduction
The fuzzy logic control has become one of the most important areas of research in the ap-
plications of the fuzzy set theory [l]-[4] since about twenty years ago, Mandani and his
co-workers [8]-[10] have successfully applied the simple_/Hzzy logic controller (FLC) to
several industrial processes such as the steam engine, warm water plant, stirred tank tem-
perature system, etc. During recent years, the literature fuzzy logic control has been grow-
ing rapidly. Most of the historically important milestones in both the developtnent and
application of fuzzy logic control have been summarized by Lee [5][6].
By converting the linguistic control strategy into automatic control strategy without
using the mathematical model of controlled systems, the FLC can be used to deal with
complex ill-defined systems. However, the work of FLC design is time-consuming and the
response trajectory of the controlled system is unpredictable because fuzzy control rules
are experience oriented and the suitable membership function should be selected by the
trial and error procedure. To overcome these disadvantages, we will introduce the sliding
mode of the sliding mode controller (SMC) to the FLC. We will call it the fuzzy-sliding
mode controller (FSMC).
The conventional sliding mode control methodology has been developed in the lit-
erature [15]-[17]. By choosing a suitable sliding surface
X = { X I cj(x) = 0 }
and applying discontinuous control action across the surface, the resulting state trajectory
of the controlled system can be transferred onto the sliding surface and then slide along the
surface.
There are two phases in the sliding mode control system. One is called the "hitting"
phase and the other is called the "sliding" phase. In the hitting phase, a hitting control law
will be applied to transfer the representation point everywhere of the state space onto the
sliding surface. As soon as the representation point hits the surface, the controller turns the
sliding phase on, and applies an equivalent control law to keep the states lying on the slid-
ing surface. In an ideal situation, when the state trajectory stays on the sliding surface, the
control system then has perfect insensitivity with respect to the external disturbances as
well as the plant parameter variations. Furthermore, even if the system is perturbed such
that the representation point is out of the sliding surface, the control system turns on the
hitting phase again, and reacts the hitting control law to bring the states toward the sliding
surface once more. Therefore, the sliding mode control imposes certain constrains on sys-
tem dynamics; in other words, one can predetermine a sliding surface to dominate the dy-
namic behavior of the controlled system.
However, the methodology of sliding mode control system has also some draw-
backs as follows:
1. The state trajectory in the hitting phase is usually sensitive to system uncertainty.
This difficulty can be alleviated by using high gain feedback to save hitting time [21] [22].
279
But there still exist some problems associated with high-gain feedback such as extreme
sensitivity to unmodclled dynamics, actuator saturation, etc. [18].
2. Due to various nonidealities such as hysteresis of switching operation, time delay
of the control system, and sampling of digital implementation, the state trajectory of a real-
ized sliding mode control system does chatter rather than slide along the sliding surface
[15]. This undesirable chattering phenomenon will cause high-frequency unmodelled dy-
namics in the controlled system.
In the course of designing FSMC, we first adopt the methodology of SMC to select
both the membership functions and the fuzzy inference rules, and then fuzzily the sliding
surface and get the feedback control gain from the fuzzy inference rule base. Because the
FSMC is composed of FLC and a SMC, it might have the advantages of both.
The simple FLC based on fuzzy set theory [l]-[4], originally conceived by Assilian
and Mamdani [11], is used to regulate the output of a process around a given set-point.
The basic configuration of the simple FLC comprises three principal components: ajuzzi/i-
er, a rule base with inference engine, and a defuzzifier as shown in Figure 1. We briefly
describe the functions of them as follows:
A. Fuzzifier
The main function of the fuzzifier is to convert crisp data into fuzzy sets, i.e.
A'=F(x)=F{XaXGX) (1)
B' = F(y)=Fiy,xGY) (2)
where A' and B' ate two fuzzy sets, and F(.) represents a fuzzification operation [2], Xg and
_Vo are the crisp input values from the process, GXand GY arc the scaling factors of x„ and
yg such that x e (7 and ysV.
In general, an FLC consists of a set of linguistic rules which define the individual control
situation. For instance, the control rules of two-input-one-output FLC have the form:
where X and y arc linguistic variables representing two process state variables, and z is a
linguistic control variable; A., B. and C, withy = 1, 2,..., /, are linguistic values of x, y, and
z in the universes of discourse U, V, and W, respectively.
The fuzzy control rule is impleniented as a fuzzy relation in C/ X F x W and is de-
fined as follows
In this work, we adopt the widely used Mamdani's mini-operation fuzzy implication func-
tion [9], and hence the membership function oiR^ is given by
Thus the overall relation, R, of equation (3) can be implemented by using the union opera-
tion to link all rules together; that is
Consider now, i{A' and 5 ' denote the linguistic values of inputs x and y, respective-
ly, then the linguistic value of output z can be obtained by using the max-min composition-
al rule of inference
C'^{A'nB')°R (7)
C. Defuzzifier
Because the control signal of a process must be a crisp value signal rather than a fuzzy
one; the linguistic value of control output, C, which is obtained from inference engine,
must be converted into a crisp value, z^, symbolically
where D(.) denotes any function of dcfuzzification. The one which is widely used and will
be adopted in the defuzzifier of FSMC is the center of area (COA) method [13].
GX
GZ
Rule Base
Fuzzifxer and c; Defuzzifier 4 ^ .
Inference Engine
GY
To speed up the inference and to save memory one can preconstruct a decision
Junction (or, approximately, a look-up table) based on Eqs. (1), (2), (6), (7) and (8) for the
FLC. For example, consider an FLC with error e and change in error Ae as two input vari-
ables, and change in control Au as one output variable, then one can formulate
The conventional sliding mode control is also called the variable structure control, which
has been developed essentially in the Soviet Union [15]-[17]. Consider a system to be con-
trolled which is described as follows:
where x = [Xj x^ ... x„]^ is an /i-dimentional vector that denotes the state of the system, f is
an n-dimensional vector, M is a scalar control input and b is an «-dimensional control gain
vector.
2:={x|o(x) = 0} (11)
where o(x) is an arbitrary function of x which will dominate the dynamic behavior of the
control system and have to be properly chosen. From Eq. (10), we see that
in which g^(x) is the gradient vector function of a(x). Without loss of generality, we as-
sume that g^(x)b(x) > 0 and is bounded.
In the design of the sliding mode control system, we must first find the equivalent
control law, u^^, which will keep the state of the controlled system staying on the sliding
surface whenever X6 Z, that is
282
o\.^^=0 (13)
g\x)[f(x) + b ( x ) « J = 0 (14)
Hence
Therefore, under the sliding mode, the equivalent closed loop system can be obtained by
substituting Eq. (15) into the original system of Eq. (10), as follows:
From Eq. (16), we see that the gradient of the sliding surface, g^x), will govern the
dynamic behavior of the control system in the sliding mode. Thus one can expect to obtain
good performance from the control system by suitable design of the sliding surface.
In the hitting phase, where xe R°-Z, a hitting control law will be found to transfer the state
anywhere to reach the sliding surface. Let us define a Lyapunov function as follows:
V = |o'(x) (17)
It is shown in [19] that if there exists a positive constant ri, such that
where r||CT| is a function of class K [23], then the state trajectory will hit the sliding sur-
face, Z, within a finite time /^ defined by
^ |o(x(0))|
/, < ^ - ^ (19)
The above results show a global sliding condition and guarantee that all initial state lying
off Z will be attracted to Z. From Eqs. (12), (17) and (18), we see that the sliding condi-
tion is satisfied if
283
or, equivalently
« = «., + «j (23)
in which k: R " ^ R " satisfies k(x) > a(x) for all xe R".
In the ideal sliding mode control, as soon as the state trajectories of the system hit
the sliding surface, the state trajectories slide along the surface. However, because of non-
idealities of the real systems, a realized sliding mode control system may have chattering
along the sliding surface. To alleviate the chattering phenomenon, many strategies have
been proposed. For instance, Slotine and Sastry [18] suggested a quasi-sliding mode con-
troller which replaces the term sgn{o) in Eq. (25) by sat{o/e), i.e.,
To date, the fuzzy logic control and the sliding mode control have been two different re-
search topics. The basic idea of combining these two methodologies together and formu-
lating a scheme of fuzzy-sliding mode control is that the appearance of nonidcalitics such
that the sliding surface of the sliding mode control system is no longer crisp. In other
words, it is possible that a state, say x^, which belongs to Z in the ideal sliding mode but
does not belong to the same set as nonidealities appear. Hence, the sliding surface turns
into a fuzzy set rather than a crisp one in the real situation, and the fuzziness of such a set
depends on the amount of nonidealities. Therefore, we can use the fuzzification operator
to fuzzily a crisp sliding surface and then obtain a fuzzy-sliding surface. Then, we use the
fuzzified version of the crisp variable a as a linguistic input variable of the FSMC. To
handle more information, we also fuzzily the derivative of a as another linguistic input
variable. Hence, an FSMC uses the linguistic variables of a and d as inputs instead of er-
ror and change in error in the conventional PLC. In order to develop the FSMC, we re-
write the FLC method and the SMC scheme as follows:
= D{[F(exGE)nF(^AexGQ\oR}xGU (28)
« = «., + u, (29)
with
u, = -k(x) sgn(a) (30)
where k: R - ^ R " satisfies k(x) > a(x) for all xe R", and a(x) = r|(g^(x)b(x))' .
In the approach of designing FSMC, we will use the fuzzy logic control law of Eq. (27) to
replace the sliding mode control law of Eq. (29); namely, the fuzzy sliding mode control
law will be formulated as
285
u{t + At) = u(t) + Ujf (31)
where Ujf is obtained by fuzzy relation and what we wish is that E{u(f)} —> M^^ as ? —> <»
(£{•} denotes statistical expectation). To establish such a fuzzy-sliding mode control law,
the designer must keep in mind some physical meaning and geometry interpretation of the
sliding mode regime. We illustrate these concepts by viewing the dynamics of a linear
system.
Referring to Eqs. (10), (11), suppose that f(x) = Ax, b(x) = b,CT(x)= c^x, and g^(x)
= c^, then, the global sliding condition for such a system is given by
acKO (32)
in which k^^, the equivalent feedback gain vector, and kj, the discontinuous feedback gain
vector, are respectively given by
where kj is the /-th element of the vector k^. Substituting Eqs. (33), (34) and (35) into the
original system of Eq. (10), one can obtain
i=(A-bkJ)x-bk/x (36)
and
<j= c^ X = c'r(A - bkJ,) X - c'^bkj x = - c^'bk/x (37)
From the viewpoint of geometry, any state x can be represented as a point in the
state space. Thus, the value |CT(X)| can be interpreted as the distance between the represen-
tation point X and the sliding surface Z. Consequently, in the physical sense, d can be
viewed as the velocity of the representation point tends toward the switching surface. From
Eq. (37) one can obtain
Hence, the magnitude of discontinuous feedback gain, k^, governs the convergent rate of
O. Kkg is increased, the speed of representation point that moves toward the surface accel-
erates and the hitting time is decreased. It also has been shown that a sufTiciently large
gain can guarantee the existence conditions of the sliding mode. In practice, however, as
the switching operations do not occur as soon as the sign ofCTchanges, the larger gains
cause the larger trajectory chattering. This is due to the switching criteria in the conven-
tional sliding regime based on a sign function, which is a crisp version and has a sharp
edge. In other words, the sliding surface in the conventional sliding mode control is a
fuzzy singleton with membership equal to 1 as o = 0 and equal to 0 as (J ;* 0. Now, let us
define the following fuzzy sets:
CT.p : a is positive
a.z : o is zero
CT.n : a is negative
such that \i^^B [0, 1], H„^G [0,1] and ^l„„e [0, 1].
It is obvious that in an ideal sliding mode, the above fuzzy sets degenerate into
crisp sets, i.e.,
CT.p={K|a(x)>0}=5:P (39a)
a.z={x|CT(x) = 0} = 2: (39b)
a n = { x | a ( x ) < 0 } = ZN (39c)
I 1, / / CT>0 ,.„ ,
1, if 0 = 0
(40b)
0, if CT^tO
, 1, i/ o < 0 ,.„ ,
^''"io; ,/a>o ('«^>
Similarly, in the quasi-sliding nrode proposed by Slotine and Sastry [18], the
switching surface becomes a boundary layer, which is also a crisp set, thus
az = {x||a(x)|<e} (41b)
]i.(5.p —
_ J 1, if a>e (42a)
0, if CT<e
(42b)
1, if a < - e (42c)
\i.a.,.
0, if CT>-e
In real situations, the switching surface is a fuzzy set rather than a crisp one. So,
the first step in the design of FSMC is to fuzzify the sliding surface; that is
a.p=F(XP) (43a)
1, if 0 > / j
-alls. if - /, <CT< 0
|Xo.«(CT) = • 1, if CT<-/, (44c)
0, if CT>0
288
Similarly, we can fuzzify Aa(x(t)) = C(x(t)) - a(x{t-At)) = oAt , change a, in the same
way, such that
I, if ACT>/C
^4o.p(Aa) = ACT//,, if 0 < ACT < /, (45a)
0' if ACT<0
From preceding discussion, we notice that a and ACT can be interpreted as the distance
from the representation point to the sliding surface and as the velocity of representation
point tends toward the sliding surface, respectively. Furthermore, we have indicated that
larger feedback gains cause faster convergence but larger trajectory chattering. To attenu-
ate the chattering, the feedback gains must be chosen appropriately. Therefore, in design-
ing the FSMC, we will use a fuzzy relation to obtain a nonlinear state feedback control
gains based on the specification ofCTand ACT. To do this, let us select two groups of vari-
ables as the input variables of the FSMC
V, = Aasgn{x), I = 1, 2 n (46b)
Then we transfer these two groups of variables to their corresponding universes of dis-
course by separately multiplying two groups of scaling factors, i.e.,
Consequently, the fuzzy-sliding mode control law of Eq. (31) can be formulated as
follows:
Thus, for an n-state system, we should establish n sub-rule bases to form an overall
rule base. Each of them has two inputs s. and cs. and one output kj. If we define S-, CS.
and KD. as the linguistic variables corresponding to 5,, cs- and k^^ respectively, the fuzzy
sets A^j, Bjj and Q ,j = 1, 2, ..., /„ as the linguistic values of 5, , CS^ and KD. respectively,
then the rule base may have the following form:
in which /?,y is the J-th rule in the i-th sub-rule base R., and /, denotes the number of rules in
R-, therefore
R = {R„R„...R,,...,R„} (50b)
Consequently,
Figure 2 schematizes the architecture of the FSMC. Note that for a n-statc system,
a FSMC consists of n distinct two-input-one-output FLC as well as a sliding function
evaluator (or called an input variables generator). Thus, the FSMC has 2n input variables.
290
Theorem If the input space of each input variable of the FSMC is divided into m parti-
tions, then the number of inference rules in a complete rule base for the FSMC
is equal to / = nm^, where n stands for the number of system states.
proof: For a system containing n state variables, x„ x^, ..., x„, the rule base of the FSMC
has n distinct sub-rule bases, namely, RB^, RB2,..., and RB„, and each sub-rule base
RB- has two input variables, 5, and CS., as well as one output variable, AX),. Every
sub-rule base RB. consists of/^ = m^ rules, / = 1, 2, ..., n, to form a complete sub-
n
rule base. Therefore, n complete sub-rule bases have I = Y, li = nm inference rules.
D
For a conventional FLC, the complete fuzzy rule base for a system has n state vari-
ables, and that each of the state variables have m partitions, as inputs has m" inference
rules. The above theorem indicates that by using the FSMC, the number of inference rules
in a complete rule base is reduced to a linear function of system state n, instead of an expo-
nential function of n for the conventional case. Therefore, the fuzzy-sliding mode control
scheme performs with great effectiveness in the control of a large scale system.
MUX
^ ^-R7>
RB: Rule Base P : Plant MUX: Multiplexer
F : Fuzzifier DF: Defuzzifier DMU: Demultiplexer
Generally, for a system with n state variables, one must establish n distinct sub-rule bases
to form a whole rule base for the FSMC. In practice, the n distinct sub-rule bases can be
established identically; therefore, one can establish only one rule base which has two lin-
guistic input variables, S and CS, as well as one linguistic output variable, KD, with re-
spect to the crisp input variables, (7sgn(x) and Aasgn(x), and the crisp output variable, k^^ ,
for all I = 1,..., «. The rule base might have the form
In this situation, the input variables are sequentially fed into the controller by a multiplex-
er, and the output variables are demultiplexed by a demultiplexer. Consequently, the struc-
ture of such an FSMC becomes a serial one as shown in Figure 3.
In the following we illustrate three examples of second order systems and show how to de-
velop the fuzzy-sliding mode control strategy. These examples are simulated by a digital
computer with a sampling interval of 0.1 sec. and with the initial condition x(0) = [ 1 0 ] \
0 1 "0"
x= X + (53)
- 1 -2 _ 1
If we choose the sliding surface for the conventional sliding mode control as follows
a = c'^x = [l l ] x
and select ^o= 5. From Eqs. (33), (34), and (35), we have
k J = (c^b)'c^A = [-l-l]
and
«^ = -ka^x = -£A:;j:,.
Figure 5 shows the response of the conventional sliding mode control systeiti. Next, let us
design a FSMC for the system of Eq. (53). In designing the FSMC, we define the asso-
ciated fuzzy sets to be shown in Table 1 and their corresponding membership to be shown
in Figure 4, where we assume that
and their universes of discourse are all assigned to be [-6, 6]. The fuzzy-sliding mode con-
trol rules are described in Table 2.
293
0.5 - 0.5
0.5
Set
GS, =GS^ = 6
GCS^ =GCS^ =12
GK, =GK^ =0.85
The response trajectories of fuzzy-sliding mode control are shown in Figure 6. Comparing
Figure 6 with Figure 5, one will notice that the chattering phenomenon in the fuzzy-sliding
mode control system is smaller than that in conventional sliding mode control system.
Example 2 To view the robustness of fuzzy-sliding mode control, let's assume that the
system in Example 1 has paranrKter uncertainty as follows:
0 1 0
X = \ +
H - A ^ i -2-)-Aa2 J I
0 1 X\
X =
-1 -2 \X2\
Again, we apply the SMC and FSMC as given in Example 1, the response tra-
jectories of both control schemes arc shown in Figure 9 and Figure 10,
respectively.
(c) (d)
Figure 5. The response trajectories of conventional sliding mode control
system (Example 1)
u 0.5 -
10
f. 5
o
-5
10
time, sec time, sec
(c) (d)
Figure 6. The response trajectories oifuzzy-sliding mode control
system (Example 1).
296
a 0.5
-10
0) 0.5
time, sec
10
-5
10
time, sec
(d)
<u 0.5
state, xl
(.0)
•Z 0.5
-0.5
-5
4.2 Applications
In this subsection, we will apply the scheme of FSMC to deal with two dynamic physical
systems. One is a highly nonlinear inverted pendulum system, and the other is a two-link
robot arm which has not only nonlinearity but also coupling effects.
gsinxi-cosxi [-;;^xlsmxi-^;;^f^
In the design of FSMC, we treat the inverted pendulum as a black box with one input
variable,/, and two output variables, Xj and JC^ , and we assume that x, and Jf^ are measur-
able. The control system is simulated by a digital computer with sampling interval At = 2
ms and initial condition x^{0) = 1 rad and XjCO) = 0 rad/sec.
A. Regulation
The regulation problem is to balance the pole in the vertical position by using the
actuactor force/ to control the motion of the cart. So, the sliding surface can be chosen as
CT = x, + X2 . By applying the FSMC as given in Example l,we get the response of the
fuzzy-sliding mode control system as shown in Figure 12. To view the robustness of
/
T3 D"^
Figure 11. The inverted pendulum system
299
control system, one may assume that the pole has load variations and external disturbanc;
the related results are shown in Figure 13 and Figure 14, respectively.
B. Tracking
Now, suppose that we want to control the states of the pole to track a desired sinu-
soidal trajectory given by
x,j = a sin(M)
x^j = dx,j / dt = aco cos(CiX)
Again, adopting the FSMC as given in Example 1, we get the system response as shown in
Figure 15. Similarly, by adding the load variation and external disturbance shown in Fig-
ure 13b and Figure 14b, we get the resulting trajectories shown separately in Figure 16 and
Figure 17.
2. Two-Join Manipulator
Consider a two-joint manipulator shown in Figure 18, where g is the gravitational con-
stant, 0, and 0j are joint angles, and u, and M^ are the applied torques. Let
Xi = 6 i , X2 =6] , X3 =02, ^4 =02 and under the assumption that there exist lumped
equivalent masses and massless links, then its dynamics can be represented as the follow-
ing state space [20][22]:
Xl ~ Xy
^ m v2.i.oft
Xl = -f(^,2>c^' . . +
+ 2P,2A:2A:, ^ „y,g
„ ^+. M,)
. ^ -M
^ ( /- p
f .t ^ x. ,2' .+ y,g + «,)
X3 —X^
with
a,, = (m, + w^) / / -t- 7M^ /j^ + Im^ I, I2 cos Xj -H J,
0.5
5 10
tfme, sec
(b)
0.4
0.2
5 10
time, sec
(b)
-50
-100
0.5 1
time, sec Angle, rod
(c) (d)
Figure 13 The responses of inverted pendulum system with load variation
(Regulation)
301
time, sec
>
<c
time, sec
(c)
>
c
<
>
c
<
time, sec
(c)
Figure 17 The responses of inverted pendulum system with disturbance
(Tracking)
303
Px2 = Wi 11 h sin X,
where m,, I., J., i = 1,2 denote the point mass, link length, and additional constant inertia
with respect to axis of rotation, respectively. This system not only has highly nonlinearity
but also has coupling effects between two links. Let's assume that the parameters of the
manipulator are given by: Wy = 0.5 kg, m^ = (l+w^) kg, /^ = 1.0 m, /^ = 0.8 m, and J,=J^ =
5.0 kg m^. Now, if we wish to control the manipulator to take a ball with the mass m^ = 3.0
kg from the coordinate (29/2, 0) to (-3"^, 0) and the initial angles arc 6,g= 1 rad, and O^,, =
0.5 rad, then, the desired paths can be exprssed as follows:
Owing to the insensitivity of the fuzzy-sliding mode, the coupling effects between two
joints can be treated as external disturbance. So, we can apply the design procedure indi-
cated in Example 1 to design two FSMCs for each joint independently, no matter whether
these two controllers are identical or not. For simplicity, we use the same fuzzy set
assignments as shown in Table 1 with / = 1,2 ,..., 4, and their corresponding membership
functions are the same as those shown in Figure 4. The inference rules in two controllers
are the same as those listed in Table 2. Now, we select the scaling factors to be
(=1
i=3
with
k,; = Mi a^gn(x)xGS,, Aa^gn{x)xGCS,)xGK,
where i = 1, 2,..., 4, and k = 1 for i = 1, 2, and A: = 2 for i" = 3, 4. Figure 19 shows the re-
sponse trajectories of 6^ and 6^.
5. Conclusions
The proposed^zzv-i'/i(/mg mode regime gives rise to several characteristics as follows:
1. The dynamics behavior of the control system can be to a large extend dominated by
a (fuzzy) sliding surface which is predetermined by the designer. In other words, like the
classical sliding mode control system, the fuzzy-sliding mode control system is also insen-
sitive to plant uncertainty as well as external disturbances.
2. Fuzzification of sliding surface will increase the robustness to the system nonideali-
ties and decrease the chattering.
305
Acknowledgements
The authors would like to express their appreciation to Professor Hsin-Hsiung Lin of Ta-
tung Institute of Technology for his careful reading and editing of the manuscript.
References
[I] G. J. Klir and T. A. Folger, Fuzzy Sets, Uncertainty, and Information, Prentice-Hall:
New Jersey, 1988.
[2] D. Dubois and H. Prade, Fuzzy sets and systems: theory and applications. Academic
press: New York, 1980.
[3] L. A. Zadeh, "Fuzzy sets", Information and Control, Vol. 8, pp. 338-353, 1965.
[4] L. A. Zadeh, "Outline of a new approach to the analysis of complex systems and de-
cision processes", IEEE Tran. on Syst. and Cyber., Vol. SMC-3, No.l, pp. 28-44,
1973.
[5] C. C. Lee, "Fuzzy logic in control systems: Fuzzy logic controller - Part I ", IEEE
Tran. on Syst. Man and Cyber, Vol. 20, No. 2, pp. 404-418, 1990.
[6] C. C. Lee, "Fuzzy logic in control systems: Fuzzy logic controller - Part 11", IEEE
Tran. on Syst. Man and Cyber., Vol. 20, No. 2, pp. 419-435, 1990.
[7] C. C. Lee, "A self-learning rule-based controller employing approximate reasoning
and neural net concepts". Int. J. of Intel. Syst., Vol. 6, pp. 71-93, 1991.
[8] T. J. Procyk and E. H. Mamdani, "A Linguistic self-organizing process controller ",
Automatica, Vol. 15, pp. 15-30, 1979.
[9] E. H. Mamdani, "Application of fuzzy algorithms for control of simple dynamic
plant", PTOC. IEE. Vol. 121, No. 12, pp. 1585-1588, 1974.
[10] E. H. Mamdani, "Application of fuzzy logic to approximate reasoning using lin-
guistic synthesis", IEEE Tran. on Computers, Vol. C-26, No.l2, pp. 182-1191, 1977.
[II] S. Assilian and E. H. Mamdani, "An experiment in linguistic synthesis with a fuzzy
logic controller". Int. J. Man Mech. Studies, Vol. 7, No. 1, pp. 1-13, 1974.
306
[12] Y. Y Chen and T. C. Tsao, "A description of the dynamical behavior of fuzzy sys-
tems", IEEE Tran. on Syst. Man and Cyber., Vol. 19, No. 4, pp. 745-755, 1989.
[13] R. Zhao and R. Govind, "Defuzzification of fuzzy intervals", Fuzzy Sets and Systems,
Vol. 43, pp. 45-55, 1991.
[14] M. Mizumoto and H. J. Zimmermann, "Comparison of fuzzy reasoning methods",
Fuzzy Sets and Systems, Vol. 8, pp. 253-283, 1982.
[15] V. I. Utkin, Sliding modes and their application in variable structure system, Mos-
cow: Mir, 1978 (English translation).
[16] V. I. Utkin, "Equations of sliding mode discontinuous system, I", Automat. Remote
Control, No. 12, pp. 42-54, 1971.
[17] V. I. Utkin, "Equations of sliding mode discontinuous system, 11", Automat. Remote
Control, No. 2, pp. 51-61, 1972.
[18] J. J. Slotine and S. S. Sastry, "Tracking control of non-linear systems using sliding
surfaces, with application to robot manipulators". Int. J. Control, Vol. 38, No. 2, pp.
465-492, 1983.
[19] J. J. Slotine, "Sliding controller design for non-linear systems". Int. J. Control, Vol.
40,No. 2, pp. 421-434, 1984.
[20] K. K. D. Young, "Controller design for a manipulator using theory of variable struc-
ture systems", IEEE Tran. on Syst. and Cyber., Vol. SMC-8, No. 2, pp. 101-109,
1978.
[21] K. K. D. Young, P. V. K. Kotovic and V. I. Utkin, "A singular perturbation analysis
of high gain feedback system", IEEE. Tran. on Automatic Control, Vol. AC-22, pp.
931-938,1977.
[22] K. K. D. Young and H. G. Kwatry, "Variable structure servomechanis design and ap-
plications to overspeed protection control", Automatica, Vol. 18, No. 4, pp. 385-400,
1982.
[23] M. Vidyasagar, Nonlinear Systems Analysis, Prentice-Hall: New Jersey, 1993.
C H A P T E R 11
1 Introduction
Since the basic work by Zadeh in the late sixties [12], [13], fuzzy logic has been
intensively applied in the area of control system theory and design, in order to set up
control environments which do not require the precise knowledge of the mathematical
model of the process to be controlled (see, for instance, [3], [8] and the references
therein cited).
Indeed, fuzzy logic, enabling the definition of sets and membership functions of
a certain clement to given sets by using natural language, appears to be suitable to
deal with all those practical situations in which a certain level of uncertainty in the
description of the system under study has to be taken into account.
T h e design of a fuzzy controller consists of a fixed number of steps. T h e aim
of the controller is that of relating the relevant variables, such as input or error
signals, usually described in a crisp way, to the control action. T h e first step is
therefore devoted to the conversion of the crisp input variables to the fuzzy form by
associating with them the corresponding linguistic values.
T h e core of the controller is represented by a set of linguistic rule. These latter
are evaluated according to the compositional rule of inference, producing a fuzzy
o u t p u t which, once transformed into a crisp variable, enables the generation of the
suitable control action.
In spite of the wide variety of applications of fuzzy controllers to industrial
processes, a certain number of problems, mainly related to the efficient design of the
look-up table containing the fuzzy rules, still remain to be overcome. A noteworthy
a t t e m p t in this direction is constituted by the possibility of combining a controller
design procedure based on fuzzy logic with a less qualitative design approach, such
307
S.G. Tzafeslas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 307-328.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
308
as that inspired to the variable structure systems theory, recently outlined in the
literature [7].
A variable structure controller (VSC) enables the definition of a suitable sliding
manifold on which the state trajectories of the controlled system are kept, from a
certain time instant onwards, in spite of the uncertainties assumed on the process
in question [10], [11]. T h e sliding mcinifold coincides with the intersection of the
discontinuity surfaces associated with the various control signals acting on the plan-
t. As a consequence, the overall control strategy, once the s t a t e trajectory of the
controlled systems lies on the sliding manifold becomes discontinuous.
T h e fuzzy version of the cleissical VSC proposed in [7] refers to single input/single
o u t p u t plants and is characterized by a set of rules which enables the determination,
at each sampling instant t = kAT, k = 1 , 2 , . . . of the discontinuous gain vector to
be introduced in the state feedback loop to realize the conventional VSC scheme.
Indeed, when the plant is single input, by using the VSC approach, it is often
sufficient to guarantee that some simple upper and lowerbounds on system dynamics
are known to attain the control objective with the desired accuracy. T h e control
strategy is a signal, discontinuous on suitable manifolds of the state space, whose
amplitude must be greater than some unknown but bounded function of the system
states and disturbances.
In the single input cetse, even when no upperbound of such an unknown function
is known, a fuzzy controller can be designed since a simple qualitative relationship
between the effect (the distance of state trajectory with respect to the sliding mani-
fold) and the cause (e.g., an overestimation of the uncertainty bounds) can be stated,
and, consequently, a suitable qualitative control action, based on rules of the type
"the control action is too low, then increase it", can be adopted.
It is worth noting, even in simple cases, t h a t this qualitative reasoning cannot
be applied to any control technique. Consider, for example, the stabilization of a
controlled system defined by
ij = xi + u (1)
Therefore, when dealing with the possible use of a fuzzy system to control com-
plex uncertain plants, it is important to choose a conventional (crisp) control strat-
egy which is suitable for being described by means of finite number of fuzzy rules
(the Ziegler-Nichols method for tuning the parameter of a P.I.D. regulator belongs,
for example, to this class of control techniques).
In this chapter, we move from the single input case and focus our attention on
the multivariable control of non linear uncertain systems. In particular, a fuzzy
controller ba^ed on a particular version of multi-input VSC strategy will be dealt
with. To this end, some assumptions on the class of controlled systems needs to
be made. More precisely, we shall confine our investigation to the class of systems
modelled by pth-order controlled differential equations on 3?"". We assume that the
full state is available for feedback purposes and introduce weak a-priori assumptions
in order to guarantee the existence of the solution to the control problem.
In order to motivate our choice, let us consider a controlled m-degrees of freedom
mechanical system
A{q)q = u + fit,q,q) (2)
where q{i) £ 3?"' is the vector of the generalized coordinates, A{q) is the inertia
matrix, f{t,q,q) represents elastic, damping, friction, Coriolis and other torques, us
well as external disturbances.
If additional constraints are not included and no coordinate transformation has
been performed, it is well known that A{q) is symmetric and positive definite. The
robust control of this kind of system, as far as both regulation and tracking problems
are concerned, hcis been addressed by Utkin [10], [11] and, more recently, by Guzzella
and Geering [5] and Slotine and Li [9]. The proposed procedure is worth being briefly
recalled in order to put in evidence how the relevant control structure is suitable for
being inserted in a linguistic (fuzzy) control environment.
To this end, consider the m-vector
V = \s''A{q)s (3)
where A' is a matrix such that K{q, q)q represents the m-vector of centripetal and
Coriolis torques.
As a consequence,
causes the reaching, in finite time, of the sliding surface s = 0 and the fulfilment of
the control objective.
This result is well suited for being dealt with in a fuzzy control environment.
Indeed, suppose that M* is not known in advance. It is clear that any control
component can be represented as
Ui = -jisign{si}
Then, a set of linguistic rules can be individuated relating the effect, the component
Si, and the cause, the choice of the parameter 7,, in order to generate a control
causing the system trajectory to reach the sliding manifold (e.g., "if Si is positive
big, then the increment of 7, is positive big"). Note that this fuzzyfication procedure
might fail if different approaches, like feedback linearization [6], were used, since.
311
in such cases, the control law strictly depends on the equations which describe the
controlled system dynamics.
On the other hand, this approach cannot be pursued if the generalized coordi-
nates are subject to a non linear trasformation
X = F{q)
X = J{q)q (9)
where J{q) is the Jacobian matrix. In that case, the differential equation governing
the evolution of the system in the new coordinates, assuming the transformation to
be invertible, becomes
X = J{q)q + J{q)q
A{q)[J(q)\-'x^Y{t,q,q)^u (11)
This equation is similar to (2), but this time the matrix A{q)\J(q)\~^ does not
result symmetric and positive definite (j4(9)[J(g)]~' is however assumed to be in-
vertible). Therefore, the previous control procedure cannot be applied.
In the next sections it will be shown how it is possible to use a suitable control
procedure, the so-called simplex sliding mode control method, in order to deal with
systems of the type (11) in a fuzzy context.
This chapter is organized as follows. Section 2 contains a presentation of the sim-
plex method for the control of multi-input uncertain systems. Section 3 is devoted
to the description of the fuzzy control evironment in which the problem considered
is inserted, as well as to the presentation of an illustrative example to complement
the theoretical discussion.
Starting from their results, a new control algorithm is proposed which proves to be
suitable to constitute the control philosophy underlying the fuzzy controller which
will be presented in the next section.
Consider a system described by the following differential equation
where x € 3i", u € 3?"*. We choose a manifold {s(x) € 3?"" : ^(x) -Cx -0} so that
when the system trajectories lie on it (i.e., i(x) = 0, if s{x) = 0), the behaviour
of the system satisfies a prespecified control obiective. The control signal role is,
therefore, that of forcing the system trajectories from any initial state to reach the
manifold in finite time. Mathematical and applicative details of such an approach
can be found for instance in [2], [4], [10], [11], [14].
In order to provide a control strategy accomplishing the above control objective,
let us consider the dynamic behaviour of the m-veclor s{x) described by
where the matrix [CB{x)] is assumed to be invertible, but neither symmetric nor
positive definite.
The traditional way of extending to the multi-input case the basic VSC strat-
egy consists in the selection of the control signals according to a componentwise
procedure. This means that each component of the vector s{x) is related to the
corresponding time derivative i(x) by the so-called reaching condition, i.e.,
that is
s{x) = —rsign{s{x)} (15)
where F := diag{%}, i = 1 , . . . , m
The control vector results to be,
However, the reaching condition can be expressed also in alternative ways, giving
rise to different control switching logics. Among these, the simplex method will be
exploited in this chapter in order to transfer the conventional sliding mode control
method into a fuzzy context. This procedure can be briefly described as follows.
1. Choose in 3?*" a simplex of vectors [ui, . . . ,Um+i] containing the origin, i.e.
m+l
y = lCB{x)]-h(x)
that is
y = {CB{x)]-'CA{x)
where -4,(x), Bi{x) are the rows of A{x), B{x), respectively. Then,
y - g{x,s,u) + u (19)
If the control is chosen as u = Uj, when y G ilj, then the origin y = 0 is asymptoti-
cally stable. Moreover, the convergence takes place in finite time and an upperhound
oj the convergence interval t' — to is given by
I -ao
R e m a r k : the control is unique for any region, and any region is univocally deter-
mined by the choice of the simplex of control vectors.
( m+l
R e m a r k : This control strategy is not unique because for each fJj we can choose
one control vector among the set {«,}, i ^ j . To solve such an ambiguity in [1] it is
proposed to choose a control sequence of the type
f Ui+i i = 1, . . . , m
ui 1= m + l
Once this choice has been made, a unique correspondence between the choice of the
simplex of control vectors and the partition of the y - s t a t e space is stated, and a
unique relationship between any region of the y state space and one of the control
vectors is established. All this facts require the perfect knowledge of the matrix
[CBix)].
315
Q Q.
m+i 1
!
under the assumption previously indicated for g^x^s^u). If the control is chosen
as u = — Z!{^t'>!i V'"i) Vi ^ 0) JTi^t'tViVi ^ If w/ien j / G ilj, then the origin is
asymptotically stable and the convergence takes place in finite time.
P r o o f . Consider y = J2^\-tj A,Ui, and, for the sake of simplicity, assume t h a t A,-
are all greater than zero. Then, define
m+l
u = - Yl V."M <P. > 0
u = -[U,]4> (21)
A = [U,]-' y {22)
318
where Uj = — J^J^t.^i/'•"•' f^> — ^' ITI^t.^^j/*• > ^> since the control vectors u^'s
form a regular simplex. Therefore, defining
it yields
- g{x, s, u) = [U,] K, + k, [[/,] M, (24)
so that, from (22),
A = m-'{-[UAK,-kAU,]M,-[U,]<l,}
= -Kj - kjMj - <f> (25)
a = - E ^' + ^i h - E /'i - E V.
i=l \ i=l,i5^J / •=!.;#>
< -(1+ao) (28)
this implies that a(-, <) = 0, for any t > t', with t' = o(j/o, 0)/(l + QQ)
Then, we are in a position to state our main result, which generalizes the previ-
ous considerations about the possibility of using a regular simplex of control vectors
319
when the state space is partitioned on the basis of an irregular simplex of vectors.
where
causes the finite time reaching of the origin of the y-state space.
a fuzzy controller solving the control problem in question can be based on such a
theorem.
When the transformation matrix is unknown, it is not possible to quantify the
actions to be applied to the regular simplex in order to properly re-positionate the
irregular simplex with respect to the previous one. Such actions consist in rigid
rotations of the simplex and variations of the modula of the component vectors.
T h e n , it seems natural to establish a qualitative relationship between the action to
perform and the difference between the actual and the expected behaviour of the
system.
Suppose to choose, as a switching logic, that expressed by Theorem 1, with the
difference that now vector y is substituted by vector s. Because of the transformation
operated by matrix [CB(x)], it is not ensured that 5 moves towards the origin of
the s-space. On the other hand, if the disturbance g{x,s,u) in (19) is small and
a regular simplex of vectors with suitable amplitudes is applied, the trajectory of
the unknown vector y is approximately a straight line parallel to the corresponding
applied control vector, according to (19). While y is unknown, the vector s(x) =
[CB(x)]y is measurable. This latter moves along a trajectory which, as long as
[C/?(j;)] were constant, would result in a straight line as well, if evaluated in a short
time interval, the angle between this line and the direction of the applied control
vector being the opposite of the angle of which the regular simplex needs to be
rotated in order to have a motion of s directed in the desired way. In this situation,
defining As{t) := s{t) — a{t — At), the action to be applied to the regular simplex is
just a rotation of an angle 0(t) = arcos[< Aa{t),u > /(||As(i)||||u||)].
Since g{x, s, u) is not known and [CB{x)] is not a constant matrix, the above pro-
cedure cannot be applied in a quantitative way, unless some a-priori known bounds
on g{x,s,u) and [CZ?(x)] are available. Also in t h a t case, as previously outlined,
some problems remain to be faced. Indeed, when the uncertainties are described in
terms of interval vectors and matrices, the possibility of losing the invertibility of
matrix [CZ?(a;)], as well as the necessity of performing a highly conservative control
design, should be taken into account.
On the contrary, by using a qualitative approach, based on the experience that
the amplitude of the control vector is required to be sufficiently high to dominate
the unknown disturbance g(x,s,u), and that the rotation angle 0{t) has to be tied
to the angle (f> between As{t) and u for sufficiently small At, it seems possible
to overcome these drawbacks. It should be noted t h a t , according to the results
expressed by Theorems 3, 4, the ideal value of the angle 0{t) is not unique, but can
321
be expressed as an interval of values such that for any 0(t) belonging to that interval
t h e finite time convergence of s to the origin takes place. This implies that when a
contraction of the norm of s occurs (i.e., A||s(<)|| = ||'S(OII - ll^(' — ^011 '^ negative)
the positioning of the simplex of control vector could be stopped.
In order to design the fuzzy controller according to the previous considerations,
we proceed in the standard way, selecting the suitable variables to be used a.s terms
in the premises of the rules. Then, the design procedure can be performed according
to the following steps:
3. Determine the membership function corresponding to each rule and the overall
membership functions associated with the output variables.
4. Deffuzzify the output variables obtaining the crisp values of the control actions.
The previous list can be particularized to the control case considered within this
chapter, in order to obtain a fuzzy simplex sliding mode control scheme for m u l t i -
input systems.
T h e input variables to be considered can be the amplitude of vector s, the vari-
ation of the amplitude (i.e., vector A||s||), and finally the rotation matrix between
vector A | | s | | and the currently applied control vector (w = Uj if s £ ilj). While ||s||
is available at any time instant, the other quantities are supposed to be evaluated at
discrete t i m e intervals kAt, k = 1,2, Note t h a t the sampling interval At could
be regarded as an input variable, as well.
T h e o u t p u t variables are the rotation matrix Rk which needs to be applied to
the previous control signal, which we call Wk-i, in order to determine the current
control Wk according to the following relationship
Wk = RkWk-l
= RkDkU
Do = I, Ro = I
Dk = Rk-iDk.i, k = l,2,... (29)
322
and the amplitude Au^ of the components of the simplex of control vectors. Both the
input and output variables are fuzzified by associting with t h e m suitable fuzzy sets
corresponding to a prespecified number of levels, with proper membership functions.
T h e above considerations can be clarified by means of a simple example relevant
to a two-input system described by the following state equation
s = Hu+ g (30)
ith
- 0 . 5 ( 1 +7sen*i/(a)) -0.866
H-' =
-0.866(1+7senV(.s)) -0.5
a {I +/3sen^i/(5))
9 =
a {I + 0sen^i/{s))
with iy{s) := (1 + s^ + ^2)'^^- Note t h a t if 7 :^ 0, matrix / / " ' is non symmetric.
T h e control vectors constituting the initial regular simplex are
T h e proposed fuzzy controller is based on the rules shown in Table 1 which refers
to the membership functions relevant to the input variables ||s||, Ajl^jl, 0, and the
o u t p u t variables 9 and Au indicated in Fig. 4, 5, 6 and 7 and 8, respectively.
T h e behaviour of the controlled variable 3, starting from the initial condition
s^ = [2, 3], is illustrated in Fig. 9 under the assumption that 7 = 0, o = 0 and
constant modulus of the applied control vector (i.e., A M = 0, which means that the
last column of the rule table is not considered). In Fig. 10 the behaviour of vector 6
is reported assuming, instead, that 7 = a = 1, ^ = 0.5, and allowing the adjustment
of the modulus of the control vector u.
In both cases the behaviour is highly satisfactorily, since the convergence of the
controlled variable to the origin of the state space is obtained with good convergence
rate in spite of the uncertainties on the non linear model of the controlled system.
4 Conclusions
In this chapter a fuzzy control approach is presented in order to deal with multi-
input uncertain non linear systems.
323
phi
20 25 30 50
control magnitude
I
a
-1 0 1
Componeni 2
I ! ! I
—1 I
- 5 - 4 - 3 - 2 - 1 0 1 2 3 4 5
Coinponenl2
The starting point is the choice of a control algorithm which does not result to
be too strictly dependent on the perfect knowledge of the mode! of the system to be
controlled. T h e most effective approach in this sense, among those proposed in the
literature, is the so-called simplex sliding mode control method, in which, however,
the transformation matrix acting on the control vectors is required to be perfectly
known.
In the present chapter it has been shown that this requirement can be relaxed
as long as a non conventional control approach is applied. T h e resulting algorithm
turns out to be characterized by the application of simple operations on vectors, like
rotations and increment of their modula, which can be easily translated in terms
of fuzzy rules, giving rise to a new fuzzy controller which enables to solve in a
satisfactory way the crucial problem of controlling uncertain non linear systems.
327
If ||.|| is ... and All^ll is ... and 4> is ... then 9 is ... and Au is ...
BIG POS POS HIGH HIGH BIG POS
BIG POS POS MED HIGH SMALL POS
BIG POS POS LOW MED BIG POS
BIG POS POS EZ LOW BIG POS
BIG POS EZ HIGH HIGH
BIG POS EZ MED MED
BIG POS EZ LOW LOW
BIG POS EZ EZ LOW
BIG POS NEG HIGH MED
BIG POS NEG MED LOW
BIG POS NEG LOW LOW
BIG POS NEG EZ EZ
MED POS POS HIGH HIGH
MED POS POS MED MED
MED POS POS LOW LOW
MED POS POS EZ LOW
MED POS EZ HIGH MED
MED POS EZ MED LOW
MED POS EZ LOW LOW
MED POS EZ EZ LOW
MED POS NEG EZ
SMALL POS POS MED
SMALL POS EZ LOW
SMALL POS NEG EZ
References
[l] Baida, S. V., and D. B. Izosimov 1985, Vector method of design of sliding
motion and simplex algorithms, Automation and Remote Control, 4 6 , 830-837.
[2] Bartolini, G., and T. Zolezzi 1985, Variable structure non linear in the control
law, IEEE Transaction on Automatic Control, 3 0 , 681-684.
[3] Chang, S., and L. A. Zadeh 1972, On fuzzy mapping and control, IEEE Trans-
action on Sys., Man, Cyber., 2, 30-34.
[4] DeCarlo, R. A., Zak, S. H., and G. P. Mattews 1988, Variable structure control
of non linear multivariable systems: a tutorial, Proceedings of the IEEE, 76,
212-232.
[5] Guzzella, L., and H. P. Geering 1990, Variable structure controllers for robots,
in Deterministic Control of Uncertain Systems, A. S. I. Zinober Editor, Peter
Peregrinus, London.
[7] Kung, C. C., and S. C. Lin 1992, A fuzzy sliding mode controller design, Proc.
IEEE Int. Conf. on Systems Engineering, Kobe, J a p a n , 608-611.
[8] Lee, C. C. 1990, Fuzzy logic in control systems: fuzzy logic controller, IEEE
Transaction on Sys., Man, Cyber., 2 0 , 404-435.
[9] Slotine, J. J. E., and W. Li 1991, Applied non linear control, Prentice-Hall,
Englewood Cliff, N.J..
[10] Utkin, V. I. 1978, Sliding modes and their application in variable structure
systems, MIR Publishers, Moscow.
ABSTRACT Pure numerical computations have been playing a dominant role in control
systems. With the increasing complexity of the systems under control and diversity of
control tasks to be handled, it became obvious that the concept of this numerical platform
mastered so far cannot be sufficient in many situations. Some techniques of Artificial
Intelligence (such as e.g., control expert systems) on one hand, and nonparametric
parallel optimization (i.e. structures such as neural networks) on the other, have been
already studied as possible computational frameworks. Following this avenue, we will
investigate the problem of knowledge representation and information processing through
the use of an integrated platform composed of neural networks and fuzzy sets. This
unified approach takes advantage of flexibility of fuzzy sets used in representing control
knowledge and learning capabilities residing within neural networks. The proposed
topology of the controller includes an internal logic-based part describing control rules
and an interface component "translating" these rules and their activation levels into pointwise
control actions using either preselected (fixed) or trainable control values. With regard to
this feature the learning of the controller is reinforced or is carried out in a fully supervised
mode.
1. INTRODUCTION
hand and visualize essential relationships between the variables. Rather than concentrating
on mastering linear or nonlinear local relationships between all the variables again handled
in a numerical pointwise manner, it might be useful to move toward a higher level of
generality. This level can tackle with information granules defined for involved variables.
By this type of information aggregation we mean a collection of associated values which
are coupled to a certain degree within the given context of the specified control task
(control situation) and labelled for further use. From a conceptual standpoint we can
indicate three distinct approaches already reported in hterature.
(i) the first one, usually residing within the area of Artificial Intelligence (AI), refers to
purely symbolic computations. The aggregates studied there and being almost of a
Boolean character, are named symbolically and used afterwards. Possessing a
certain syntax, these aggregates being processed as purely symbolic quantities give
rise to a so-called Qualitative Modelling (Q^) and Qualitative Control, (QC), cf.
[2] [3]. Obviously, the very fundamental nature of the AI techniques does not leave
any room for numerical processing. In many control situations this approach cannot
be fully accepted mainly due to significant and irreversible loses of information
caused by this symbolic quantification,
(ii) the second approach deals with aggnsgates having a form of closed numerical intervals
(as well as hypercubes forming their multidimensional counterparts). These quantitites
create a basis for the development of a certain cognitive perspective at which the
control processes are looked at. In this setting, the existing techniques of interval
calculus widely apply, cf. [5]. This way has been exploited in the past mainly for
purposes of performing much more thorough sensitivity analysis and modelling
systems using sel-theoretic perspective (e.g., set-theoretic models with unknown-
but-bounded disturbances). Within the above framework numerical processing
capabilities are pnsserved. The increased generality is obtained by adjusting (increasing)
the size of the information aggregates (granules) through a suitable increase of the
widths of the intervals. Definitely, the control algorithm becomes robust due to the
introduced interval-valued character of the variables. The obvious shortcomings
manifesting here is that border values of the intervals might be difficult to establish.
The interval-based knowledge representation one can easily control the level of
specificity (generality) of the concepts. It is, however, not capable of distinguishing
between the values (points) belonging to the same interval.
In the third approach we will consider a combination of (I) and (ii) that is carried out
as follows.
332
(iii) Referring to the lack of the focussed concentration (selective discrimination) occurring
in the previous framework we will replace intervals by fuzzy sets. Fuzzy sets can
eliminate the above mentioned disadvantage due to their membership functions acting
as elastic constraints and highUghting a degree of membership to the given granule.
They can carry out a detailed discrimination process within the same scope of
cognition therefore properly addressing the question of the selective focus of attention.
The higher the grade of membership, the higher satisfaction of the given property
(constraint). Subsequently, some elements in the granules become strongly elevated
while the others substantially suppressed. An immediate incorporation of fuzzy sets
in QSi and QC as introduced e.g. in [2] allows to cope with the semantics of the
conveyed terms and take care of well-defined syntax features.
Fuzzy sets, as generally set-theoretic oriented concepts, are not extremely powerful
at numerical computations. The collection of the existing and commonly utilized operations
is not well suited for effective and fast calculations. These operations are not capable of
coping efficiently with a probabilistic orrepetitivecharacter of information.
In designing fuzzy controllers, the general control strategy is synthesized on a basis
of some individual and scattered pieces of knowledge such as e.g., the previous records
of control acdons or more or less complete control protocols. Sometimes these protocols
are biased by a lack of precise data which are replaced by some linguistic statements.
These situations call for the use of fuzzy sets as a relevant formalism to cope with
linguistic and/or incomplete information.
There are many diverse conceptual as well as applied platforms which can directly
benefit from the methodology of intelligent control. The essence of the principle becoming
an essential part of our scheme lies in using fuzzy sets to address essential representation
issues and afterwards make use of neural networks in order to realize all relevant
computations. Furthermore we should be aware that the appliedrepresentationscheme
has a direct impact on efficiency of learning by focussing its field of performance. It
means that by accepting the linguistic labels (i.e. nonlinear membership functions) a
preliminary nonlinear preprocessing takes place. This stage highlights some potentially
fruitful andrelevantregionsof search to be explored within further detailed learning. In
this way the optimization search is not solely guided internally by a prespecified universal
performance index (as it happens in almost all the existing learning schemes). One has to
be aware that the "universal" performance index makes the learning insensitive to the
problem specificity.
The concise summary of the relationships between neural networks and fuzzy set
333
Fuzzy Sets
(knowledgerepresentationlevel
of control)
Neurocomputations
(conputational aspects
of control)
N
by forming a union of their Cartesian products of the form u (Aj x Bj). The inference
n= l
is based on a standard max-min composition operator [6] [10]. This way of aggregation
and inference has been broadly implemented giving rise to many application-oriented
versions of the fuzzy controllers. One should be aware that this design indicates a
strong resemblance of the fuzzy controllers to associate memories that utilize the
Hebbian-like encoding and correlational recall.' Furthermore, as studied in [6],
the knowledge incorporated into the structure of the controller has to be consistent.
Even the fuzzy controller can tolerate some discrepancies of the rules, starting from a
certain level, these shortcomings tend to affect the performance of the controller. If
an appropriate learning mechanisms were in place these inconsistencies would be
easily ruled out The set-theoretic operations (like max and min) are not capable of
eliminating some exceptions and inconsistent facts. The facts, even being very rare
and quite meaningless, could easily override a vast number of consistent observations
that have been already "coded" into the structure of the controller.
(ii) In the domain of neural networks the learning capabilities of them are significant.
Usually their representation features are not enhanced. The first problem arises
when input variables refer to different physical variables and, what makes the entire
story worse, these variables can take on numerical values from distant ranges (say
10"^ as opposed to 10*). The learning using the raw data is almost impossible from a
numerical point of view. This calls for normalization and/or discretization of the
input variables as a significant preprocessing phase.
In the next section we will focus on a general topology of intelligent controllers
enhancing the general methodological links as visualized in Fig. 1. In Section 4 we will
set up a general architecture of the controller and proceed with underlying learning
schemes.
In this section we will study a general scheme of the controller, identify its basic
functional blocks and finally come up with its detailed scheme including underlying
mathematical formalism.
'To make this analogy clear lei usreplacethe union (u) by summation and the minimum
operator (x) by a standard product operator.
335
Let us start with a multivariable control structure comprising of several input variables
(variables of the system under control) and control variables. Fuzzy sets (FS) and
neurocomputations (NN) are directly involved in different parts in the diagram as shown
in Fig. 2.
system
variables
control
variables
Fig.2. Functional scheme of the controller involving Juzzy sets and neural networks
Its individual blocks are described as follows. Firstly, all the systems variables are
preprocessed by fuzzy sets contributing to the cognitive perspective suitable for this
specific variable. [7] [8] This nonlinear transformation situated there enables us to come
up with a vector of numbers in [0,1] intervals and build all the inputs of the neural
network. By this mapping mechanism, all physical quantities of the variables are translated
in a nonlinear manner into the corresponding dimensionless quantitates situated in [0,1].
Note that this translation provides a mechanism of nonliunear data quantization.
Subsequently, the outputs of the network indicate degrees of activation of control labels
of all control variables. This request calls for another conversion block (C) the role of
whose is to convert any information coming from the processing block into a single
numerical quantity.
The more detailed structure of the controller giving a better insight into the functional
blocks and their performance is shown in Fig. 3.
336
input data
- pointwise defined
-interval-valued linguistic labels single-layer linguistic labels Conversion block
— > • of system -x->- neural -^ of control (pointwise control
- fuzzj sets variables network variables representation)
- Incomplete
It is worth noting that the topology of the network is influenced by the specification
already done with regard to the representation part. The number of the layers and nodes
of the neural network depends upon:
(1) the number of the linguistic labels associated with the individual system variable; the
number of the input nodes depends on all the linguistic labels defined for the
variables. Each input node correspond to a single linguistic label. Denote their
number by "n".
(2) The number of the linguistic labels describing control variables uniquely determines
the number of hidden nodes. Let its numbers be equal to " h". They form a hidden
layer of the network.
Before proceeding with learning to be completed by the network, we will briefly
describe the way in which the input nodes are activated as well as highlight the manner in
which thefinalnonfuzzy value of control is determined.
Considering the linguistic label as uniquely described by its membership function A,
where A(x) stands for degree of membership at element x, we express the level of
activation of A caused by any fuzzy (or nonfuzzy) information X accordingly
sup [min(A(x), X(x))] (1)
where supremum is taken over all the elements of the universe of discourse in which A
and X have been already defined. Note that the above expression is nothing but the
possibility measure of X taken with respect to A [13]. It plays a role of a scalar index
pointing out a degree to which A is "fired" (activated) by X. As already mentioned, the
preprocessing block can handle various forms of informadon:
337
Possibility
1.0
Obviously, if x„ e [x-, x+], then this type of information implies higher activation
of the corresponding node in comparison to that generated by the pointwise datum.
The output layer realizes a nonfuzzy control action. For the t-th control variable, u„
it is computed by specifying a nonfuzzy representation of fuzzy sets of control (expressed
by the hidden layers) and treating them as the connections from the hidden layer to the
t-th node of the output layer. The operation realized at this output is expressed as
ut = (utiyi + Ut2y2 + - + uthyhVCyi + y2 + - + yh) (2)
where Utj are connections between the t-th output node and the j-th node of the hidden
layer while y,, yj, .... y„ are the signals generated by this layer.
338
- when all the nodes in (2) are activated, the resulting control u, becomes a mean
value of the representatives, (connectives) uy
h
"j=i
In this section we will discuss a detailed architecture of the neural network. Essentially,
it consists of two components. The two first layers are constructed by the two types of
logical computing nodes (neurons). These types of neurons will be referred to as logical
neurons. The last layer of the network is built with the use of a nonlinear mapping
realizing (2) which is also viewed as a component carrying out a transformation from the
level of the logical relationships to the physical level of the control variables. For the
sake of clarity in the remainder of the section we will discuss the controller with a single
output (the extensions to the multidimensional case are quite straightforward).
The overall architecture of the controller is illustrated in Fig. 5
It consists of two hidden layers. The first hidden layer is composed of AND neurons
while the other utilizes OR-type processing units. Essentially, these two layers encapsulate
a collection of "if-then" control statements. The AND neurons summarize all the inputs
(subconditions) by ANDing them. The conclusion (control) part is aggregated ORwise.
The detailed formulas describing the network are summarized below. Note that the
realization of the AND neurons involves the usage of the triangular norms (t- and s-norms)
that result in building a s-t composition.
339
hidden2 output
input
transformation
condition laver conclusion layer
output-hidden layer:
yiUi+y2U2+...+yhUh
' yi+y2+...+yh
hidden^-hiddeiij layer:
yi = .S (wij t Zj)
j=i
i=l,2,...,h
hidden,-input layer:
Zj = jT (vji s xi)
j=1.2,...,p
340
Now we will address the problem of supervised learning. This mode of learning
implies the use of a collection of the training samples organized as inputs x,, x^, ..., x^
and associated outputs viewed as target values t,, tj, .... t^. The learning pertains to
updates of the connections (w,v,ii) via a steepest descent method,
A_connections= -a :r ^^-^
oconnections
The Mean Square Error (MSE) will be used to optimize the network:
N
Q= X[tk-N(xk)F
where N(x^) (= y^) denotes the output (control) produced by the network and obtained
for the input x^.
Two schemes of learning will be analyzed:
(i) reinforcement learning. [ 11] In this type of learning the logic-based part of the
network will be modified while the connections ui, U2, ..., Un between the hidden
and output layer remain unchanged .
The connections W|j are adjusted on a basis of the gradient of Q, .
3w
The detailed formulas read coordinatewise as
Wij = Wij + a ( t - N ( x ) ) - ^ - ! ^
OWij
with
and
5 ^ = 5 — ( T (Vii s xi)
ii) All the connections of the logical part of the network as well as the control values
ui, U2, ..., Uh are trained. The values u; play now a role of "optimal" prototypes
similar to that obtained in Adaptive Vector Quantization (AVQ), cf. [9]. The learning
process itself reminds the search of an optimal partition of the control space U (in
which all the control variables are situated). The resulting general learning scheme
is the same as already presented for (i) with an additional update scheme describing
a sequence of modifications of Uj's,
Ui = U i + a ( t - N ( x ) ) ^ p ^
dui
i = 1, 2, ..., h
with
9N(X) ^ yj
aui ^
5. CONCLUSIONS
The paper clarifies the role of coherence between the processes of knowledge
representation ad its pro.essing that is necessary to develop intelligent controllers. By
changing the scope of interest and the level of precision of the linguistic labels, we can
easily develop a suitable perspective required to cope with complexity of the system and
specificity of the control actions. Furthermore, as it has been clarified, by changing the
number of the labels we can easily switch from almost symbolic level of processing (very
few labels representing a very coarse space partition) to the very fme and almost pointwise-
like numerical grains of information. At this point the processing of control knowledge is
342
worked out on a numerical level with the aid of logic-oriented neural networks. Two
learning scenarios are also studied. The fuzzy set interface represented by the collection
of linguistic labels is advantageous from at least three major reasons,
the controller can handle precise, incomplete, or noisy data pertaining to system's
variables.
the connections of the neural network can be easily interpreted as a qualitative
conu-ol protocol involving conditional statements of the format "if state the control".
These statements are generated in an automatic manner thus the scheme can be
viewed as an efficient tool for automatic rule generation or adjustment of some
preliminary sources of knowledge. This may allow us to avoid a serious botdeneck
occurring in many knowledge-based systems and related with the tedious and
error-prone knowledge acquisition processes,
finally, the labels heavily influence the topology of the network determining the
number of its nodes in the input as well as in the output layer. Therefore some
relationships between information granularity and learning capabilities of the conU-oUer
as well as its robustness can be clarified and a reasonable trade-offs established. In
this manner we can talk about focussed (external) learning implied by fuzzy sets.
ACKNOblLEDGMENT
REFERENCES
[1] K.J. Astrom, J.G. Anton, K.E. Arzen, "Expert Control", Automatica, 13, 1986,
277-286.
[2] B. D'Ambrosio, Qualitative Process Theory using Linguistic Variables, Springer-
Veriag, New York, Beriin, 1989.
[3] K. Forbus, Qualitative Process Theory Ph.D. Thesis, MIT, Cambridge, Mass.,
July 1984.
[4] IEEE Control Systems Magazine, (Special section on neural networks and control).
Vol. 8, no. 2, 1988.
[5] R. Moore, Interval Analysis, Prentice-Hall, Englewood Cliffs, N.J., 1966.
[6] W. Pedrycz, Fuzzy Control and Fuzzy Systems, Research Studies Press, J. Wiley,
This page intentionally blank
343
1. INTRODUCTION
The field of parameter estimation goes back to Gauss (1795) who has presented the
classical least squares method in his book : "Theoria Motus Corporum Coelectiwn". Least
squares was then studied by Legendre (1806) in his book: "Nouvelles Methodes pour la
Determination des Orhites des Gametes" . In Gauss' own words : "The most probable
value of the uknown quantities will be that in which the sum of the squares of the
differences between the actually observed and the computed values multiplied by numbers
that measure the deegre of precision is a minimum".
Since then, an enormous effort was made by scientists and engineers to develop
further the method and apply it to sophisticated situations (least squares in function spaces,
least squares filtering, etc.) and applications of industrial and real-life nature.Regarding the
fuzzification of the least squares problem, the authors have picked-up the work of the
Hungarian mathematician Celmins [1]. He has developed an elegant but quite involved,
computationally, method along with a software package (called COLSAC) that implements
it. The authors present an alternative (less general) least-squares method where use is
made of the concept and the algebraic properties of the so called L-R fuzzy numbers [2].
Minimizing their distance with respect to the unknown parameters is a classical
minimization problem and has a small computational demand. The chapter continuous by
formulating and solving the fuzzy state estimation of a particular discrete time state-space
model with fuzzy disturbances and initial conditions [3-7]. Similar problems have been
considered by several authors under various mathematical formulations.
For example, it is worthwile to mention the work of Lee [8] where a state estimation
problem is solved for a class of distributed-parameter systems with uncertain parameters.
The parameters are assumed to be arbitrary time functions known to be in a closed and
bounded region. The resulting estimation algorithm provides an assured accuracy and a
"guaranteed" error estimator which gives an upper bound of the estimation error for any
allowed variation of uncertain parameters [9]. Other works on fuzzy estimation include
[10-20].
The structure of the chapter is as follows. Section 2 starts with a general discussion
of the fuzzification of the classical least squares problem, proceeds to a brief review of
Celmins' fuzzy least squares method, and goes on to the formulation and solution of the
authors' fuzzy least squares problem. Section 3 deals with the state estimation problem of
347
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 347-368.
© 1994 Ktuwer Academic Publishers. Printed in the Netherlands.
348
where ej=Yj-yj is the ith error and Cj, i=l,2,...,n are given weighting coefficients. The
optimal estimate is found by solving the equation dE/da=0 , i.e
' dE
a: -r-=0
da
If (xj.yj) are normally distributed and f(.) is linear, then the least squares estimate of
a coincides with the most probable estimate of a, since the total error E is the negative
exponent in the probability distribution of a given (xj,yj), i=l,2,...,n.
Now we shall try to insert "fuzziness" in the above classical least squares problem.
Fuzzines can exist either in the data of the problem or in the parameters of the model.
Before proceeding to the formal derivation, let us give the following real example.
A student tries to make his reading plan in order to take the exams of the "fuzzy
control" subject. From his own experience he knows the following:
-With "about" 2 hours/day study he solves "about" 2 problems
-With "about" 3 hours/day study he solves "about" 4 problems
-With "about" 4 hours/day study he solves "about" 7 problems
-With "about" 5 hours/day study he solves "about" 11 problems
-With "about" 6 hours/day study he solves "about" 13 problems
Question : If the student devotes "about" 3.5 hours/day for his study, how many
problems will solve?
This problem is actually a problem of fuzzy estimation. In relation to this problem
one can observe the following:
(i) The student (the human) is a dynamical system that receives stimuli (inputs) and gives
replies (outputs). Here both the inputs and outputs are fuzzy (they contain the word
"about"). Thus we have a fuzzy system and we want to estimate its model,
(ii) The human replies very easily to questions (problems) that involve fuzziness, even if
he knows nothing about theories ( e.g. the least squares theory).
A plaussible human answer to the above question is that: "With "about" 3.5 hours
study per day the student will solve "about" (more or less) 6 problems".
349
y=a+bx
dE/d("about-a") ?
To overcome the above difficulties one can follow two approaches; namely (i) to
reduce the problem into a classical one, (ii) to keep the fuzziness and treat it directly
throughout the problem solution. When using the first approach the fuzzy problem is
approximated by a classical one, the solution of which is presented as an approximation to
the solution of the original fuzzy problem. In the second approach the fuzziness is not
eliminated but kept throughout the solution. Thus the question arises: What is better, an
exact solution to an approximate problem or an approximate solution to an exact problem?
Here we shall concentrate on the second approach.
{ W;IICxll;F;W=IICxl|2}.
where llx-alla=lx-al/s^ is the associated norm, a is the dominant value, and Sjj is the spread.
If A is a fuzzy vector then
HA(x)=max(0, l-[(x-A)Tp^-l(x-A)] H
= l-min{l, [(x-A)Tp^-l(x-A)] ^^}
= l-min{l,IIX-A'llA)
where the norm is ||X-A i|j4^ = [(x-A)Tp^-l(x-A)]-^ • with A vector of dominating values,
and P ^ diagonal matrix with elements the spreads of sdj.
Pictorially a fuzzy 2-dimensional vector with conic membership function is shown:
H^^
1 -•
Celmins defines the distance D(A,B) between two fuzzy vectors A and B:
D(A,B)=minmax{h(X,A),h(X,B)}
where h(.,.) is a metric between a classical vector X (X^ R"^) and a fuzzy n-dimensional
vector A.
Celmins' formulation of the fuzzy least squares is then the following:
Minimize
i=l
subject to
L((m-x)/a) , x<m
{ R((x-m)/p , x>m
where m is a classical number, a and P are parameters, and L(.), R(.) are functions of
special type (satisfying the conditiqns of Def A.l in the Appendix). Symbolically an L-R
fuzzy number M is denoted by :
M=(m,a,P)
352
Now, let F(R) the set of all fuzzy L-R numbers that are defined on R. On the set
F(R) one can define a linear structure, since the following properties hold:
a) (m,a,P)+(n,Y,8)=(m+n,a+Y,P+6)
p) t(m,a,3)=(tm,ta,t3) , t>0
Y) t(m,a,p)=(tm,t3,ta) , t<0
Also, on the set F(R) one can define a norm d[.,.]2 via which the distance of two
fuzzy L-R numbers Xj=(xj,aj,3j) and X2=(x2,a2,f^2) can be determined:
d[Xi,X2]2=(xi-X2)2+(xi-X2-(ai-a2))2+(xi-X2-(3i-|32))2
The new formulation of the fuzzy least squares problem is then the following:
GIVEN: Fuzzy R-L data {X-^X{) , i=l,2,...,N where XpCxj.cj.di), Yj=(yj,ei,fi) and
a linear fuzzy parametric model Z=a+bX where a,be R or F(R)
FIND: The parameters a,b so as to minimize the total error
N
^^d[Z„Y-f
E=
i=l
where Zj=a+bXj, and d[Zi,Yj] is the ith error (distance of the L-R fuzzy numbers Zj and
Yi).
It is a model, that when b in Z=a-t-bX is an L-R fuzzy number, then the product bX
is only approximately L-R fuzzy number [2], i.e .
(m,a,P)- (n,Y,8)=(inn,mY+na,m8-i-nP).
Also, if the L-R numbers used are of the special type (m,a,a), then the
computational load is reduced considerably.
In the following we shall treat three cases, namely;
CASE A: Model Z=aH-bX , a,be R , Xe F(R)
CASE B: Model Z=A-fbX , A,Xe F(R) , b e R.
CASE C: Model Z=A-i-BX ,A,B,XG F(R).
and so
d[Zi, Y/=d[a+bXi, Yi]^=d[(a+bxj, bdj, bci),(yi, e;, f;)]^
2 2 2
and the parameters a and b are determined by classical minimization of the real valued
function E=E(a,b), i.e
Parameter a: ^ = 0
9a
Parameter b: M ^ = 0
do
The resulting algebraic system of two equations in the two unknown a and b is:
N N
3Na+b5^(3xi+di-Ci)=5^(3yi+fi-ei)
i=i i=i
N N
aJ(3xi+di-Ci)+b2[x^+(xi-c/+(xi-d/]
i=i i=i
N
=5^[(Xi-Ci)(yi-ei)+(Xi+di)(yi+fi)+Xiyi]
i=l
E=2_^{(a+bXi-yi)\[a+bXi-yi-(bdi-ei)]^+la+bXi-yi-(bCi-fi)]^}
i=I
aE(a,b) ^ ^ 9E(a,b) „
_ _ _ ^ 0 and -3b-=0
give
3Na+bJ(3xi+di-Ci)=5!l(3yi+fi-ei)
i=l i=l
N N
a2^(3xi+di-Ci)+b2[x?+(xi-Ci)V(xi-di)^]
i=l i=l
==2L[(xi+di)(yi-ei)+(xi-Ci)(yi+fi)+Xiyi]
i=i
Thus the parameters a, a^, 02 and b are given by the solution of the following
systems of equation:
^ . 0 ^ = 0 ^ = 0 ^=0
da ' dtti ' da2 ' db
355
(II) b<0
A+bXp{a, tti, a2)+(bxi, bd^, bci)=(a+bxi, tti+bd;, a2+bci)
d[Zi, Y/=d[A+bXi, Yi]^=d[(a+bxi, aj+bd;, a2+bCi),(yi, q, f;)]^
2 2 2
=(a+bxi-yi) +[a+bXi-yi-(a]+bdi-ei)] +[a+bxi-yi-(a2+bCi-fi)]
N
E(a, ai,a2;b)=^d[Zi, Yif
i=l
For simplification we assume here that the fuzzy L-R number A has the form A=(a,
a*, a*), i.e a2=a2=a*. Then, the algebraic system of equations that are satisfied by
a,a*, and b is:
For b>0
«*=[Xfi+Z^i)^2N)-b[21d,+2cij/(2N)
Na 3^Xi/N+^(di-ei)/N +Na*(^5^d,+^Cij/(2N)
+b2^f(Xi-Ci)\(Xi+di)Vxf]
For b<0
a= /N
Na[35^x./N+X(di-ei)/N -Na* X ^ i + S c i ^ N )
=2^[(xi-Ci)(y,+fi)+(xi+di)(yi-ei)+Xiyi
356
A+BXi=(a, tti, a2)+(b, Pj, |32)(xi,Ci, di)=(a, ttj, a2)+(bxi, bc^+XiPi, bdi+x^pj)
=(a+bxi, ai+bCi+xjPi, a2+bdi+Xip2)
d[Zi, Yi]^=d[A+BXi, YiJ^=d[(a+bXi, tti+bCi+XiPi, a2+bdi+XiP2),(yi, e„ f)f
2 2 2
=(a+bXi-yj) +[a+bXi-yi-(a,+bCi+XiPi-ei)] +[a+bXi-yi-(a2+bdi-XiP2-fi)]
and the problem is to select a, a j , 02; b, Pj, P2 so as to minimize the criterion
E=2d[Zj,Yi]2. The solution is found by solving the equations:
Example 2.1
We are given the following L-R fuzzy data:
Y.=(ype,f,) Xi=(Xi,c,di)
(4, 0.6, 0.8) (21, 4.2, 2.1)
(3, 0.3, 0.3) (15, 2.25, 2.25)
(3.5, 0.35, 0.35) (15, 1.5, 2.25)
(2, 0.4, 0.4) (9, 1.35, 1.35)
(3, 0.3, 0.45) (12, 1.2, 1.2)
(3.5, 0.53, 0) (18, 3.6, 1.8)
(2.5, 0.25, 0.38) (6, 0.6, 1.2)
(2.5, 0.5, 0.5) (12, 1.8, 2.4)
The problem is to fit to them the model Y=a+bX (a,be R) or the model Y=A+bX
(Ae F(R), be R). Applying the procedures described in Sections 2.4 and 2.5 one can
easily verify tha the results are the fuzzy straight lines:
Y=1.052+0.147X
Y=(1.201, 0.180, 0.180)+0.136X
formulate the theory we need to define the concepts of N-dimensional fuzzy sets and N-
dimensional Gaussian fuzzy sets.
We know that a fuzzy set defines on the reference hyperset R a straight line section
the points of which loose, towards its ends, the property to belong to it. Analogously an n-
dimensional fuzzy set defines on the hyperset R" a closed volume the points of which
loose the propety to belong to it towards the boundary surface.
Definition 3.1
Analogously to the 1-dimensional fuzzy sets, A and B are said to be n-dimensional
(n-D) fuzzy sets if
where |iyi^(x) and Hg(x) are n-dimensional membership functions (see Ch. 1 , Def. 2.1).
Some useful operations of n-dimensional fuzzy sets (fuzzy vectors) are the following:
PI. Vector Sum of n-D fuzzy sets
If A and B are n-D fuzzy sets, then their sum
is also n-D fuzzy set. This sum is a generalisation of the vector sum of two
classical n-D vectors.
P2. Crisp translation of an n-D fuzzy set A in the direction of a
crisp vector y
Geometrically this means the translation of the "volume", defined by the
fuzzy set A, by the quantity y. The translated fuzzy set A has exactly the
same shape with A, i.e
with HA'(x)=^A(x-y).
A'={(X,^A(X)) | X 6 R V A ' ( X ) : R " ^ [ 0 , 1 ] }
P3. Direct image under P of the n-D fuzzy set A
Definition 3.2
A n-D Gaussian fuzzy set A with center ue Rn , variance Q and height a is
defined as :
A={(X, ^A(X)) I xe R " , ^ A ( X ) : R " - ^ [ 0 , 1 ] }
where
358
^^^^'^=1" ^g-(x-u)V'(x-u)^^
The operations of n-D Gaussian fuzzy sets follow as special cases of the operations
of arbitrary n-D fuzzy sets. Thus, let A=(a, u i, Qi) and B=(P, U2,Q2).
Then
GI. Translation of A by the vector v
A'=(a, Uj-v, Qi)
( - ' Qi Q2
A©B= min(a,|3), u^+Uj , — + - ;
V '^ ' ^ a 1-a
where a a parameter wih 0<a<l.
G3. Direct image of A under P
AnB=(Y, u*, Q)
where
7=aexp[-(u*-uj) Qj (u*-u,)
u*=[pQr'+(l-p)Q2']"'[pQ7'Gi-i-(l-p)Q2'G2]
Q=(l-d^)[pQr^+(l-p)Q2']"'
d^=(ui-G2)'^[p"'Q,+(l-p)"^Q2]"Hu,-U2).
359
Before going to the fuzzy state estimation problem, we recall that a classical n-D
vector can take each time as value only one point of R", and thus it covers all R". For
convenience the differences of classical and fuzzy variables and vectors are summarized in
the following table:
Variable Value
Classical (crisp) point of R
Fuzzy Section of R
Classical vector Point of R"
Fuzzy vector Subsetof points of R"
3.2 Fuzzy State Estimation Problem
Consider a discrete linear dynamic system with state equation:
x(k+1 )=Ax(k)+Bu(k)+Gw(k), x(0)=x(to)
and output equation:
z(k)=Cx(k)+v(k)
where the initial condition ^{IQ) is a n-D fuzzy set on the reference hyperset R" , the input
u(k) is exactly known, the input disturbance w(k) ia a m-D fuzzy set on R™, and the
output noise v(k) is a p-D fuzzy set on RP. The matrices A,B ,G and C are all known and
real-valued and have appropriate dimensions for the operations to have sense.
The fuzzy state estimation problem is the following:
GIVEN: (i) A system described by the above state-output eaquations
(ii) The set of fuzzy control signals U={u(0),u(l),...,u(k-1))
(iii) The set of fuzzy ouput-signals Z={2(l),z(2),...,z(k)}
FIND: An estimate of the fuzzy set of the state x at the time instant k, i.e an estimate of
x(k).
Problem Solution
Here the term "estimation" means the determination (estimation) of all parameters of
the fuzzy set including its membership function. The algorithm to be described makes use
of the operations described in Sec.3.1 and involves two steps.
Step I: Let x(k-l/k-l) be the estimate of x(k-l) on the basis of the data up to time k-1.
Then the one-step ahead predicted state estimate x(k/k-l) is given by :
x(k/k-1)=Ax(k- 1/k-1 )-t-Bu(k-1 )-i-Gw(k-1)
Clearly, this ives the set of all posible states x(k/k-l) which is reachable from
x(k-l/k-l) through the given dynamics of the system.
Step 2: This step computes the corrections to x(kA:-l) that must be made in the light of the
information transferred by the new output measurement z(k). More specifically: We solve
the output equation z(k)=Cx(k)-i-v(k) with respect to x(k), i.e
Now, the estimate x(k/k) of x(k) on the basis of data up to time k, is obtained as the
section of the quantities:
x(k/k-l) and-C-l[v(k)+(-z(k))].
This is so, since the estimate must belong to both of these fuzzy sets.
360
Therefore
x(k/k)=x(k/k-l)n-C '[vtk]©(-z(k))]
x(k/k-l)=Ax(k-l/k-l)eBu(k-l)©Gw(k-l)
x(k/k)=x(k/k-l)n-C"^[v(k)e(-z(k))]
x(k/k-l)=Ax(k-l/k-l)®Bu(k-l)®Gw(k-l)
^Ax(k--i/k-i)(x)=Ux(k-i/k-i)(A~ x)
b) The product Gw(k-l) of the matrix G by the m-D fuzzy set w(k-l). Thus
I^Gw(k-i)(x)=^w(k-i)(G'' x)
c) Translation of the n-D fuzzy set Ax(k-l/k-l) by the crisp vector
Bu(k-l). The result is:
M•Ax(k-l/k-])®Bu(k-l)('')=^^Ax(k-l/k-l)(''-Bu('<^-l))
=l^x(k-i/k-i)(A"'x-Bu(k-l))
d)Sum of the n-D fuzzy vectors Ax(k-l/k-l)©Bu(k-l) and Gw(k-l), i.e
I^Ax(k-l/k-l)®Bu(k-l)®Gw(k-l)^'')
=sup(M-Ax(k-i/k-i)®Bu(k-i/x-q)A^w(k-i)(G"'q)}
=sup{^^(k_]/k-i)[A"'x-Bu(k-l)-q]A^^(k_,)(G'^q)}
x(k/k)=x(k/k-l)n-C'nv[k]®(-z(k))]
a) Translation of the p-D fuzzy vector v(k) by the crisp vector -z(k), i.e
M-v(k)®(-z(k))(x)=^v(k)(x-(-z(k)))=^^(k)(x+z(k))
b) Multiplication of the crisp matrix -C"^ by the p-D fuzzy vector v(k)©(-z(k)), i.e
M-c-'[v(k)ffi(-z(k))](x)=^^v(k)f-C~'x^-z(k)]
Thus finally
l^x(k/k)(x)=|ix(kA-l)'^^^v(k)f-C"'x+z(k)]
4.CONCLUSION
In this chapter we have studied the fuzzy least-squares parameter estimation problem
and presented two methods. The first one is due to Celmins and the second to the present
authors. The first seems to be computationally very demanding, whereas the second
method is applicable only to the class of L-R fuzzy numbers but it has computational
efficiency similar to the classical least-squares method. Then our attention was focused to
the state estimation problem of systems subject to fuzzy disturbances and fuzzy initial
conditions.
An estimation algorithm was presented which has high theoretical value but is
practically limited by the high computational demand. The area of fuzzy parameter and
state estimation seems to be quite open and provides challenging theoretical and practical
problems.
d- Vxe[a,bJ , |iis,(x)=l
e. ^N(X) 'S strictly decreasing in [b,d]
f. Vxe [d,+oo), |aN(x)=0
From this definition it follows that N is a fuzzy real number if and only if N is a
convex and n normal fuzzy subset of R.
Some examples of fuzzy real nunbers are shown in Fig. A. 1
i L
H(x) H(x)
1
i
c a 4 b
(b)
l\
1 \
1 \
1
-N={(z,^N(y))|z=-y}
i/N={(z,tiN(y))lz-i/y}
It is obvious from the above definitions (and from Fig. A.2) that the operations of
the fuzzy numbers have a convolutional structure, and so they are computationally
involved. For example, if Supp(N), and Supp(M)=m then Supp(MxN)=mxn.
It is therefore necessary to develop a special type of fuzzy numbers, the operations
of which have computational complexity equivalent to that of classical operations. Such a
class of fuzzy numbers are the so-called L-R fuzzy real numbers anf are examined below.
Definition A.3: Let L(x) and R(x) be two functions satisfying the conditions of
Def. A.l. Then as L-R fuzzy number M around the classical number m, is defined a fuzzy
number M with membership function:
L((m-x)/a), x<m
^IM(X)=-^
R((x-m)/p), x>m
where m is the modal value and a, P, are spread parametes. Since the L-R fuzzy number is
fully determined by the triad of parameters m, a and p it is symbolised by: M=(m, a, P).
Example A.l
Let L(x)=l/(l+x2) and R(x)=l/(1+ | x |) with a=2 , P=3 and m=5. Then, the fuzzy
number "nearly 5" has the membership function:
L((5-x)/2)=l/[l+((5-x)/2) ] . x<5
^5M=
{ R((x-5)/3)=l/[l+ I (x-5)/3 I ] , x>5
0 5 X
As shown by the following Theorem A.l, the operations of fuzzy L-R numbers
reduce to numerical operations of standard numbers [2].
365
Theorem A.l
Let two L-R fuzzy numbers M=(m, a, p) and N=(n, y, 6). Then one can show [2]
the following:
/ . Fuzzy addition
2. Fuzzy opposite
3. Fuzzy subtraction
The subtraction has sense only between L-R and R-L , L-L and L-L , R-R and R-R
numbers (not between R-L and R-L numbers).
M-N=M-i-(-N)
4. Fuzzy inverse
5. Fuzzy multiplication
Three cases are distinguished:
Case A: If m>0 and n>0, then
6. Fuzzy division
We now proceed to extend the operators MIN and MAX in the fuzzy algebra. In the
classical algebra these operators are defined axiomatically as follows: Let a,Pe R. Then
by definition:
366
max(in,N)
kl k2 k3
Fig.A.4 Two fuzzy numbers and their maximum
Clearly, this cannot be answered through the axioms of classical arithmetic, and so a
new ordering relation is defined as follows [2]:
Definition A.4 Let
Max(M,N)={(z, HMAX(Z)) I ze R}
where
367
Hj^z),ze(-oo,kiMk3,+oo)
I |jfj;z),z€[ki,k3]
REFERENCES
1. A. Celmins, Least-Squares Model Fitting to Fuzzy Linear Data, Fuzzy Sets and
Systems , Vol.22, pp. 245-269, 1987.
2. D. Dubois and H. Prade, Operations on Fuzzy Numbers, Int. J. Systems Sci.,
Vol. 9, No.6, pp. 613-626, 1978.
3. H. Sira -Ramirez, Fuzzy State Estimation in Linear Dynamic Systems, Proc. IEEE
Conf. on Decision and Control, Vol.2, pp. 380-382, 1980.
4. H. Sira -Ramirez, Evolution of Fuzzy Sets in Linear Dynamic Systems, Proc. 1979
Int. Conf. on Cybernetics and Society, Denver, Colorado, Oct. 1979.
5. D.P. Bertsekas, Control of Uncertain Systems with a Set-Membership Description
of the Uncertainty, MIT Electr. Syst. Lab. Rept ESL-R-447, Cambridge, June
1971.
6. F.C. Schweppe, Recuirsive State Estimation: Unknown but Bounded Errors and
Systems Inputs, IEEE Trans. Auto. Contr. Vol. AC-13, No.l, pp. 22-28, 1968
7. H. Witsenhausen, Set of Possible States for Linear Systems Given Perturbed
Observations, IEEE Trans. Auto. Contr. Vol. AC-13, No.5 pp. 556- 558, 1968.
8. K.Lee, Modelling and Estimation of Distributed Systems with Uncertain
Parameters, In: Distributed Parameter Systems: Modelling and Identification (A.
Ruberti, Ed.),pp. 325-334 1978.
9. S.S.L. Chang, Control and Estimation of Fuzzy Systems, Proc. IEEE Decision and
Control Conf, pp. 313-318, 1974.
10. P. Diamond, Higher-Level Fuzzy Numbers Arising from Fuzzy Regression Models,
Fuzzy Sets and Systems, Vol. 36, pp. 265-275, 1990.
11. IFAC Report,, Round Table Discussion on the Estimation and Control in Fuzzy
Environments, Automatica, Vol.11, pp. 209-212, 1975.
12. S. Joszef, On the Effect of Linear Data Transformations in Possibilistic Fuzzy Linear
Regression, Fuzzy Sets and Systems, Vol.45, pp. 185-188, 1992.
13. H. Tanaka and H. Ishibuchi, Identification of Possibilistic Linear Systems by
Quadratic Membership Functions of Fuzzy Parameters, Fuzzy Sets and Systems,
Vol.41, pp. 145-160, 1991. .
14. A. Bardossy, Note on Fuzzy Regression, Fuzzy Sets and Systems, Vol.37, pp. 65-
75, 1990.
15. K.Jajuga, Linear Fuzzy Regression, Fuzzy Sets and Systems, Vol.20, pp. 343-353,
1986.
16. H.Tanaka and J. Watanabe, Possibilistic Linear Systems and their Applications to
the Linear Regres.sion Model, Fuzzy Sets and Systems, Vol.27 pp.275-289, 1988.
17. H.Tanaka, S. Uejima and K. Asai, Linear Regression Analysis with Fuzzy Model,
IEEE Trans. Syst. Man Cybern., Vol.5 SMC-12, No.6, 1982.
368
Shyi-Ming Chen
Abstract
In this paper, vvc present a methodology to deal with fuzzy reasoning based on the
matching function S. The singlc-input-singlc-output (SISO) fuzzy reasoning scheme and
the multi-input-single-output (MISO) fuzzy reasoning scheme are discussed in details.
1. Introduction
Much knowledge residing in the knowledge base of an expert system is fuzzy rather
precise. Thus, a powerful expert system must have the capability to deal with fuzzy
reasoning. Fuzzy reasoning can be viewed as a process to infer a possibly imprecise
conclusion from a set of possibly imprecise premises [6, p. 320]. Many fuzzy reasoning
methods have been proposed, such as [1], |2], [3], [4], [5], [6, p. 29l-341|, [9], [10], [12],
In most of the above references, the following single-input-single-output (SISO) fuzzy
reasoning scheme is discussed:
Rule l : l F X i s A i T H E N Y i s B i
Rule 2: IF X is Aj THEN Y is B2
• (1)
Rule p: IF X is Ap THEN Y is Bp
Fact: X is Ao
Consequence: Y is Z:^
where X and Y are linguistic variables [15], and Ao, Ai AT, ..., Ap, BQ, Bi, B2,..., and Bp
are fuzzy terms (linguistic terms), such as "slow", "medium", "fast", etc.
The fuzzy terms usually are represented by fuzzy sets [13]. Roughly speaking, a fuzzy
set is a class with fuzzy boundaries. Let U be the universe of discourse (i.e., a collection
of objects), f/={«], H2> •••> "«}• A fuzzy set A of L' is a set of ordered pairs {(i/j,/4(HI)),
(H2./4("2)), •••. ("/i,//i("«))}. where/4, Z^: U -* [0, 1], is the membership function,/^(w,)
369
S,G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 369-386.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
370
indicates the grade of membership of «, in A, and 1 s / s /;. Let A and B be two fuzzy sets
of the universe of discourse U, U={iii, 112, •••, u,,}, and lei/A and//? be the membership
functions of the fu/./y sets A and B, respectively, whereZ^: U—*\0, !],/«: U—*10, 1],
A={{iti,J\(iti)) I u,E.U}, and B={{Ui,/ij(Ui)) I iiiizU}. The fuzzy sets A and /^ arc called
equal (i.e., A=B) if V iiitiU,fA(iii)=fij{ii,). The union operation between the (uzzy sets A
and B is defined by
Let U be the universe of discourse, U =|0, 100], and "speed" be a linguistic variable
[15], and let "slow", "medium", and "fast" be three fuzzy terms. According to [9], these
fuzzy terms can be characterized as fuzzy sets whose membership function curves are
shown in Fig. 1.
Tspeed
40 55 70 Speed
Fig. 1. A diagrammatic representation of fuzzy speed.
Consider the fuzzy reasoning scheme in (1), where A], AT, .., and Ap are fuzzy sets of
the universe of discourse U, U={ui, 112,..., ««}, and B\, B2 Bp arc fuzzy sets of the
universe of discourse of V, V={vi, VT, •., v,„}. For a given system's observation fact "X is
Ao", where Ao is a fuzzy set of (J, U-{t4i, HT, ..., «„}, and the membership function of Ao
^^fAi),/An'- f^—*\0, 1|, the system can perform inference by either compositional rule of
inference (CRl) [14] on each of the rules or by approximate analogical reasoning (AAR)
[ 12) on selected rules from the knowledge base to reach a conclusion "Y is /Jo", where Bo
is a fuzzy set of the universe of discourse V, V={vi, VT, .., v',„}. A comparison of CRI
based fuzzy reasoning methods has been discussed in [10] by Mizumolo and
Zimmcrmann. The approximate analogical reasoning (AAR) schema is based on
similarity measures. In [16], Zwick el al. have made a comparative analysis of 19 kinds of
similarity measures between fuzzy sets. They pointed out that some similarity measures
(such as S4, c(»,Ao^ and A.) ha\'c relatively good performance, and the Euclidean
distance measure was found not to be a gtxxl measure. However, in [12], Turksen ct al.
pointed out that the conclusion in [16] may not applicable to all cases and suggested that
users should experiment with different similarity measures before deciding on an
appropriate measure for their t)wn use.
In this paper, we present a new methodology for pcrfontung fuzzy reasoning based on
the matching function S [3], [7]. The proposed methcxl belongs to the AAR scheme. The
single-input-singlc-output (SISO) fuzzy reasoning scheme and the multi-input-single-
output (MISO) fuzzy reasoning scheme arc discussed in details.
371
In [7], Ke el al. presented a matching function ^S' to measure the degree of similarity
between two vectors. Let a and h be two vectors in R", where i? is a set of real numbers
between zero and one, i.e..
where S{a, ft)e [0, 1], which can be considered as the similarity measurement between
the vectors a and h . The value of S{a, h) indicates the degree of similarity between a
and h . The larger the value of S(a, h), the more the similarity between the vectors a
and b.
Because a* h ^ I allftlcos0, " 0 " is the angle between a and h , and because cosO =
1, (2) can be rewritten into
- - Iall/71cos0
S(a, h) _ - - (3)
Max(\a\\a\, \b\\b\)
The relationship between the vectors a and b can be distinguished into the following
three cases:
\a\\b\co?.d \h\co&B
Case 2: If Ic/I > l/?l, then S(a, b) =
lallal \a\
\a\\b\cos6 lalcos^
Case 3: If lal < 161 , then .S'(^, 7)) =
\b\\b\ \b\
Therefore, we can see that
1) If rt and h are identical vectors, then S{a, b) = \.
2) The larger the length difference between a and /;, the less the similarity between
a and b . If \a\ » \T?\ or 1^1 » ki\, then Si'i, h) ^ 0.
372
3) The larger the angle 6 , Ihc less the similarity between a and b . U 6 -* n,
then S(a, ^) a 0, where n is an even number.
Based on the matching function S, we can measure the degree of similarity between
fuzzy sets. Let U be the universe of discourse, U={ui, ti2,..., ««}. and let A and B be two
fuzzy sets of t/, i.e.,
where fl/ fc |0, 1],fc,e [0, 1], and 1 s / s //. By using the vector representation method,
the fuzzy sets A and Bean be represented by the vectors A and B, respectively, where
Based on the similarity function S, the degree of similarity between the fuzzy sets A and
B can b)e measured,
I»5
.S'(A, B) Max{A* A, B'B)
where S{A, B) 6= fO, 1]. The larger the value of S(A, B), the more the similarity between
the fuzzy sets A and B.
Let's consider the following generalized modus ponens (GMP):
Rule:IFXis/i| THEN Y is Bi
*
Fact: X is A^
— I
Consequence: Y is fi]
where X and Y are linguistic variables, Aj and A\ are fuzzy sets of the universe of
discourse U, U-{n\, 112, .., «„}, and Bj and Bi are fuzzy sets of the universe of discourse
« *
V, V={v\, VT, ..., v,„}. Assume that the membership function of the fuzzy sets A] , Ai, Bj,
« , arc / . , / , / . , and / , respectively, where / . : U-*\{\ 1], / : U-^\0, 1),
f . : V-^|0, 1], /' : V-*[0, 1], and the I'uzzv sets Aj', Ai, and ZJ,* have the followmg
"i ' ^1
forms:
AI = {OIU -fl), (U2, XT), ..., («„, X,,)}
-41 :^ {(Ml, Vl), (U2, Vl), - , (««, .Vn)}
A = { ( ^ 1 . -1)' (*'2, ^2) (V,„, -y„)}
373
where J:,£ [0, 1], ViG [0, 1], zjE: [0, 1], 1 s / s H and \& j ^m. Let Aj and A^ be the
*
vector representation of the fuzzy sets A^ and Ai, respectively, where
Then, based on the matching function S, the degree of similarity between the fuzzy sets
A, and A\ can be measured,
~^ — A •A
S(A,. A,) = - ^ ^ ^
Max{Ai • Ai, A, • A,)
where 5( A;, A,) E [0, 1]. Let 5(A|, Aj )= /:, and let the deduced consequence of the rule
is "Y is 5, ", where the membership function of the fuzzy set D is as follows:
* * *"
It IS obvious that if Aj and Aj are identical fuzzy sets (i.e., A\ =Ai ), then S(A\, Aj )=1
*
and Z^ is equal to/?].
Consider the following single-input-single-output (SISO) fuzzy reasoning scheme:
Rule 1: IF X is Ai THEN Y is S,
Rule 2: IF X is Ao THEN Y is B2
Rule p: IF X is Ap THEN Y is Bp
Fact: X is Ao
Con.scquence: Y is BQ
where Ao, Aj, A2, ...,and A„ are fuzzy sets of the universe of discourse U, (/={«], 112,...,
«„}, and Bo. ^1-^2. •••, ana ^p are fuzzy sets of the universe of discourse V, K={V],V2,
..., v,„}. Assume that
A, = {(Hi, Jr,l), (U2, Xfi), .... (M,„ X,,,)}
Bj - {(v\, yjx), (V2, ypj,..., (v,„, Vym)}
374
where \ ^ i ^ p and 1 s y s p . By using Ihc vector representation method, the fuzzy set A,
can be represented by the vector A^ ,0 ^i s p, where
S{\^, A\)=ki ^> the deduced consequence of rule R[ is "Y is B^ ", where
«i ={(vi, A:,*yii), (V2, *i*.Vi2), •.•, (Vm, ^i*.ViJ}.
i'(A)' ^ ) =^^2 =^ the deduced consequence of rule ^^2 '^ "Y is ft ", where
/ ^ = {(*'l. ^ 2 * V 2 l ) . (l'2, ^ 2 * y 2 2 ) . - . (Vffl, ^2*y2;«)}-
.S'( A)' ^p)=^p =* "ic deduced consequence of rule Rp is "Y is B^", where
«^ = {(vi, Vvpi), (V2, kp*yp2)...., (v,„, V V " ) } -
where ^,fci [0, 1] and \-^i ^p, and the deduced consequence of the above SISO fuzzy
reasoning scheme is "Y is HQ", where
and " U" is the union operator of fuzzy sets. That is,
where
B={{VuZi),{V2,Z2),-,{V,n,Z,n)y
By using the vector representation method, the fuzzy sets A^ , A,,..., A^ , and A/ can
«• *•
Based on the matching function .V, we can get the following results:
S{AI Ai)=ki
.V(A,*, Ai)=ki
376
where ki^E [0, Ij, and 1 s / s / . Letc = Min{k\, ...,kiy, Ihe deduced consequence ol the rule
is "Y is B*", where
• (4)
where Aoi, An. A21,..., and A„i are fuzzy sets of the universe of discourse f/, U={u\, 112,
..., «„},..., At)/, Ai/,A2/,..., ana Ap/are fuzzy sets of the universe of discourse T, 7={/i, h,
..., If}, dnd Bo,B\, B2, ••, and Bi are {\i7.z\ sets of the universe of discourse V,V={vi,
V2,..., v,„},i.e.,
^ 1 = { ( V i , J i l ) , (V2,2i2), . . • , ( v , „ , . r i J }
B2 ={(Vi, :2i), (V2, :22). •••> (Vm. Jf2/n)}
By using the vector representation method, the fuzzy sets AQI, A n , A21, •••, A ^ j , . . . , AQ/.
Ai/, A2/,..., and Api can be represented by the vectors A Q ] , A J J , / ^ J , ..., A ^ j , ..., AQ/ ,
A]/, A2/,..., and Apj, respectively, where
Let
Cp=Minikpu...,kpi).
Then, based on the previous discussion, we can get the following results:
*
the deduced consequence of rule Rp is "Y is B„", where
In this case, the deduced consequence of the above MISO fuzzy reasoning scheme is "Y
is Bo", where
£^ =^"' U f t U ...UB^p
and " U" is the union operator of the fuzzy sets, i.e.,
where
si=Majiici*zn,C2*Z2\, ...,Cp%i)
.V2 = MaJ((C]*Zi2< '•2*^22, •••. f'>*2p2)
observation facts and the antecedent portion of rule /?, is larger than or equal to X..
Otherwise, the rule R, can not be fired. For example, consider the following generalized
mtxlus ptmens:
«" *• *
Let A] , Aj ,..., Aj , and Aj be the vector representation of the fuzzy sets A], Aj,..., A/,
and A) , respectively. Supposed X is a threshold value, XJE[0, 1 J, and the fuzzy set B has
the following form:
where Zi e[0, 1] and 1 ^ / s /«. Let C = M ' « ( 5 ( A ' , AI ),..., S(Ai, A,)). If ca A, then the
rule can be fired, and the deduced consequence of the rule is "Y is B"\ where
B*={(vi.0),(v2,0),...,(v,„,0)}.
Consider the SISO fuzzy reasoning scheme shown in (1) or the MISO fuzzy reasoning
scheme shown in (4). If there are some rules that were fired based on the threshold value
A,, then the deduced consequence of these rules would be combined to get a deduced
consequence "Y is HQ" of the fuzzy reasoning scheme. Finally, we can find an
approximate linguistic term for BQ. This can be accomplished by using the linguistic
approximation methcxl described in [11, p. 56] by matching BQ against a "linguistic term
base", where all allowable linguistic terms are defined by membership function curves
and the linguistic term which has the minimum Euclidean distance from BQ would be
selected as the approximate linguistic term for BQ, where the Euclidean distance between
any two fuzzy sets is calculated as follows. Let F and G be any two fuzzy sets of the
universe of discourse U, where
and let F and G be the vector representation of the fuzzy sets F and G, respectively,
where
380
F- (/,. h /«>
3. A Simple Example
In this section, we use an example to illustrate the fuzzy reasoning process. Consider
the following single-input-singic-output (SISO) fuzzy reasoning scheme:
Rule l : I F X i s A i T H E N Y i s B i
Rule 2: IF X is Ao THEN Y is B2
Rule 3: IF X is AT THEN Y is S,
Rule 4: IF X is Aj THEN Y is B4
Rule 5: IF X is A5 THEN Y is B5
Fact: X is AQ
Consequence: Y is £()
where X and Y are linguistic variables, AQ, A [ , AT, A3, A4, and A5 are fuzzy sets of the
universe of discourse t/, U={u\, «2,..., H14}, B\, Bi, BT„ B4, and B^ arc fuzzy sets of the
universe of discourse V, V={i'i, vo,.... V14}. Assume that Ao=A3. These fuzzy sets are
shown as follows:
Ao= {(«,, 0), (H2, 0), (K3, 0), («4, 0), («5, 0),
(H6, 0), (H7, 0), (H8, 0.5), (My, I), ( « K ) , 0.5),
(HH.0),(«I2,0),(«I3,0),(H,4,0)}
A, = {{III, 1), {112, 1). («3, 1), (114, 0.5), (H5, 0),
(Hf, 0), («7, 0), (MR, 0.5), («9, 1), («I0, 0.5),
(«ll,0),(«i2,0),(H,3,0),(Mi4,0)}
A',= {(M,, {)), (Ml, 0), («3, 0), (H4, 0.33), (M5, 0.67),
(«6, 1), ("7. 0.67), (Mg, 0.33), (M9, 0), («io, 0),
(M,|,0),(M,2,0),(M,3,0),(M,4,0)}
A3 = {(M,, 0), (M2, 0), (M3, 0), («4, 0), {lis, 0),
(M6, 0), (M7, 0), (Ms, 0.5), (Mg, 1), (MK), 0.5),
("11,0),(«12,0),(M13,0),(«14,0)}
A4= {(«!. 0), (H2, 0), (H3, 0), (M4, 0), {II5, 0),
(M6, 0), (M7, 0), (MR, 0), (Mg, 0), (HJO, 0.5),
(Ma,l),(Mi2,0.5),(M,3,0),(Mi4,0)}
381
A5= {(«!, 0), (112, 0), («3, 0), («4, 0), (//5, 0),
(H6, 0), (H7, 0), (lis, 0), (M9, 0), (Mio, 0),
( H „ , 0 ) , ( / / i 2 , 0 . 5 ) , ( / / n , 1),(«14, 1)}
flT= {(vj, 0), (vo, 0), (V3, 0.5), (V4, 1), (vs, 0.5),
(V6, 0), (V7, 0), (Vs, 0), (Vg, 0), (Vio, 0),
(V„,0),(V,2,0),(V,3,0),(V,4,0)}
fi4= {{Vl, 0), (V2, 0), (V3, 0 ) , (V4, 0), (V5, 0),
(V6, 0), (V^, 0.1), (Vg, 0.4), (Vg, 0.7), (V,o, 1),
(Vll,0.7),(V,2,0.4),(Vi3,0.1),(Vi4,0)}
B5= {(V,, 0), (V2, 0), (V3, 0), (V4, 0), (V5, 0),
(V(, 0), (V7, 0), (Vg, 0), (Vg, 0), (Vio, 0),
(Vii,0,5),(Vi2, l),(Vi3, 1),(V,4, 1)}.
By using the vector representation method, the fuzzy sets AQ, A^ , A2, Aj,, A4, A5, Bi, B2,
Bj,B4, and B5 can be represented by the vectors AQ,Ai,A2,Ai,A4,A^,B^,B2,B^,
B^ , and B5 , respectively, where
Ui U2 U: U4 Us Ue U? Ua U9 UIOUP U'2UI3U:
membership fupctions
i.Bl B2 B3
V; V2 V3 V4 V5 V6 V7 Va V9 Vio V i : V : j V.3 VM
.V(^, Aj) = _ ^ 1 ^ =0
S(A^,^). _'^L^^ = 0.
When the threshold value X, is set to 0, then all the rules can be fired. According to the
previous discussion, we get the following results:
( v n , 0 ) , (Vi2,0),(v,3,0),(vi4,0)}
B, = {(vi, 0), (V2, 0), (V3, 0.039), (V4, 0.078), (vs, 0.039),
(V6, 0), (V7, 0), (Vg, 0), (V9, 0), (Vio, 0),
(Vll,0),(Vi2,0),(Vi3,0),(Vi4,0)}
R^ = {(vi, 0), (V2, 0), (V,, 0), (V4.0.1), (V5, 0.5),
(V6, 0.8), (V7, 1), (Vg, 0.8), (V9, 0.5), (vio, 0.1),
(Vll,0), (V|2,0),(Vi3,0),(Vi4,0)}
li ={(Vl, 0), (V2, 0), (V3, 0), (V4, 0), (Vs, 0),
(V6, 0), (V7, 0.0167), (Vg, 0.0835), (vg, 0.1336), (v,o, 0.167),
(vii, 0.1336), (V]2, 0.0835), ( V B , 0.0167), ( v ^ , 0)}
BI = { ( V I , 0), (V2, 0), (V3, 0), (V4, 0), (V5, 0),
(V6,0),(V7,0),(V8,0),(V9,0),(Vio,0),
(Vll,0),(V,2,0),(Vi3,0),(Vi4,0)}.
where /^, is the reasoning result from rule 7?,, I s / s 5 . Therefore, the deduced
consequence of the fuzzy reasoning scheme is "Y is BQ", where
^ = B,' U ft U £^' U ^4 U B^
= {(vi, 0)" (V7, 0), (V3, 0.039), (V4, 0.1), (V5, 0.5),
(V6, 0.8), (V7, 1), (i^, 0.8), (V9, 0.5), (vio, 0.167),
(vn, 0.1336), ( v p , 0.0835), ( v ^ , 0.0167), (vi4, 0)}.
The reasoning result of the SISO fuzzy reasoning scheme is represented by the dotted line
shown in Fig. 3. Based on the linguistic approximation method, vvc can sec that the
approximate linguistic term of the reasoning result/^ is B3.
384
membersMip f u n c t i o n s
II A1 A2
Ui U2 UJ U 4 U5 Ue U7 Uo U9 U:oUr U)2Ui3Ui
membership fLnc'.ions
iiBl B2
When the threshold value k is set to 0.15, then by the previous discussion, only the
rules Ry, and lix can be fired. In this case, we get
B'I ={(vi, 0), (1-2, 0), (u,, 0), (V4,0), (V5, 0),
(V(„0),(V7,0),(V8,0),(V9,0),(V,0,0),
(Vll,0), (V'io,0),(v,3,0),(Vi4,0)}
ft = {(V,, 0), (v., 0), (v^, 0), (V4, 0), (vs, 0),
(V(„0),(v7,0),(v8,0),(v9,()),(vio,0),
(V,],0), (V,2,0),(V,3,0),(Vi4,0)}
/^' = {(vi, 0), (V2, 0), (v^, 0), (V4, 0.1), (V5, 0.5),
(V5, 0.8), (V7, 1), (vs, 0.8), (V9, 0.5), (vio, 0.1),
(Vll,0), (V,2,0),(Vi3,O),(Vi4,O)}
li = {(vi, 0), (V2, 0), (v,, 0), (V4, 0), (V5, 0),
(V(^ 0), (V7, 0.0167), (V8, 0.0835), (v^, 0.1336), (vjo, 0.167),
(vii, 0.1336), (vp, 0.0835), (vn, 0.0167), (V14, 0)}
385
li, = {(V|, 0), (V2, 0), (\'3, 0), (V4, 0), (V5, 0),
(V6, 0), (v^, 0), (vg, 0), (V9, 0), (vio, 0),
(Vll.O), (V,2,0), (Vi3,0),(V]4,0)}.
where B ' is the reasoning result from rule /?,, I s / s 5 . Therefore, the deduced
consequence of the fuzzy reasoning scheme is "Y is BQ", where
membership functions
membership functions
MBI B2
Vi V2 V3 V4 Vs Vs V7 Ve V9 V;o V i i V 1 2 V 1 1 V:.>
4. Conclusions
References
I i I Z. Cao, A. Kandel, and L. Li, "A new mtxiel of fuzzv reasoning," Fuzzy Sets and
Systems, \ ol. 36, pp. 311 -325, 199().
|2) T. C. Chang, K. Hasegawa, and C. W. Ibbs, "The effects of membership function on
fuzzy reasoning," Fuzzy Sets and Systems, vol, 44, pp. 169-186, 1991.
| 3 | S. M. Chen, "A new approach to handling fuzzy decisionmaking problems," IEEE
Transactions on Systems, Man, and Cybernetics, vol. 18, no. 6, pp. 1012-1016,
1988.
(4| S. M. Chen, J. S. Ke, and J. F. Chang, "Knowledge representation using fuzzy Petri
nets," IEEE Transactions on Knowledge and Data Engineering, vol. 2, no. 3, pp.
311-319, 1990.
[5] Y. Ezawa and A. Kandel, "Robust fuzzy inference," International Journal of
Intelligent Systems, vol. 6, pp. 185-197, 1991.
| 6 | J. Giarralano and G. Riley, Expert Systems: Principles and Programming, Boston:
PWS-KENT Publishmg Company, 1989.
|7| J. S. Kc and G. T. Her, "A fuzzy information retrieval system model," Proceedings
of 1983 National Computer Symposium, Taiwan, 1983, pp. 147-155.
[8j A. Kandel, Fuzzy Mathematical Techniques with Applications. Addision-Wesley
Publishmg Company, 1986.
(9| C. C. Lcc, "Fuzzy logic in control systems: fuzzy logic controller-part I," IEEE
Transactions on Systems, Man, and Cybernetics, vol. 20, no. 2, pp. 404-418, 1990.
110] M. Mizumoto and H. J. Zimmermann, "Comparison of fuzzv reasoning methods,"
Fuzzy Sets and Systems, vol. 8, pp. 253-283, 1982.
[111 K. J. Schmucker, Fuzzy Sets, Natural luinguage Compulations, and Risk Analysis,
Rtx;kville, MD: Computer Science Press, 1984.
|12| I. B. Turksen and Z. Zhong, "An approximate reasoning approach based on
similarity measures," IEEE Transactions on Systems, Man, and Cybernetics, vol.
18, no. 6, pp. 1049-1056, 1988.
(131 L- A. Zadch, "Fuzzy sets," Information and Control, vo]. 8, pp. 328-353, 1965.
[14| L. A. Zadeh, "Outline of a new approach to the analysis of complex systems and
decision process," IEEE Transactions on Systems, Man, and Cybernetics, vol. 3, no.
1, pp. 28-44, 1973.
[15| L. A. Zadch, "The concepts of a linguistic variable and its application to
approximate reasoning (1)," Information Science, vol. 8, pp. 199-249, 1975.
|16| R. Zwick, E. Carlstein, and D. V. Budescu, "Measures of similarity among fuzzy
concepts: a comparative analysis," International Journal of Approximate Reasoning,
vol. l,pp. 221-242, 1987.
C H A P T E R 15
I. Introduction
The present chapter deals with the problem of determining the p a t t e r n class and its
multivalued s h a p e / b o u n d a r y from sampled points (training samples). Once these are
computed, some salient features of the class can then be extracted which are useful
in making decisions about a course of action {e.g., identification, classification and
p a t t e r n description) to be taken later. This will also reduce the storage requirement
of the complete pattern class.
It may be noted t h a t in most of the real life pattern recognition problems, the
complete description of a pattern class is not known. Instead, a few sampled points
are usually available which are assumed to represent the class. Hence determining
the p a t t e r n class and its shape from sampled points is an important problem in
p a t t e r n recognition.
There are various approaches described in the literature for determining the shape
of a p a t t e r n class from sampled points [1-6]. These methods are mostly heuristic in
n a t u r e and they provide an exact boundary or shape of the p a t t e r n class. One of
the inherent observations about these algorithms is that the b o u n d a r y of the class
is restricted by the sampled points. This need not be true because the resulting
boundary leaves certain regions not confined in it, although it should be. So, it is
necessary to extend the boundaries to some extent to handle the possible uncovered
portions by the sampled points. T h e extended portions should have the following
two properties:
(i) As the number of sampled points increases, the extended portions should de-
crease.
(ii) T h e extended portions should have less possibility to be in the p a t t e r n class
than the portions explicitly highlighted by t h e sampled points.
T h e second property leads to define a multivalued or fuzzy (with continuum grade of
belongingness) boundary of a p a t t e r n class. T h e basic concept of one of the existing
methods [1-6] is described below in short for illustration.
Edelsbrunner et al. fl] introduced the notion of the a-shape of a finite set of
points, for arbitrary real a. This notion is a generalization of the convex hull. Given
a set 5 of t points in a plane, the convex hull of S may be defined as the intersection of
all closed half planes that contain all points of S. T h e a-hull of S is the complement
of the union of all open discs of radius not less than 1 / Q (for arbitrary negative real
387
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 387^18.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
388
a) which contains no point of S. The shape of the set is determined using a-huUs.
If the selection of a is proper then the shape of the set obtained by the method
resembles the intuitive shape of the planar set. These Q-hulls have been found to be
extremely useful in computational geometry [7]. The selection of a is very important
for the method, but the authors did not provide any criterion for the selection of a.
Also this method does not provide the multivalued shape of a pattern class.
We have formulated here a methodology for providing multivalued (fuzzyl shape
of a pattern class from a set of sampled points. The procedure can be viewea in two
phases. Phase I is concerned with the decomposition of sample set into groups of
nearly N-dimensional parallelepiped (henceforth parallelepiped) shape. The decom-
position is based on the boundary variations (geometric complexity) of the pattern
class found in the sampled points. A window approach has been incorporated here
to find the boundary variations of the sample groups or sets. Phase II determines
each of the subclasses corresponding to the groups separately, puts them together,
and finds the multivalued shape of a pattern class. The effectiveness of the procedure
has been demonstrated on some artificially generated data sets in IR^ and ff^. The
practical applicability of the method has been demonstrated on a real life speech
data set. The convergence of the estimated classes to the original ones has been
verified using HausdorfF metric and a newly defined metric.
In section / / , some basic concepts along with the block diagram of the proposed
method are discussed. Section III deals with the procedure for decomposing a
training sample set. Section IV provides an approach to determine the multivalued
boundary of a pattern class. Experimental results are provided in section V. It
is to be mentioned here that the sections / / / , IV and V concern with only two
dimensional feature space. The extension of the procedure to higher dimensional
space (M ) and its implementation to the pattern classes in ]F^ are provided in
section VI. The convergence property of the proposed algorithm is discussed in
Section VII. Section VIII deals with the conclusions. ^
The relevance of the properties (i), {ii), (Hi) and (iv) of Definition 1 is provided in
[6]. Let ]B= {A : A satisfies Definition 1}. iB is the collection of all classes in Hi'^.
Any A £ E is referred to as the pattern class. •
b. Accuracy factor
It has been argued in the previous section that the boundary of a pattern class
obtained from its sampled points should be extended to some extent to highhght
the possible uncovered portions of the class by the sampled points. An accuracy
factor (St) based on the number of sampled points (<) is considered here for the said
extension to manage the uncertainty. St satisfies [8]
1 . 1 ,,.
juN < ^'< ^iuum ^^>
so that as t increases, i5( —> 0 and tS^ -+ oo. Since St decreases with the increase of
t, the accuracy of the obtained boundary also increases with the increase of t. The
selection of an accuracy factor is guidea by the inequality (1) and its justification
can be found in [9, 10]. •
c. Coverage factors
Each individual sampled point represents a covered area of the pattern class in the
feature space. Hence, in order to find the possible uncovered portions of the class
by the sampled points, the boundaries are extended to some extent. The extended
portions should have less possibility to be in the pattern class than the portions
explicitly highlighted by the sampled points. To decide on the amount of extension,
a factor, named as coverage factor, is defined below corresponding to each feature
axis.
Let Xi.,X2,- • • ,Xi,... ,Xt be the training samples where X,- = (x^j, a ; , j , . . . , x,-^,..
.,Xi^y and X, denotes the jih feature value of the fth sample. Let MAXj and
MINj denote the maximum and minimum feature values respectively in the sample
set corresponding to the jth {j = 1,2,..., iV) feature i.e.,
MAXj — max Ix; 1 and MINj = min | x , )
•' 1=1,2 t ' ^ i=l,2,...,( ••
The coverage factor for the set of sampled points corresponding to the j t h feature,
denoted by £j (j = 1,2,..., A'^), is defined as
£j = {MAXj - MINj) X St (2)
where St is the accuracy factor. When the number (f) of sampled points increases,
the value of the accuracy factor {St) decreases, and correspondingly the values of the
coverage factors (e/s) also decrease and the accuracy of the boundary increases. •
class, then the portions are referred to as the holes. The intuitive idea behind holes
of a pattern class can be put mathematically by the following definition.
Definition 2 : A -pattern class A is said to have k holes if
B. Block diagram
The block diagram of the proposed multivalued shape determining procedure is
shown in Fig. 1. It consists of two parts, namely the decomposition and the fuzzy
processor. The decomposition section deals with the decomposition of the sample
set into some groups of nearly parallelepiped shape. The fuzzy processor determines
each of the subclasses corresponding to the sample groups separately and all these
subclasses are then combined to compute the multivalued shape of the pattern class.
DECOMPOSITION
Boundary Pattern
Variation Class
Calculator Sub-divider
_L
Membership Boundary ,'^ Multivalued \
Function H Pattern
Estimator J Decider V Class ,/
FUZZY PROCESSOR
The decomposition section consists of three blocks jis shown in Fig. 1. The hole
detector block decomposes the training sample set with holes into groups to find
the hole information. The boundary variation calculator block finds the values of
the boundary variation factors corresponding to all possible boundary directions.
These boundary variation values are analyzed in the pattern class sub-divider block
to decompose (if necessary) the sample set into groups.
391
The theory of fuzzy sets [11-13] has been used in the fuzzy processor section to
extend the boundary of the sample set and also to relate every point in the whole
feature space to its possibility to be in the pattern class. The membership function
estimator block decides about the compatibiUty/membership functions to represent
each of the subclasses corresponding to the sample groups. The boundary decider
block determines each of the subclasses separately, puts them together and finds the
multivalued shape of a pattern class. Jk
The following two sections (i.e., sections III and IV) describe in detail the opera-
tions of different blocks of Fig. 1 for pattern classes in IR^. A
III. Decomposition
The decomposition section consists of three blocks, namely hole detector, boundary
variation calculator and pattern class sub-divider. It takes the training samples of
a pattern class as an input and decomposes the training sample set into groups
if the sample set is found to be not nearly rectangular in shape. To obtain this
decomposition, a procedure based on some overlapping windows is adopted here.
Before describing the operations of various blocks of decomposition, the approach to
generate the windows is furnished below.
Generation of windows : The two features under consideration are referred to as
the first (Fi) and second {F2) feature axes respectively. In the proposed approach,
one of the axes is considered as the base axis and the other axis is considered as the
height axis. Consequently the coverage factors cori'esponding to the base and height
features are referred to as the base coverage factor (£{,) ana height threshold factor
(eh) respectively.
The sampled points are first of all arranged in ascending order according to the base
feature values. The first window starts with the first sample of the ordered sample
set and it includes all those samples one after another in ascending order until its
base coverage length exceeds £),• Assume that the generated first window ends with
the A;th sample of the ordered sample set. Then the second window will end with the
(^-f l)th sample and to find the starting point of this window, it proceeds backward
from kth sample until its base coverage length exceeds £;,. Similarly other windows
are constructed by including one new sample at the end and excluding some samples
from the beginning of the previous window such that the base coverage length would
atleast be £(,. The last window ends with the last ordered sample i.e., sample with
the highest base value. Thus, some overlapping windows of the sample points are
generated utilizing the sample base values and the base coverage factor (ej).
The maximum and the minimum height values are found from each of the windows
and these are taken to be the upper (u) and lower (/) boundary values respectively
of that window. The combination of tne upper boundary values highlights the upper
boundary of the training sample set and the combination of the lower boundary
values provides the lower boundary of the training sample set.
The way in which the boundary of a training sample set is obtained using the
aforementioned procedure for generating windows, is explained below considermg a
typical pattern class [Fig. 2(a)]. A hypothetical training sample set is shown in
Fig. 2(a), where the locations of samples in the feature space are shown by cross
( x ) marks. Initially, Fj is considered as the base and few windows are generated.
In such windows, Fj is considered as the height feature. A typical window is shown
by dotted lines in Fig. 2(6) where the samples in the window are shown by tick
392
{^/) marks. The boundary values corresponding to the lower and upper boundary
directions (referred to as 2| and 2„ respectively) are also marked for the window.
Based on these boundary values, the rough boundaries in the coded directions 2/ and
2„ are dra^vn in Fig. 2(c). To find the boundaries in the lower and upper directions
of Ft (i.e., directions 1| and 1„), some windows are first of all generated considering
F2 as the_base feature. A typical window with its sample points and boundary values
is shown in Fig. 2{d). Pig. 2(e) shows the rough boundaries in the coded directions
1| and 1„.
s )s X * ^ X X
^ a< .t * "
Combining the boundaries in figures 2(c) and 2(e), the complete boundary of a
pattern class is obtained [Fig. 2(/)]. To incorporate the possible uncovered portions
of the pattern class by the training set (Fig. 2(a)], these boundaries are extended
to some extent (depending on the coverage factors). To visualize the said extension,
the extended boundary of the pattern class is conceptually drawn in Fig. 2(/). The
extended portion should have lower possibiUly to be in the class than the portions
explicitly highlighted by the sample points. The extended regions decrease with the
increase of the sample size (<). •
A. Hole detector
The concepts about holes of a pattern class are provided in section II.A.d, The
adopted procedure to find holes from a set of sampled points is discussed below.
Procedure : The procedure considers Ft axis as the base feature and the F2 axis
as the height. Correspondingly gj and £2 are considered as the base coverage factor
(£i) and height threshold factor (e/i) respectively. Windows are now generated using
the aforementioned approach. The sample points in each window are then arranged
393
in ascending order according to the height (Fj) sample values. If the difference of
height values of any two consecutive samples within a window exceeds £/,, then a
hole is assumed to be present between the said sample pair. Let h! and h" be the
height feature values corresponding to two such sample points. To illustrate this
finding, a pattern class with a hole is shown in Fig. 3 in which Px and Pj are the
sample points which satisfy the above condition and the corresponding window is
shown using dotted lines.
To detect the hole, the sample set is decomposed into two groups according to
whether the height values H.e., F2 values) are less than {h' + h")/2 or not. The
decomposition leads to finding two groups where none of them possess a hole. In
Fig. 3, the line with dashes indicates the split.
The aforementioned routine is repeated until every sample group is found to be not
containing any hole. When the subclasses corresponding to the sample groups are
combined in the boundary decider block [section 4], the holes will be excluded from
the final shape of the pattern class. It is to be realized that this procedure detects not
only the holes elongated across F2 feature axis but also the holes elongated across
Fi feature axis. Note that the proposed procedure can not detect the holes with
sizes less than Si and/or £2 corresponding to the feature axes Fj and F^ respectively.
Hence the detectable minimum hole size depends on coverage factors £1 and £21 which
in turn depends on the accuracy factor {St) and finally depends on the sample size
it).
It is also to be mentioned here that the procedure decomposes a training sample
set not only for a hole, but also for some particular type of concave boundaries.
For example, sample sets having shapes like 'C', ' ± ' or 'Z' may also be decomposed
although these do not possess any hole. But this does not create any problem in
the overall shape determining procedure, since such sample sets anyway had to be
decomposed in the pattern class sub-divider block. A
the windows from a sample set and consequently to find the boundary values are de-
scribed earlier. It may be recalled that the maximum and the minimum height values
of each window are considered as the boundary values in the upper and lower bound-
ary directions respectively corresponding to the height feature axis. The approach
to calculate the boundary variations is discussed below.
Calculation of boundary variations : To describe the approach, let us consider
the boundary variation values in a particular direction, say d {d E {!/, !«, 2;, 2u}.
The procedure to obtain the boundary values has already been discussed previously.
Henceforth, it is assumed that there are w windows and their boundary values are
//; {i = 1,2,... ,w). A boundary variation factor, denoted by V^ in direction d is
defined as
V, = EW-'^.-i)' /el (3)
where Ch is the height threshold factor for the direction d. Here the division factor
E.\ is used to make the variation factor Vj unitless.
Now let MAXH and MINH be the maximum and minimum of the boundary
values i/,'s (i = 1,2,... ,tu) respectively. If the difference of MAXH and MINH
does not exceed £fi (i.e., if {MAXH ~ MINH) < eh] then the sample set is assumed
to be nearly rectangular in shape. In such a case, the variation factor Vj is assumed to
be zero i.e., make Vj = 0. Otherwise the sample set is considered as decomposable
in the direction d.
Initially, Fi axis is considered as the base to generate the windows and F2 axis is
considered as the height to find the boundary variation factors V21 and V2 for the
coded directions 2; and 2,^ respectively. Similarly, by reversing the roles of Fj and
F2 axes above, the boundary variation factors Vi, and Vi„ for the coded directions 1(
and lu are calculated. Thus, the boundary variation factors for the four boundary
directions are obtained in this block.
VD = max{Vi„Vi„,V2„V2„}.
If Vf) = 0, then the sample set is assumed to be nearly rectangular in shape and it
is not further decomposable.
Otherwise i.e., if VD > 0, then it is assumed that the sample set is not nearly
rectangular in shape and it is to be decomposed into groups. Now from the direction
of decomposition (i.e., D) the windows with their base and boundary values, and the
corresponding height threshold factor Sh are recalled. The samples are then arranged
in ascending order according to the base values.
For making a cluster of windows, the maximum boundary value is found. The
starting window for the cluster is taken as that window which have the maximum
boundary value. The position of the starting window is noted. The following windows
395
from the starting windows are assigned one after another in the cluster until the
differences between the boundary values of the current window and the starting
window exceed the height threshold factor (ch). Similarly, the preceding windows are
also put in the window cluster. The samples lying in the window cluster are assigned
to the first sample group.
The aforesaid routine is repeated on the remaining windows until all the windows
are exhausted. This leads to the formation of window clusters. Every window cluster
results in a group of sample points. Thus, the given training sample set is decomposed
into a few group of sample points.
The decomposition procedure is appUed on the sample groups repeatedly until all
the groups are found to be nearly rectangular in shape. It is to be observed that
the larger the number of groups, the greater will be the accuracy of the shape of the
multivalued pattern class obtained. ^
/ a a \ \ S(x;^t,^ua) if x < a
x ( a : ; a , ^ , , / J . , 7 , , 7 u ) = | il^^aj^,-,^) \i x> a (4)
where
0 if X < a
1 (f5f) if a < X < 6
S{x;a,b,c) = < (5)
1 if X > c.
Note that unlike the standard S function [11-13], here 6 ^ a+c
2 •
MAXk +Mmk
Qk, =
MINU (6)
Ilk MINk 1.^, = MAXk,+ej ;
j = l , 2 ; fc = l,2,...,Tj
where Cj is the coverage factor for the j t h feature [Eq. (2)]. ^
397
B. Boundary decider
In the previous subsection, membership functions corresponding to all the seimple
groups along each feature axis are determined. Using these functions, each of the
subclasses corresponding to the sample groups is estimated and those are finidly
combined to obtain the estimated (multivalued) shape of the pattern class. All
planar points in the feature space are labeled with their degree of possibilities to be
in the cla.ss. To show the shape of a pattern class in the plane, the whole feature
range is divided into small rectangles and these are referred to as the Feature Space
Cells (henceforth FSCs). The size of all the FSCs are same and these are made as
small as possible such that each FSC can be distinguished in the feature space. Thus,
all these FSCs are labeled in terms of their possibility values to be in the pattern
class. The FSCs with zero possibility vaJue are considered to be outside the class.
The method of obtaining these possibility values is described below.
Procedure: Let (xi,X2)' be a typical feature values of such a FSC. Suppose fXk.
denote the membership value of the FSC corresponding to A:th [k = 1,2,...,?/)
subclass (i.e., fcth sample group) and jth {j = 1,2) feature, fik^ is calculated from
the corresponding TT function i.e.,
Now, the possibility, say 0, of the FSC to be in the estimated pattern class is
defined as the maximum of the membership values of the subclasses. That is,
0= max { Uk ] (9)
it=l,2,...,TJ '^ ^^ ' ^ '
Let r be the number of subclasses for which the combined membership values
(/i/t's) of the said FSC are positive. To incorporate the effect of the neighboring
subclasses with positive membership values in the estimated pattern class, tne value
of 0 is increased to 6^^'^ for T > 1. That is, when the possibihty values of the FSC
(/it's) are positive for two or more subclasses, then it indicates that the said FSC
has the possibihty to He in those subclasses, which in turn increases the possibility
of the FSC to be in the finally obtained pattern class. •
A method has been described above to find the possibihty Vcdue {&) of a FSC to
be in the pattern class. Note that 0 < 6 < 1. If the value of $ is zero, then the
FSC is considered to lie outside the pattern class. Otherwise the FSC belongs to the
pattern class with the possibility 6.
To obtain the complete shape of the pattern class, the aforesaid routine is repeated
for every FSC in the feature domain. Thus, all the FSCs are labeled with their
possibility values to be in the pattern class, and as a result, the multivalued shape
of the pattern class is obtained. 4
398
1 ^^•^ ^H X)
•a"
.a i
*3 iM
SI a
15 6^
a O
1 » s • o
«3
6)
•I
5" TJ
CI.
p
1
. —
' '
us
|i«
400
Figure f : (si) A set of training samples from the class in Fig. i>{a).
(02) Three decomposed groups of the trainiiig set in (ai).
I I i 1 1 T
F, In m
B. Speech data
To examine the practical applicability, the algorithm has been implemented on a
set of Indian Telugu Vowel sounds in consonant-vowel-consonant context uttered by
three speakers in the age group 30 to 35 years. Fig. 8(a) shows the typical feature
space of six vowel classes /S/, / a / , / i / , / u / , / e / and / o / with 72, 89, 172, 151,
207 and 180 samples respectively corresponding to the features Fi and F2. Here Fi
and F2 denote the first and second formant frequencies which were obtained through
spectrum analysis of the speech data. Details of feature extraction procedure can be
found in [13, 14]. The classes are seen to be overlapping and their boundaries cire
ill-defined (fuzzy).
The proposed algorithm has been applied on each of the six vowel classes separately
where the total available data are assumed as the sampled points. That is, for the
classes /6/, / a / , / i / , / u / , / e / and / o / , the number of sampled points are 72, 89,
172, 151, 207 and 180 respectively and correspondingly the accuracy factors are
considered as 0.12, 0.12, 0.10, 0.10, 0.08 and 0.10 respectively. Fig. 8(6) shows the
obtained multivalued shapes corresponding to the vowel classes / a / and / u / ; Fig.
8(c) shows the shape corresponding to the class / e / ; Fig. S{d) shows the estimated
shapes corresponding to the classes / i / and /of, and Fig. 8(e) shows the shape
corresponding to the class /6/. In figures 8(6)-(e), the FSCs with possibility values
(6) > 0.5 are represented by the corresponding vowel characters; the FSCs with 6
satisfying 0 < ^ < 0.5 are represented by dots and the FSCs with ^ = 0 are shown
as blanks. Observe that the estimated vowel classes are good representations of their
original classes. 4
A. Decomposition
A pattern class in ]R is represented here by a set of sampled points. The decom-
position section detects the geometric structure of the pattern class from the sample
set. The iV features axes under consideration are referred to as the first (-Fi), second
{F2), ..., A'^th {FN) axes respectively. Initially, the window generation procedure
402
g
«
O
a
JS 9^
.s s
o
o
BfwmWwmWffrf
9«!6«ie«!)
•Hi
• n^fl390 KSfffiG«l 163 E^fliffi3fiBj*«S^
H
.
1 - i g ^ f f i i Q f S B^rgaiigffi
. .Bj^SI . .
i m ^
u o
"i - 0) «
33P3
: : : [^^^ P 3 3 3 3. 3.31333
3333
3333 •
•i I '
,,.
m
-333
333S
•333 3332 3 ' •
a33S 3 • . 03
-sua 333S
•333 333333 3 3 3 3 3 3 3 ^ 3 3 3 3 3 3 3 • •
•333 333333 33333 3 3 3 3 3 3 3 3 3 3 • •
H
333333 3333:5 3 3 3 3 3 3 3 ^ 3 3 • •
. -33 3 3 3 3 3 3 3333:3
09
9
•H 3:
|i|
403
O t<3
"o o
.§
m
5 S fo
—r" s
o
a o
^ § £^
o --^
o i«
so ^ ^
g S-S
3 I
£ - 3 0
s - ^_g
HiiilHHHjl
::i:-!J:!:!:!:!:i:!,;--jiiii3it£
s
•H X
s
404
from a sample set in R'^ is described below. T h e proposed methodology for finding
the shape is then described using the window generation procedure.
Generation of windows : A window based approach is adopted here to find the
boundary variations of a sample set in iR^ and the approach is more or less the
same as t h a t described for M in section / / / . Initially, one of the feature axes is
considered as the base feature and the corresponding coverage factor is referred to
as the base coverage factor (ej). T h e training samples are first of all arranged in
ascending order according to the base feature values. Then, depending on the base
feature values and the base coverage factor Cb, t h e same procedure, as stated for M^,
is followed here to generate windows. These windows are such t h a t base coverage
length of each window would atleast be £t.
Here for a particular window, the base feature values of the samples are assumed to
be the same and t h e rest (A'-l) features for the samples may take any value. So the
generated windows are visualized here to belong in ]R^~^ feature space; although the
original sample set belongs to iR . I n this sense, the proposed method of generating
windows always results in reducing the dimension of the sample set by one. In this
context, note t h a t , the initial training sample set may itself be considered to belong
to an TV-dimensional window.
Based on the aforesaid concepts, the multivalued shape determining procedure is
extended to JR^. T h e approach has been applied on a sample set in JR'^ in the
following way. To start with, the feature Fi^ is considered as the base feature (axis)
and correspondingly a few windows m M ~ are formed. For each of these windows,
another axis F;, is considered as the base feature and windows in R'^~^ are formed.
Repeating this process sequentially with Fi^, F;,, . . . , i^„_i as the base features
axes, some one-dimensional (l-D) windows are generated eventually. T h e values of
the left out feature F^^ are considered here as the height values. T h e maximum
and the minimum height sample values are found for each window and these are
taken to be the upper and the lower boundary values respectively for t h a t window.
T h e combination of the upper boundary values of all l-D windows, generated by
sequentially taking i ^ j , F^j, . . . , Fi^_^ as the base axes, highlights the upper boundary
of the training sample set in the Fi^^ feature direction. Similarly, the combination of
the lower boundary values provides the lower boundary of the sample set across F,-„
feature axis.
a. Hole detector
the holes are excluded from the final shape of the pattern class. •
'1(2)
3 : 1 3 : d rr3 : d
vl(2) EK
*=i
(2) '-)' /4
ji = l , 2 , . . . , 9 i and de {/,u}.
Similarly, by taking F3 as the base feature, every sample window VKj' (gener-
ated by assuming Fj cis the base on the original sample set), gives rise to g^Jgs
407
(a) ie)
i ' '••,..
w (f)
If F3 is first considered to be the base, the initial sample set results in qs 2-D
windows «'ith Fi and F2 as the feature axes. A typical windows is marked in Fig.
9(e) and its 2-D view is provided in Fig. 9(/). Corresponding to each of the §3 2^
windows, the boundary variation factors K 5 A ' ' * , ''^li)" ' ^^ ^*(2)' ' ""^ •^3(2)
(is = 1,2,..., ga) are determined.
Thus, there are in total 4(gi + f2 + fs) boundary variation factors corresponding
to a sample set in I^. All these variation factors are analyzed in the next block i.e.,
pattern class sub-divider block. •
This block analyzes the boundary variation factors to determine whether the training
samole set is to be decomoo.sed or not. To decide this, it finds the maximum of aJl
408
B. Fuzzy processor
A training sample set is decomposed in the previous section into few groups of
nearly parallelepiped shape. Here the subclasses corresponding to these groups are
determined separately and finally these are combined to obtain the multivalued shape
of the pattern class.
b. Boundary decider
To show the shape of a pattern class in the feature space, the entire feature range is
divided into small units of parallelepiped shape and these small units are referred to
as the Feature Space Cell or FSC. The size of all the FSCs are same and these are
made as small as possible such that each FSC can be distinguished in the feature
409
Space. All these FSCs are labeled in terms of their possibility values {0} to be ia the
pattern class.
Let (a^i, a;2,..., x^)' he a typical feature value of such a FSC. The membership
value (nkj) of the FSC corresponding to tth (k = 1,2,,..,»|) subclass (i.e., tth
sample group) and j t h (j = 1,2,... ,iV) feature is calculated from the corresponding
w function using Eq. (7). The combined membership (^i) of the FSC corresponding
to i t h {k = l,2,...,i]} subclass is defined as the geometric mean of pk/s i.e.,
The
. . „ possibility
r - " - — . - value
~ — (0)v / of the FSC is then determined using° -Eq.
- (^)- Depending
on the membership values •alue of the FSC in the neighboring subclasses, the value of 0
is similarly modified as done for M^. Thus, by labeling all the FSCs of the feature
domain with their possibility values, the multivalued shape of a pattern class in JR'
is obtained. The implementation of the extension concepts is provided below. •
5"?
s^m-.- V
:??'' i l i i s . •• \ • •\. C.J.2
\ / " p l k ^ ••' .
' M 3
O oj .
** B J?
+3
^ el i2
- mm. .|iiftV:^. S « A
,:^r/-'m±.\-.-' •
IK^il, o |:;Q>
>J o -
. , O lO
i' "o
' 'T3 d
vv-v-/7;V..
• y :,• , i ;.....;••
/^;;; ;rv'
.£?3AI
h "o -as
2 >
m "-S
g «
"go
R E A
S o .
•• a
»-l !:• d
o "A'
h
411
A. Experimental verification
Two artificially generated pattern classes have been considered in this section to
demonstrate the convergence property of the proposed algorithm. The first one is
a disc [Fig. 13(o)] with radius 2 and centre at (3,3). Five different sets of data
are chosen randomly from it with sizes 50, 100, 150, 200 and 250. The algorithm
was applied assuming these five data sets as sample points with accuracy factors
(Si) 0.16, 0.13, 0.10, 0.08 and 0.07 respectively. The corresponding estimated
multivalued classes are shown in figures 13(6)-(/).
The other considered pattern class is a sphere [Fig. 14] with radius 2 and centre at
(3,3,3). Four different sets of data are chosen randomly from it with sizes 150, 300,
500 £ind 1200 respectively and the values of 6t are considered as 0.20, 0.16, 0.13 and
0.10 respectively. Here also three levels of extracted classes are shown corresponding
to 0 > 0.5, 0 > 0.25, 0 > 0. Figures I5{a)-{d) show the estimated classes with
412
fa.) (W
(C) id)
(e) if)
Figure 13 : (a) A circular class; (l>)-(f) correspondiag estimated versions based on 50,
100, 150, 200 and 250 samples respectively.
413
Figure 15 : (a)-(cl) Estimated versions of the class in Fig. 14 with if > 0.5
based on 150, 300, 500 and 1200 samples respectively.
414
"^8
3 o "5
Oin>
' m '~' -w
<- r^ o
«
> 9
O a)
o,
o "S-
I A i
3<S8
•^^ -^
^•^—'-^ T 3
"^ S
5 6?
PC4
09
3o
O'-' >
' ^ 13 •?.
SJ O SJ
>T3 O.
, 43 «
T ) m <a
? _ I*
.^ IN g.
wo S
2."* o
iH •
bO
tlOCi
415
9 > 0.5 based on the selected sample sets of sizes 150, 300, 500 and 1200 respectively
corresponding to the pattern class in Fig. 14. Figures 16^a)-(<f) show the estimated
classes with 9> 0.25 based on 150, 300, 500 and 1200 traming samples respectively.
Figures n{a)-{d) show the estimated classes with ^ > 0 based on 150, 300, 500 and
1200 training samples respectively.
It can be seen from these results that as the sample size (i) increases the estimated
classes are gradually converging to the original pattern class. A
a. Hausdorff metric
Normally, to find the similarity between sets, a distance measure is often used. Haus-
dorff metric [15] has been used here for this purpose.
Let {X, d) be a metric space. For any compact subset of ,4 of A^, define
where inf means infimum. Note that 6(y, A) is finite and BXQ G A such that S(y, A)
= d{xo,y). Now the definition of the Hausdorff metric is given below.
Definition 3 [15]: Let A and B he two compact subsets in M^ . Then the distance
between A and B, denoted by Dist[A,B), is defined as
The boundary of the disc [Fig. 13fa)] is approximated by 180 equally spaced points.
This set of 180 points is considered here as the set B. Note that the estimated classes
are multivalued. Hence, in order to apply this measure, three levels of estimated
boundary based on the possibihty values (#), namely 0 > 0.5, (? > 0.25 and # > 0
are considered. The values of the Dist measure are shown by a graph in Fig, 18(a).
> > 0.5 > .1 0-25 S > 0 I i > O.S • > 0.25 jl > 0
HI
--0 son 1200
Sampl« Size (Estimated Class) Siie {Estimated CiMJ)
(a)
Figure 18 : Values of Dist measure (a) between the estimated classes in figures
13(l»)-(/) and the actual class in Fig. 13(n); (b) between the actual class
in Fig. 14 and its estimated versions.
The distance measure Dist has also been applied on the estimated sets or classes
[figures 15(a)-(d), 16(a)-(d) and 17(a)^(d)] with the original set in Fig. 14. ^ The
boundary of the sphere is approximated by 5675 equally spaced points and this set
of 5675 points is considered here as the set 5 . Here also three levels of estimated
boundary based OE the possibihty values (fl), namely 0 > 0.5, 0 > 0.25 and 0 > 0
are considered. The values of the Dist measure are shom'n by a graph in Fig.
18(6). *
Here the first term denotes the average similarity of the dements of A to B, the
second term denotes the average similarity of the elements of B to A and the last
term Dist{A,B) denotes the overall similarity between A and B. Sim is a metric
[9, 10]. •
417
The defined metric Sim has been considered as another criterion for goodness of fit
in the proposed shape determiniDg procedure. It has initially been applied between
each of the five estimated sets (with three levels of boundaries as # > 0.5, 0 > 0.25
and 0 > 0) [figures 13(ii)-(/)] and the original set [Fig. 13(a)] in the same way as
the previous case. The values of the Sim measure are shown in Fig. 19(a).
(4 W
Figure 1§: Values of Sim measure (a) between the estimated classes in igures
13(i»)-(/) and the actual class in Fig. 13(a); (b) between the actual class
ifl Fig. 14 and its estimated versions.
The metric Sim has finally been apphed between the pattern class in Fig. 14 and
its estimated multivalued classes [figures 15(a)-(d), 16(a)-(d) and 17(a)-(d)]. The
values of the Sim measure are provided in Fig. 19(6).
Hence, the convergence property of the proposed shape determining procedure is
established both experimentally and analytically. •
VIII, Conclusions
In this article, we have described an approach to compute the multivalued shape of
a pattern class from its sampled points. The concept of fuzzy set theory has been
found to be appropriate here. In estimating the shape, the portions not covered by
the sampled points are assigned some fuzzy membership values denoting the degrees
of their belonging to the actual class. Therefore, unHke the conventional approaches
the proposed method does not attempt to provide crisp boundary from incomplete
sample set. The effectiveness of the methodology has been demonstrated on some
artificially generated data sets (in JR^ and JR^) and also on the real life speech data.
The multivalued shapes can be converted to the usual crisp versions by considering
only the feature points (FSC) with possibihty value (0) > 0.5, say to be within the
classes. The convergence of the estimated shape to the original one has been verified
both experimentally and analytically. la this context, a new metric Sim has been
defined which can be used to find the similarity between any two finite sets.
The parameters (like accuracy factor, coverage factors etc.) required for determin-
me the shaoe of a class are determined automatically from the sampled points. One
418
Acknowledgement
T h e authors gratefully acknowledge Professor Sankar K. Paliov his valuable guidance
and suggestions in this work.
References
[1 aH.plane,"
Edelsbrunner, D. G. Kirkpatick, and R. Seidel, "On the shape of a set of points in
IEEE Trans. Inform. Theory, vol. IT-29, pp. 551-559, 1983.
[2 R. A. Jarvis, "Computing the shape hull of points in the plane," in Proc. IEEE Comp.
Soc. Conf. on Patt. Recog. and Image Process., pp. 231-241, 1977.
[3] S. K. Akl and G. T. Toussaint, "Efficient convex hull algorithm for pattern recognition
applications," in Proc. 4th Int. jt. Conf. on Patt. Recog., Kyoto, pp. 483-487, 1978.
[4] J. Fairfield, "Contoured shape generation forms that people see in dot patterns," in
Proc. IEEE Conf on Cybem. and Soc., pp. 60-64, 1979.
[s; G. T. Tousscdnt, "Pattern recognition and geometrical complexity," in Proc. 5th Int.
Conf Patt. Recog., Miami Beach, Florida, pp. 1324-1347, 1980.
[6; C. A. Murthy, On consistent Estimation of classes in M^ in the context of cluster
analysis. PhD thesis, Indian Statistical Institute, Calcutta, India, 1988.
[7: F. P. Preparata and M. I. Shamos, Computational Geometry : An Introduction. New
York: Springer Verlag, 1985.
[8] U. Grenander, Abstract Inference. New York: John Wiley, 1981.
[9] D. P. Mandal, A Multivalued Approach for Uncertainty Management in Pattern Recog-
nition Problems Using Fuzzy Sets. PhD thesis, Indian Statistical Institute, Calcutta,
India, 1992.
[lo: D. P. Mandal, C. A. Murthy, and S. K. Pal, "Determining the shape of a pattern class
from sampled points in iR^," Int. J. General Systems, vol. 20, pp. 307-339, 1992.
[n L. A. Zadeh, "Fuzzy sets," Information & Control, vol. 8, no. 3, pp. 338-353, 1965.
L. A. Zadeh, "An outline of a new approach to the analysis of complex systems and
[12; decision processes," IEEE Trans. Syst., Man & Cybem., vol. SMC-3, pp. 28-44, 1973.
[14 S. K. Pal and D. Dutta Majumder, "Fuzzy sets and decision making approaches in
vowel and speaker recognition," IEEE Trans. Syst., Man & Cybem., vol. SMC-7,
pp. 625-629, 1977.
[is: K. Kuratowski, Topology, vol. I. New York: Academic Press, 1966.
[16: D. P. MEUidaJ, C. A. Murthy, and S. K. Pal, "Formulation of a multivalued recognition
system," IEEE Trans. System, Man & Cybem., vol. SMC-22, pp. 607-620, 1992.
This page intentionally blank
PART 5
APPLICATIONS
CHAPTER 16
1 Introduction
In the last twenty five years a number of new control techniques, such as fuzzy logic, have been introduced [1].
Fuzzy logic has met with some success in the process control industries and a number of papers have been published,
typically Daley [2]. The work presented here enhances the published work by investigating the application of parstllel
processing to real time fuzzy logic in a project carried out by the authors. INMOS transputers [3] are used as the
building blocks for the multi-instruction, multi-data parallel (MIMD) computer employed in this project.
2 System Hardware
The project utilizes a small Mitsubishi industrial robot, the RM-501 [4], with the proprietary control unit removed
to allow direct access to the actuators, optical encoders and joint boundary detection switches. Counter cards
have been introduced to register the output from the optical encoders employed to measure joint position and each
actuator is controlled directly from a transputer, with voltage as the manipulated variable.
A three INMOS transputers network resident in a SENSION B004 system [5], was atdopted because it would
allow adequate software developement for the project envisaged. The "operating system" was the Transputer
Developement System [3] with Occam II [6] as the programming language. A 8048&'microproce88or based IBM
compatable computer served as the front end to support the transputer network.
The three transputers used in the network are a T400 transputer resident upon the GBUS board [6] providing a
memory mapped interface to the robot, a T&OO root transputer and a T800 host transputer. The host transputer is
connected by an interface card to the IBM computer used to provide the necessary user interface and disc storage
facilities. The digital to analogue cards (DACs), counters and joint boundary detection switches are connected to
Peripheral Interface adapter (PIA) cards which are memory mapped onto the INMOS T400 transputer by the GBUS
card, allowing the PIA cards to be accessed by the appropriate read and write operations to specific locations in
memory. The PIA cards conform to the interface protocol implemented in the GBUS and are called GESPIA cards.
A schematic representation of part of the system used is shown in Figure 1.
3 Fuzzy Logic C o n t r o l
Zadeh [1] first introduced the concept of fuzzy conditional statements and using these statements Mamdani [7]
developed his fuzzy logic controller. The Mamdani approach to the construction of a fuzzy logic controller has
formed the basis of the approach adopted by the authors. In this work the domains of discourse over which the
fuzzy logic controller is defined are integer and finite. The input domains are error ( UE ) and rate of change in
error ( URCE ) with the output domain being change in control action ( Ucv )• The integer range was selected for
practical reasons as the information read from the robot optical encoder counter cards and written to the DACs is
421
TOANSPUTEH
COMMUNICATION LINKS
MEMORY MAPPED
INTERFACE
SESPIA'CWO- GESHA-CARO'^
integer, If desired, the discttssioii can readily be geoerdized to domains delned over the reaj number universe. At
each sample interral it there afe two controller inputs, error (e(i) e Mjs) and rate of change in amt (rce(t) 6 UHCE)
with change in manipulated control variable (w(i) e Mcc?) « the output. Eate of change in erioi and not change in
error is used in order to avoid the need to rcdeine the fuziy sets when a differeot sampling rate (1.) is used. The
variables e{k) and rce(k) are calcmlated from the demand {d(k)} and the joint state (t(i)), ndng the expressions:
By adopting u ( i ) m the output there is no risk of large changes in the value of the manipnlated controlvariable,
even in the presence of a "rogue" rule. The manipnlated control variable (c(i)) is evaluated from the expression:
A fuszy logic controller is built mp from a number of rules, typically in this project between 69 and 289 rules. The
i'th rule has the form:
(4)
The output at sample interval k ftom all the rules are combined iiaiiig the maximiiin operator y to give a single
fuiiy set, FGH) £ Ucu- The deterministic cbaoge in the output control variable u{k) is usually calculated from
FC{k}. The membeirsbip grades of FC{k) are defined by the expiession:
K
i=\
The deterniinlstic change im control vaiiable output at sample interval t is calculated from FC{k) by the defrazifi-
cation process, DF{,)t
u,{k) = DF(FC(k)) (8)
E x a m p l e 1 Consider Ikt example of the cftararferisfie fimctian shown in Figure S, defined ever an integer iemsim
of discourse. The pammeiers required to define this faneiion asssming a tmpazoiial temflaU are (5,14,25,37).
MEMBEHSHIP GRADE
DOMAIN OF DISCOURSE
The development of this method of storing and using fuzzy sets allows for the defining of fuzzy sets directly over
the appropriate domains of discourse. Hence, the problem of finding the appropriate values of the input and output
gains that would otherwise be required is removed. However, the definition of the appropriate fujssy sets and rules
is still critical.
In the software developed during the investigation, tlie template parameters were stored in a data base, termed the
Fuzzy Set D a t a Base (FSDB). This data base consists of an area of memory where the parameters are held together
with a process to manage this aiea of memory and to execute the appropriate read and write operations. Consumer
processes request data from the management process, which replies with the requisit data. The relationship between
the FSDB and a consumer processes is illustrated by Figure 3, where a process is represented by an ellipse and the
area of memory holding the parameters by a tank. The linking of each process is by the point to point communication
channels [10] supported by the Occam language, which are shown as arrowed lines crossed by the symbol L This
representation is similar to that defined by MASCOT [11], a real-time software design methodology. For the system
424
V
^^IW.fl^|f A^fj:-^
Fym mi
Joint angie/radians
Figure 4 Gomparision of joint responses produced by firaiy logic controUers using triangular amd
rectangular output fuzzy sets
described, it was found by experiment that the shape of any output fuizy set is idatiYely unimportant and hence
it was decided to use rectangular output fuiiy sets. This is illustrated by the graphs presented in Figures 4 and 5
which compares the effects of using rectangular aiid triangular diaracteristie output sets. The results displayed in
Figures 4 and 5 indicate that although two controllers may differ in their internal structuie they can still behave
similarly. This is an area worthy of further investigation as such differences can have signiicant affects upon the
computational efficiency of the controller, it is always preferable to nse computationally efficient algorthms in all
real time control applications.
V„.:.;;,,/^-:
-5
: ii\J' ^
10
m
' • " ' - . ' " " ' -
- " •• '"1
Figure 5 Comparision of «>nt«jl actions produced by fuiiy logic controllers using triangulax and
rectaiigulac output fuzsy sets
m
2. Find the smidletit element yeU such that pA{y) = ". This element in t a m e d the reference point of fumzy set
A, Ref{A).
This scheme is realiiable because of the operational domains over wbich the fuz^y logic controller operaks. Gener-
alisation to variables of type real is a simple operation, if desired. This method was chosen for three reason:
• Ease of implementation.
• Computational requirements.
• Rigour of approach.
E x a m p l e 2 Consider ike three fuzsy sets shown in Figure 6. T*e reference peimto forfiazf sets A, B md C are *,
go mi S4 mspecUvely. The erier ani lafcfj of A,B and C with rasped to eack other is 1, S and J.
This ordering scheme can lead, in certain circumstances, to some minor anomalies depending upon the nature of
the fuzzy sets involved. Any "reasonatle" set of "meta-niles" governing the definition of th« fuszy sets and the
rules employed in the controller will not allow for such anomalies, though their presence will still be of no practical
importance.
From equation 4 it is seen that each rule Mi is built, up from three fuzzy sets Bj,, iiCJS,-, and CUi,. The labels for
these fuzzy sets can be written in an ordered triple which then describes the rule. For example, rule Ri m represented
by the ordered triple (ii,i2,>s). The fuwiy logic control policy, defined by a number of rule., is then described by
an equivalent number of triples. Storing these triples in a structured manner aids the search for rules contributing
to the controller output, a more detailed description of wbich is given in Section 3.4. A segment of a typical rule
base is illustrated in Figure 7, a cross indicating that no rule is present. This segment of a rule base is equivalent
to the following fuzzy conditional atatements, where the subscripts are the labels shown on the appropriate edge of
the grid in Figure 7.
I F Eo A N D RCEa T H E N CUw E t S E
I F El A N D RCEt, T H B N CUa E t S K
I F Eo A K D RCEi T H E N CUm E L S E
I F El A N D RCBi T H E N Ct/ia E L S E . . .
426
4 8 12 1« 20 24 28 34 DOMAIN OF DISCOURSB
u(m) = DFMOM{FC(k})
1 "
= Z-I]l*FC(i)(e«(«)) (10)
The centre of moments defuziification process usee all the membership grades of FC(k) in calculating the deteimistic
cotttrol output. If t i e output fuzzy set is thought of as a two dimensional shape, the deterministic output produced
427
by DFCOM(.) 18 the centroid of the shape produced by the action of the maximum operator on the outputs of the
individual rules and is defined by the expression:
The advantages and disadvantages of the previous two methods are well documented [13]. The centre of gravity
defuzzification strategy is a novel defuzzification strategy developed for this project. Its difference lies in the use
it makes of the outputs from the individual rules and manner in which u){k) is calculated. The DFcoAti) and
DFnfOM{-) defuzzification strategies wilt allow the output from a single "rogue '* rule to affect the deterministic
output in a disproportionate manner. This results because a single rule can drastically alter the shape of FC{k)
and hence the value of uf{k), see equations 4, 7 and 8. A better approach is to calculate the deterministic output
in a manner which places more reliance upon a consensus between rules. This approach reduces the effect a single
rule can have on the value of Li;(jb). The centre of gravity defuzzification strategy achieves this by using the output
from each rule and not their product under the maximum op«rator, FC{k). The output u>{k) is the centroid of the
individaul outputs of the rules and not the centroid of FC(k). Improved computational efficiency is an additional
benefit since the deterministic output can be calculated directly from the inputs to the controller and the fuzzy
set templates. Analytical expressions can be easily determined and two examples of which are presented. This
defuzzification technique has three important advantages over the methods more frequently applied:
• Ease of computation.
• More representative of the decision reached by individual rules.
• More robust in the presence of badly chosen rules.
Example 3 Lei u(ib) € Ucu (^ ihe deterministic change in control action at sample interval k and the moment of
the output for rule /J, be m , ( i ) , as illustrated in Figure 8. let the p contributory ruUa be labelled Ri,..., Rp. The
moment of the output with respect to ui(ir) is m,(ik). Hence, for all the rules by definition ofu{k) as the centroid:
£"».(*) = 0 (12)
For the Hi, the value of mi(k) can be calculated. For the output shown in Figure 8, a rectangular output fuzzy set:
J2m,{k) = 0
i=l
P
Hence 5]|/.7.i(it)[2-Ai(t)] K * ) - a,J = 0 (18)
428
MEUBEnSHIP GWOE
DOMAIN OF DISCOURSE
m.m
Figure 8 Centre of gravity defuszificafcion strategy for rectangular output fussy sets
MEMBERSHIP GFIADE
h,-4-.
L ^
wik)
DOMAIN OF DISCOURSE
Figure 9 Centre of gravity defuszification strategy for triangular oMtput fuzzy sets
429
It can be seen fiom the two previous examples that the expression derived for u(k) can be updated as eaeli
contributory rule is discovered. T h b explains the improved computational efficiency of the novel defuziification
strategy. The DFcoci-} is far less sensitive to the presence of a rogue rule, and can in some ciicumstances behave in
a manner similiar to the DFcomi-), as illustrated hj Figures 10 and 11, although some divergence in behaviour will
ol>viouBly ooaur as the nmmber of contributory rules increase. The calculation time for the DFcoai-} is not affected
by the shape of the output fujiy sets when the template protocol is used, as indicated by the results presented in
Table 1. It is possible to derive expressions for the determini»tic output for other output ckaiacteristic function
shapes if so desired.
Joint slala/radians
0.05
10 20 30 40 50 60 70 80 « 100
Sample intareal
Flguee 10 ComparisioD of joint states produced by a fuzzy logic controller using DFCOGQ and
DFCOM(^)
FC(k) Update tediiuque TViajigular output fuzzy sets Rectangular output fuzzy sets
Modal evaluation timc/ms 0.64 0.64
Minimum evaluation time/ms 0..51 0.51
Maximum evaluatioo time/ms 1.1,5 1.09
Range in evaluation times/ms 0.61 0.58
Table I Comparision of evaluaiion times using rectangular and triangular output fuszy sete
430
Control action/¥olls
10 20 30 40 50 60
Sampi9 inlereal
F i g u r e 11 Comparisioii of control actionH produced by fnszy logic comtfoller using DFcoaQ and
DFCOMQ
I I I I I
R3
Is
membetsiip grade for a fmsiisy set, the rules employing that fuazy set are not evaluated because the output fuzzy sets
will have meiiil>e»hip grades that are all equal to zero. For example, enppcBe the input to the controller at sample
period k is:
Suppose tliat:
!*«(,(«(*)) = 0 where t'l = 2
Then fn(jt) = »
Hence, |Jrc,(*)(<=") = 0 V CM 6 Ua)
Then if foi some other rule B j , jEj, = Si it cam be seen that hj{k) = 0, hence:
ffFCiityiev) = 0 V c« e Ucv
That is, rule Rj will make no contribution to the output. Tliis will be true of amy other rule using the fuzzy set ,i?j,.
A similiar arguetnent will hold for the fnj«y sets deined over UHCE- T M " explanation illustrates why the SSP will
mo¥e to seaidi the rules in the next row or column, dependant upon whether the fuzzy error set or the fuzzy rate
of change in error set haw a zero membership grade for the current input. The SSP anchor point is static and has
been set to the (0,0) location in the EDH. For this example, twelve fuzzy set membership grades were evaluated by
the SSP during the search of the B B l .
IttO)
jp.0.7} (0,0.31
_m_ P.e>
Tke first coniHion is satisfied what all of the rsks havt teea tetied; a properly mhviucei as a defavH stop conditim.
Far ofif fttrtieatar infvi there tj a n»aj;ifn»in M%mher of cemtriisioTy raks defined I j the order of the cantmller and
the iefnzzificaiion process. When the maximim is reached the search is terminaiei.
The mnderlj/mi reasoning iehini lit third coniiliou is more cempki and results from the, relationship hetween
fuzzy sets. Let e(k) € Ue ie an error input to ike conlraller and E„ £ UE be the first fszzjj set defined ever the
errvr domain that has a non-zero membership §rmie for e.{k). Then any other fsziy set, ssf E'l £ UE, which has
a noa-zero memierakip grade for e{k) will be "clo.se' to E„; thai is, tis rank I is 'similiar' to w. Let A he a /sizf
set defined ever a rfotnain 0/ discourse U. Then the number of ampperl sets which iBtersect Supp{A) is denoted Jf
S,(A).
E x a m p l e 5 Consider the fuzif sets illustmted by Pignre 17 and iepnei omr the error domsin; thef are typical of
those in the controller developed in this project. For the input e(k) = 13, liE,ieih)) = 0.5, and Si{Ei) = 2 for the
fnzzg error sets shown, it is only necessarg to evahaie the membership grades of ike .^ets Ei,E2,E4 and E§; a
rtsslt dieieiei ig the order of the fmzzji logic coniroUer.
MEMBERSHIP G R » E
8 « 10 1 ^ 16 22 24 26 28 32 36 38 42
I DOMAIN OF DISCOURSE
F i g u r e 17 Deinition of S/(-)
The ESP is the most sophisticated of the three searches irnnlemented. The approach utilizes the B D l ordering
in a manaei similar to the SSE The additional knowledge, however, of the location of past contributory rnlea is
also employed to locate an anchor point from which to start the search for the next group of contributory rales.
A default anchor point position is available when no previous contributory rules exist, this occurs during the first
sample period of the controller.
In addition to employing this extra information, the actual pattern of the search is different. The search does
not step through rows or columns of the EDB but ripples out from the anchor point, fust searching for non-zero
membership grades close to the anchor point, llien progressing to the more distant fuzzy sets. The iow diagram
given in Figure 18 illustrates the operation of the RSP and Figure 19 gives a more detailed description of the stop
condition. Sets that have labels smaller than that the corresponding set at the anchor point are described as to the
left and those to the rif Al have higher valued labels. Separate software roiitiiiei have been developed to search to
the left and to the rigM. The operation of tlie routine M G H T . S T E P is shown in Figure 20 and the structure of
the l E F T - S T E P routine is similiar. The C O N T I N U E process tests the stop conditions illustrated in the iow
diagram Figure 17.
The stimulus for a change in search pattern came from the experience gained with the SSP. This experience
indicated that contributory rules in successive sample periods are bnilt up using the same or neighbouring fazzy
sets. The labels for the fuszy sets used in deining a typical sequence of contributory rules, generated by the controller
incorporating the RSP, are shown in Table 2. Starting from an anchor point, the HSP looks for error sets that have
a Eon-xero membership grade for the input, the RSP Irst testing the membership grade of the error set at the
anchor point and then on either side of this location. Pollowing this, the RSP evaluates membership grades of sets
more distant from the anchor point. This process is illustrated by Figure 21 on a typical collection of fuzzy sets. In
435
,TBU6 ^
TRUE
Sample interval E fuzzy set label RCE tuay snot label CIJ fuzzy set label
838 12 1 7
839 12 1 7
1 8
n
840 12 1 7
11 1 9
841 12 1 7
12 2 7
11 1 9
11 2 8
step 1 the aachof point fussy set (C) is t ^ t e d for a oon-sero membership grade for the input. Following that, the
search process jumps left, in step 2 and then tigh-i in step 3. Having found a fuKzy set with a non-sero membership
grade for the input, the mMch proems continues the search, for any other non-^ero membership grades. Due tc> the
ordering of the fuzsy sets^ the search continues by jumping to the right again. The EDB search is terminated using
the same stop criteria as outlined for the SSP.
The examples given show the operation of the different rule search processes and illustrate that each proems
km a dramatically different operation time of eaacli process. In Figuras 12j 14 and 16 the aumber of r u l « which
ar<j evaluated are 42, 12 and 6 respectively. The ratio of these nurabere aie of the same order of magnitude as the
calculation times required for the fusgy logic controllei incorporating these different search strategies^ the rule h a ^
remained the same in each CIK©.
The r^iilts given in Tables S and 4 were generated WiSing two rule ba*^ containing, one containing 81 and the
other 289 rules, Prraented in the tabira are the modal, minimum and maximal values of the controller calculation
periods for combinatioiia of rule search and rule base sise. The results show that the FSP is the least efficient of
the thrM! search processes, the SSP is more efficient and the RSP is the most efficient with the least variation in
calculation kimm; the benefit to be gained from using a more sophisticated gearch strategy. The evaluation time of
the fuzsy logic controller with the ESP is shown to be relatively unafffi«;ted by the sise of the rule data base.
One property of the ESP is that the time spent searching the EDB will be relatively independant of the number
437
DOMAIN OF MSCOUHSe
of rules or sets prewJiit m the fuzzy logic controller. This results ia the fiizgj logic controller haYing a relatively
constant evaluation time whatever the size of the BDB, an important property in the application to the control
of dynamic systems. This property will be of added importance in seltorganising fuisy logic controllers where the
number of rules apd fumj sets can change.
Table 4 Search process evaluation times for a 289 element rule baae
3.S C o m p e n B a t i o n for A c t u a t o r D y n a m i c s
The Fuzzy logic Controller (FLC) so far discussed has e(ilr) and rce(i;) as inputs, hence it is not capable of
compensating for the characteristics of the robot joint actuator. A typical speed-voltage curve for a robot dc-servo
motor actutor is given in Figure 22. The dead-uone represents upto 20% of the total voltage range [16] and can
adve»ly affect the system response if not compensated for by the control policy. The use of fuzzy logic offers two
distinct methods for introdacing the necessary compensation.
In the extended fusay logic controller, the individual rules of the fuzsy logic controller are altered to take into
consideration the characteristics of the joint actuator. The additional variable required for incorporation into the
rules is the value of the manipulated control variable c(k). The more sophisticated fuzsy conditional statecoents
438
MOTOR VELOCtlY/
RADIANS PER SECOND
VOLTAOE/VOLTS
required in this case have been discussed by Zadeh [1]. In this particular case the inputs at the k'th sample interval
are e(ib), ri:e(ib) and c(k — 1), the latter being the value of the manipulated control variable in the previous sample
interval. The fuzzy conditional statements used are of the form;
IF (error IS POSITIVE SMALL) AND
(rate in change of error IS POSITIVE M E D I U M AND
(current control variable IS NEGATIVE SMALL) THEN
(change in control action IS NEGATIVE MEDIUM)
where, POSITIVE.BIG S UE
POSITIVE-MEDIUM £ URCE
NEGATIVE.SMALL £ Uc
NEGATIVE.MEDIUM £ Uc
Using previously introduced notation, the characteristic function of the output from the i'th rule is defined by:
An alternative approach is to divide the fuzzy logic control policy and the corresponding calculations into two
separate processes or components, creating a distributed fuzzy logic controller (DFLC). This approach is based
upon the chain of reasoning, discussed by Zadeh [1], with each component equivalent to a block of reasoning within
the chain. This form of controller has been termed distributed fuzzy logic controller, since the control policy for
each joint is computed by two distinct algorithms.
The first of these calculates, or reasons, from an initial premise based on knowledge available. This knowledge is
stored in the rules used by this process. The initial premise consists of the values of e{k) and rce{k); the resulting
output is a suggested change in control action (^/(ib)) which is used to produce the suggested new manipulated
control variable C(i) and cui(k), where:
c/(i) = c(t - 1) + cu/(t) (21)
The second of these processes continues the reasoning using as additional knowledge the known behaviour of the
joint-actuator shown in Figure 22. The input, or premise, for this second block is the output of the first process.
The fuzzy conditional statements in the second component of the DFLC have the form:
IF ( ci(k) IS SMALL ) AND ( cu/(i) IS POSITIVE SMALL ) THEN
( cu(t) IS POSITIVE M E D I U M )
439
The diitributed controUei can now be seen to consist of two sets of fnsisy conditioaal statements. This method
allows complex control policies to be implemented by a series of simpler processes. This apptoach conformi to the
Zafcdeh [1] idea of how the human mind operates, An initial premise, generated from an input, is modifled hj later
reasoning based upon additional information.
3.6.1 C o n t r o l of Variable C h a t t e r
In the initial implementation of this component, the manipulated control variable was switched directly across the
dead zone using the following algorithm:
0.3
; : .- || . ..: : ; • • :•'• "•' •': . ' " : ' -
-,• - . - - . / . ^ i . . - . , • .
0,2 -•• - / t
HWIBWBMIBMMiBHI
These results were generated from an input demand to the shonlder joint of 0.343 radians. An oscillation of
±0.07 radians around tlie desired value, with a period of approximately 0.5 seconds, can be observed in the data
displayed in Figure 23. For negative voltages the output torque from the motor will oppose the gravitational force
exerted on the shoulder joint which explains the negative bias in the values of the manipulated control variable. The
440
dead band has been highlighted in the graph, the zone between -3.? volts and +4 volts, where shatp changes in the
manipulated control variable can be observed due to the action of the second component of the DFLC.
The problem of chatter was overcome by using the l i s t component of the DFLC to alter the value of c(i) IB
the small zone around 0 volts, [Di„J?J C Uc- By experimenl., the authors discovered that Di = - 0 . 8 volts and
De = +0.6 volts were adequate for all joints. The benefit of this approach can be obseived in Figure 25, wlieie
the conditions remain the same x for Figmr« 23. It should be noted the changes in value of e{k) are both smaller
in magnitu<le then those illustrated in Fignre 23 and ijHiost entirely negative in value during the period that the
second component of the DFLC is active.
These results demonstrate acceptable steady state response is Mliieved by the manipulator under test, in contrast
lo the steady state holding problems experienced by other researcheis. Sormally at the steady state a fazy logic
controller will allow the manipulated control variable to vary in the dead zone. The aim of the second component in
the DFLC was not entireley to negate the presence ot the dead zone but fa minuiiiBe any disadvantages with respect
to the performance of the system. The full control policy implemented in the second component of the DFLC,
including the technique to overcome chatter, is illustrated in Figure 24.
K g e p e 24 Flow diagram representation of the control policy of the second component in the DFLC
0.35 20
^
0.3 ^15
0.25 ^10
0.2
5
Q.15
0
0.1
•" /I .\ '• . l\ '', f " -5
O.CS
0
L_l^
0 50 1C»150 2 » 2 5 0 300350400450S<MS50
.10
SamplB int#real
— Joint angto/radians —Control wariaMs/voltss
«(m ,
L.»»-j =%f*)
STATE t W CONTROLIEF
INPUT
COirmOL ACTION
l i e that comprises ti»e DFLC and PID components because of the lesulte obtained earlier in tlie project. It is
interesting to note that during the development phase of the of the HC, both PD and PI algotithme were used as
the deterministic controller components and the results showed little difference in lesponse.
The FLG omtlined earlier does not require any modiication to be made and can be simply used directly in the
hybrid controller. Similiaily, the control policy executed by the DFLC can be left undianged. There is, however,
one major dificulty. In operation it was found that a HC containing the DFLC as the (mif controller component
will experience relatively large "kicks* in the value of c{je) as the system approaches the steady state value and this
442
UPDATE.DQNWtOC AC 1K>N
leads to oscillatory uiotion of umaeceptably large amplitude. This b e h a w u r is moat obvious in the waist joint.
The magmitiide and frequency at the waist joint ia illustrated by the dashed lime in Figure 28. This behaionr,
however, is BOt as pronounced in the remaining robot joints and can be attributed to the low reaction torgue of
the waist joint actuator as the Joint reaches the selected point. There are two possible solutions to this problem.
Joint angto/radians
The irst involves tcsttncturing the hybrid controller so botli control components are active together with adjustable
"weights" on each component output value. The value of the manipulated variable will then be obtained from the
sum of these components. Suppose GFazz(4>')} " d GDMTie(k)} are the chosen weights, as shown in Figure 29,
443
MEMBERSHIP QRAOE
ERROR DOMAIN/RADIANS
The second solution, which was adopted, is computationally less demanding and more elegant. In this latter
approach, the various components of the distributed fuzzy logic controller are switched off in a predefined sequence
as e{k) tends to 0. It was found by trial and observation that if the second component of the distributed fuzzy logic
controller was deactivated when the joint luigle B{k) was within the range of ±0.004 radians of the desired value,
the previously observed *'kicks^ did not occur. The result of using this approach is shown by the continous line in
Figure 29.
It was assumed that the robot joint could be modelled adequately using a second order equation, since similiar
models have been used successfully [14]. The assumption is reasonable in this particular case because the nnodel
only represents the behaviour of the joint in a small region around the steady state value. The model properties
were evaluated using a technique based upon the logorithmic decrement application to under-damped system [15],
and were estimated to be ( = 0.6 and w„ = 5.0. Using these values, an initial set of PID controller coflicients were
derived. By trial and observation these values were tuned to the robot, yielding:
The PID velocity controller was implemented, both for safety and for bumpless transfer. The problems encountered
in implementation were created by the digital feedback signal from the optical encoder and the limitations of the
Occam language which introduces rounding errors. Both these factors reduced the stability margin. It was further
noticed that the performance of the velocity algorithm resulted in different outputs depending on the number of
rounding or truncation operations that occur. The problems that arise in the digital implementation of algorithms
when dealing with rounding errors and variable sample rates is discussed by Moroney [16] and an explanation is
given for the relatively poor response of digital controllers. The expression used to implement the velocity PID
algorithm use in the HC is:
-'(*) = t. (e(t) - 2e(t - 1) + e(k - 2)] + K, [e(J:) - e(t - 1)] + Kie{k)t. (25)
The supervisory process of the hybrid controller consists of two routines and these are error driven. The first routine
switches between the DFLC and the PID components of the HC using the boundary values discussed earlier. The
second routine ensures both of the components ( DFLC and HC ) cannot be evaluated together.
444
jQ'mi angle/radians
— SFLC " - H 0
Joint angle/radians
— SFLC ^-HFLC
Figure 31 Comparision of the beliaviout of the shoulder joint in the presence of a disturbance
It should be noted that all the experimental results were obtained using the lame data bases were appK)priate,It
can be seen that the response of the joint is superior when the FLC algorithm is employed. Not only ate the
oBcillations experienced by the shoulder joint smaller, the steady state response is much reduced.
It can be seen from these results that the performance of the "pure" fuzzy controllers ia superior to the hybrid
controller. This result confirms that the fuzzy controllers can respond more positively to both digital input data
and externa] perturbations.
4 Project Software
The experimental investigation of the proposed controllers required the development of appropriate software. The
origianl design problem was decomposed into a number of simpler tasks before coding was undertakene. The design
goals were:
• Control the major robot joints using fuzzy logic.
• GBUS board to be operated by a distinct process.
• Simple interface between the controller software, user and host computer filer system.
• Modular structure, provision for the upgrading of the individual units comprising the software package.
• Joint controllers.
• Collision detection. Actuator overload detection for each joint.
An "engineering approach" was adopted to the design of the concurrent control software. A combination of techniques
were adopted, such as a top-down design methodology using MASCOT like primitives, flow diagrams and pseudo
code. The approach to testing is bottom-up, allowing for the resolution of communication deadlocks and other
common problems common to Occam software. A discussion of these techiques is presented by Henderson [18].
HOST
l TFIAMSPLrrER T800
TFWWSPUIEB
FROM CONTROL
INTERFACE
keeping" aclivities. For example, rearranging the data in the anay when new template parameters are added or
deleted. It ensures that the data is held contiguoysly, an airangtnent which will simplify the operatioE of the data
base and lience improve the run time efficiency of the program.
The template paramteis within the data base are stored in rank order. The array offset is equal to the ranking
of the fuzzy set. The deletion or addition of a fuzzy set can result in changes to the ranking of the other fuzisy sets.
To maintain contiguous use of the array this implies mo¥iiig template parameters. This action is carried out by the
management system which governs the area of memory holding the array.
% ^
'*i
FBOM
vt ^
FROM
CONTHOL CONTBOL
INTERFACE
TO OirTPUT PROCESS
5 Concliision
The results given in this paper show that fiijiy logic can be used in robotic control and that the steady slate problem
often encountered in the application of fuisaj- logic control can be overcome through the use of a hybrid controller.
A fuzzy logic controller caw be designed to operate in the real time enviromemt with high sampling rates without
modiication of the controller's structnre. It has been shown that parallel processing using a transputer network
caa be successfully applied to real time control in robotics. The design and development of small to medium siaed
programs (8000 lines), written in the language Occam, can readily be cli¥i8e€l for applicatioES requiring real time
operation.
References
[1] Zadeh, L. A.
Outline of a New Approach to the Analysis of Comp-lex Systems and Decision Processes.
IEEE. S M C , January 1973, 3(1), 28-34.
p ! Daley, S. and Gill, K. F.
Attitude control of a spacecraft udng an extended self-organising fuizy logiccontroller.
Proc. Int. Eng., 1987, 20(12), 97-106.
p ] INMOS htd.
The Transputer Developement System.
Prentice-IIall, 1988,
[17] Moroney, P.
Issues in the Implementation of Digital Feedback Compensation.
The MIT Press, 1983.
[18] Henderson, M. I.
Use of Parallel Processing in the Fuzzy Logic Control of a Robot.
Doctoral Thesis, May 1993.
The University of Leeds, England.
C H A P T E R 17
1 Introduction
The Robotic Institute of America (R.I.A.) gives the following definition of robots:
" A robot is a reprogrammable multifunctional manipulator designed to move ma-
terial, parts , tools or specialized devices through variable programmed motions for
the performance of a variety of tasks" [55]. Based on this definition it is apparent
that a robot must be able to operate automatically. This implies that in most of the
robots it is possible to distinguish the following major subsystems: a manipulator
(mechanical unit which can be compared to the skeleton of living beings), sensors
and actuators (sensory organs and muscles of living beings), a controller (the brain),
appropriate power supplies, and very often a computer system which takes care of
the monitoring and control functions relative to the robot operation and which al-
lows exchange of d a t a between the robot and human operators a n d / o r other parts
of the manufacturing process in which the robot is performing some specified tasks.
T h e motions of the manipulator must be controlled and the control system obeys
the same basic principles as for control of motions of any mechanical system from
simple servomecbanisms up to complex machines or vehicles. It implies that positions
and velocities or displacements of the various parts of the mechanical system must be
monitored and that related data must be transinitted to the control system. Then the
latter is able to determine the driving forces a n d / o r torques which must be applied
to the mechanical system in order to force the actual positions and displacements to
track the desired ones.
T h e next section of this chapter is dedicated to the dynamics and control of robotic
manipulators as they are relatively complex mechanical systems. Then the basic
principles of control of motions in mechanical systems are reviewed in the following
section. T h e two major classes of conventional position controllers are introduced in
a unifying presentation. This presentation aims at introducing fuzzy control in the
following section, fuzzy control appearing then as a natural extension of multilevel
discontinuous control. T h e advantages of simplicity and reliability of discontinuous
' T h i s paper present.s research results of the Belgian Programme on Interuniversity Poles of At-
traction initiated by the Belgian State, Prime Minister's Office, Science Policy Programming. The
scientific responsibility rests with its authors.
451
S.G. Tzafestas and A.N. Venetsanopoulos (eds.j
Fuzzy Reasoning in Information, Decision and Control Systems, 451-491.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
452
control are retained by fnzzy control, but not its main drawback which is contin-
uous cycling between different types of operation. Therefore fuzzy control can be
viewed as an intermediate class between discontinuous and linear control systems,
resulting in an acceptable compromise between advantages and drawbacks of both.
Positions and velocities or displacements are the usual state variables defining the
state of a mechanical system. Control laws based on measurements or estimates of
those variables allow some changes in the dynamics of the system, in particular the
stabilization of unstable or neutrally stable systems, and give them the capability
of reproducing desired motions with an accuracy which depends on the gains of the
control system. However in such control systems static errors due to steady-state
loading forces cannot be avoided. T h e only way to cope with such disturbances and
reject their effect on the system is the introduction of a reset action in the control
system. This can be achieved through parallel controllers using control laws basetl
on triples instead of pairs of data. However it may result in some deterioration of the
dynamic performances and lead to more difficulties in the design of fuzzy controllers.
Self-organizing controllers are a possible solution to the latter problem. Nevertheless
there is another way since an indirect reset action can be introduced via model-based
control schemes. These allow a neat separation of the two basic tasks of the control
system: following the desired trajectory (tracking) on one hand and reducing the
effect of disturbances (regulation or disturbance rejection) on the other hand. Such
control schemes consist of two control loops, one of them including a basic position-f-
velocity controller and the other one a model of the system. The basic controller
a n d / o r the model can be implemented as numerical or fuzzy systems as it is shown
in the fourth section of this chapter.
Position or displacement control is not the only type of control which may be re-
quired in the performance of robotic tasks. In some of them one has to control
contact forces between the robot-end-effector and the robot environment. Often the
complete control system of a robotic manipulator is hybrid in the sense that displace-
ments are controlled directly along some directions in the robot workspace and forces
along others. Then force control is generally implemented as an external control loop
around usual displacement control loops. Here again use of fuzzy logic is possible.
Other applications of fuzzy logic in robotics are also mentioned or presented briefly
in this chapter, and some paths for future developments of fuzzy control of roliotic
manipidators and more generally of mechanical systems are evoked in the concluding
section.
fine positioning and the desired orientation. Joints may be revolute joints producing
pure rotary or rotational motions or they may be prismatic joints allowing linear or
translational relative displacements. Each joint defines a robot axis around or along
which there is a relative rotation or translation of the subsequent linkage of the chain
with respect to the antecedent one. These axes are usually but improperly called
degrees of freedom (DOFs) by engineers. Most of the manipulators fall into one of
the following classes; Cartesian, cylindrical, spherical or jointed devices, according
to the coordinate system which is appropriate for describing the configuration of the
major axes. Fig. 1 shows a six-DOF jointed-arm manipulator. Very often indeed the
number of DOFs is five or six, three for positioning and two or three for orientation
depending on the tasks the robot is designed for. However the number of DOFs
may be as small as four or even two in a few robots; on the contrary it is larger in
others which have redundant axes giving the manipulator more flexibility in adapting
its own configuration to a particular task. Besides in some applications the base of
the robot is carried by a vehicle allowing large linear displacements in at least one
direction.
• M{q) and C[q,q) are the inertia matrix of the manipulator and the matrix
relating centripetal and Coriolis forces to the generalized velocities,
In all the previous applications it was assumed that neither the end-effector nor
any other moving part of the manipulator comes in contact with any object in the
robot environment. Then the implementation of the control system need only joint
position and velocity measurements and some knowledge of a dynamical model of
the manipulator. Other tasks such as assembly, deburring, polishing, cooperative
work with other robots, . . . require the control of the position and also of some forces
exerted by the end-effector upon an object. This can be achieved via two different
schemes: impedance control and hybrid control. Impedance control aims at adjusting
the mechanical impedance relating the end-effector displacement to the forces due to
contact with the environment of the robot. Impedance control may be passive or
active. In the first case a kind of elastic suspension including springs and dampers is
inserted in the connection between the robot-end-effector and the manipulator wrist.
In active impedance control the feedback controller utilizes measurements of posi-
tion and velocity of the end-effector with respect to the object and measurements of
the contact forces between them. A noteworthy difference between position and force
456
desired trajectories
driving torques positions and velocitie;
for
of of
joint 1
Controller joint 1 joint
for joint 1
joint 2
Desired C'ontroller joint 2 joint 2
joint trajectory for joint 2 Robot
generator
joi it n
Controller joint n joint n
for joint n
Sensors
for joint n
Sensors
for joint 2
Sensors
for joint 1
control is t h a t position control in general should be very accurate while force control
may be looser; usually all that is required is that the actual contact force or torque
lies within a given range. Then qualitative control and fuzzy concepts appear to be
very natural in that field. Hybrid control includes feedback loops providing separate
control of displacements along some directions and forces in others. It should be noted
that many of the previous control systems require measurements in the workspace and
actuation of the joints; this implies a lot of on-line calculations for transformations
from the workspace into the space of joint coordinates and vice-versa. Then some
implementations of the control system are two-stage structures including inner loops
for joint-position control and outer loops for force, impedance or hybrid control in the
robot workspace. For this reason joint position control is the main topics which will
be considered in the followings. Thanks to the simplifying assumptions explained
in the previous section it is generally sufficient to consider separately each joint,
by regarding changes in the manipulator dynamics, joint interactions, friction and
gravitational forces and possible loading forces, as external disturbances the position
controller of the considered joint has to cope with.
457
\, Interna 1
>' Disturbances
.
Tl
\ , <? r (1
) ' M J J
dq
Tt = '
M.- = r„.-r^ (•2)
where q and q are position and velocity, subscript d refering to the desired position
and velocity, Cp and e„ are respectively the position and velocity errors, A'p and A'„
the corresponding controller gains, T„^ and r; are the driving and loading torques
respectively and M denotes either the mass or the moment of inertia of the system.
T h e closed-loop system can be described by the following second-order differential
equation relating the position error e to the desired acceleration and the loading
torque:
M • E + K^ • e + Kp • e = M • qi + Ti (4)
458
External Disturbances
(Id +
Setting this equation in the standard form of a second-order linear differential equa-
tion:
e -I- 2Cw„ • e 4- uj^ ' '^'' ^ 'i'l + ^ "^i (5)
it is well known that the complementary solution [74] of this differential equation is
a linear combination of two independent functions;
It appears that u;,i determines the response time of the closed-loop system
K
and C = is related to the aperiodic or oscillatory feature of the transients
2Miu„
and to their decay ratio. As indicated by these relationships the characteristics of
the dynamics of the closed-loop system can be adjusted independently by a proper
choice of the values of the proportional and derivative gains Kp and / \ „ . It is also
obvious from Eq. 4 t h a t the steady-state error due to a static loading torque is in-
versely proportional to the servo stiffness A'p.
Then the accuracy of the position control system is directly related to the servo
stiffness A'p. ffowever it can be improved by adding in the control action an estimate
•f( of the loading torque a n d / o r a feedforward term M -qri corresponding to the desired
motion {M is an estimate of the inertia of the controlled system). On the contrary
the dynamic error could be higher if the derivative term in the control action contains
only a velocity feedback — A'„(j instead of the velocity error A'„ • e„.
459
Another way towards the rejection of disturbances, more precisely towards the reduc-
tion of errors due to slowly varying disturbances, is the introduction of an integral
action in the controller:
signjus} = -f 1 if .s > 0
signjs} = - 1 if s < 0 (8)
It can be seen on Fig. 5 and 6 that the phase plane (e, e) is divided by the switching
line s — (] in two parts inside each of which the driving torque is constant. Neglecting
the loading torque and other inhomogeneous terms such as the desired acceleration
the equation of the closed-loop system can be written down as follows:
• de de 1 . r i /„x
M
e+ • e^ • signj^s} = constant (10)
In other words, due to the fact that a constant driving torque is always applied either
in a positive or in a negative sense, the controlled system is bound to be driven back
and forth in a cyclic manner, the trajectory in the phase plane of the closed-loop
system being a sequence of alternating arcs of parabola.
460
0.5-
-0.5
-1.5
0.5 -
-0.5-
the trajectory corresponding to simple cases such as free response or step response is
composed of two arcs only, the first one starting from the initial point in the phase
plane and the second one passing through the origin of the phase plane. This would
be the minimum time response for any driving torque bounded to take its values
in the interval {—TM,+TM), as shown in [74]. Then the switching line in the phase
plane must be made of the two arcs of parabola passing through the origin of the
phase plane (Fig. 5), and the switching function can be represented by any of the two
following expressions:
Remark that this control mode could also be presented as a limit case of a P-U con-
trol system: r„i = A'„ • (a; • e -|- e), where the gain A',, is increased to infinity but
the controller output is bound to stay in the interval {—TM^+Tm). When Ky goes to
infinity the transients of the closed-loop linear system are combinations of two ex-
ponential modes: one of them vanishes infinitely fast and the other one corresponds
to a time-constant approaching oj~^. In practice however due to the bounds on the
control variable the first part of the transient is a parabolic trajectory in the phase-
plane. This is what really occurs in discontinuous control: at first the system is forced
towards the sliding mode motion and the latter corresponds to an exponential decay
characterized by a time-constant w " ' .
Anyway the motion will ultiinately reach the origin of the phase plane, but then,
if the controller is not switched off, in other words if the driving torque can only take
the two values —TM and +TM, there will be a limit cycle, that is an oscillation around
the origin of the phase-plane. Theoretically both this limit-cycle and the possible
chattering motion are oscillations with infinitely high frequency and infinitely small
amplitudes. In practice, due to unmodeled dynamics such as small time-constants,
dead-times, .. . these oscillations will have finite amplitudes and frequencies. In par-
ticular, in case of sampled-data systems the minimum period of the limit-cycle is twice
the sampling interval. Such oscillations may be acceptable in other applications of
ou-off controllers such as residential beating-cooling systems. In position control of
mechanical systems they are generally not acceptable because of problems such as
p r e m a t u r e wear of some parts of the system, excitation of undesirable vibrations, . . .
A remedy for such a behaviour could be the use of a three-position controller with
462
between the two above "on" positions, a neutral or dead zone where the driving
torque is set to zero:
This may eHminate the limit-cycle, in general it does not avoid sliding motions the
sliding line s = 0 being now replaced by two parallel lines s = +t and s = —f. In
order to eliminate chattering superposed to these sliding motions it is necessary to
smooth the transitions from 0 to +rM or —TM'-, this can be achieved by replacing the
sign-function in the control law by a sat-function:
Tm = TM • sat
0 ;i4)
where:
= -1 if .s < —f
.s .s
iat{-} = - if \s\ < e :i5)
( f
= +1 if s > e
but this is nothing else than coming back to a linear control with bounds on the
magnitude of the controller output.
In the same way the values of the controller output can be indicated by qualitative
labels and assignment of these values can be expressed by literal "If... t h e n . . . e l s e . . . "
statements, such as:
Then the control law is expressed in a pure verbal formulation consisting of a set of
conditional statements. This set is nothing else than a rule base. In other words the
previous discontinuous or logical controller can be viewed as an elementary qualita-
tive controller and it is a first step towards a pure qualitative control system with
arbitrary numbers of classes in the universes of discourse of the input and output
variables and a proper number of rules in the controller rule base. Pure qualitative
control has a major drawback: due to the abrupt transitions between adjacent classes
of the controller inputs there are discontinuities in the controller output whatever the
number of classes in the universe of discourse may be; this may induce sliding mo-
tions and chattering of the controlled system. For instance, if .s is still the single
463
input variable of the controller and there are A' classes in its universe of discourse
the only change with respect to the two- or three-position controller considered above
would be that there are now N — 1 parallel shding lines in the phase plane (e, e). In
order to eliminate chattering it is therefore necessary to smooth transitions between
adjacent classes in the universes of discourse of the input variables. In quantitative
control systems this can be achieved by relating the controller output to the inputs by
suitable continuous mathematical relationships, which can be linear or not. Such a
smoothing can be obtained in a way which may seem more appealing and closer to the
human way of thinking by introducing fuzzy logic in the interfaces between the rule-
based logical controller and the real world. Then it could be said that fuzzy control
is an intermediate stage between simple logical control and more complex continuous
control, trying to achieve a trade-ofF between the advantages and drawbacks of these
two kinds of control systems: simplicity, robustness, ease of maintenance and low
cost are nice features of logical control, but chattering is undesirable in many tech-
nical or industrial applications; at the opposite continuous control is more complex,
more sensitive to disturbances and to perturbations in the controlled system, more
expensive and requiring more skilled technicians for operation and maintenance, but
it results in a smooth dynamical behavior of the plant.
• fuzzy control requires only a qualitative understanding of the static and dy-
namic characteristics of the plant to be controlled, using imprecise, vague and
possibly ambiguous linguistic representations of the real world; this allows the
implementation of a control system without the need of an explicit m a t h e m a t -
ical model of the plant;
• fuzziness in the control system gives it some capability of coping with external
disturbances a n d / o r internal perturbations, as well as with uncertainties or
variations over time of the plant parameters.
with applications of fuzzy logic in robotics. Those papers can be classed in several
categories:
• control of motions of unmanned vehicles and mobile robots, see e.g. [4, 7, 16,
26, 35, 40, 42, 44, 48, 57, 9 1 , 90, 95];
• use of fuzzy logic in vision systems and other sensory devices for robotic appli-
cations; see e.g. [17, 39, 62].
In the foUowings only the application of fuzzy logic to the control of servo systems
and robotic manipulators will be considered, with the emphasis on position control
of a one-DOF servo system. As said in previous sections, the control system of a
multi-link manipulator is a collection of joint controllers each of which is dedicated
to a single joint and drives it individually.
controlled process is neutrally stable and the other one providing rejection of distur-
bances and reset of the controlled variable in steady-state. Then, with a P-controlJer
or a P I incremental controller if the controlled plant is neutrally stable, with any
of the three controllers if it is stable, the closed-loop system is stable as long as the
control action is not too strong, in other words as long as the gain of the equivalent
"linear" controller is not too high.
The situation is quite different in the position control of most mechanical systems.
If there is no viscous friction providing some damping in the controlled system the
model of a one-DOF mechanical system is usually a double integrator. Then, as it
can he seen by relationships (4) and (5), in order to stabilize the closed-loop and give
it appropriate damping, the control law must include a term related to the velocity
error or to the actual velocity, both in linear and in discontinuous control systems.
This is the same for a fuzzy control system. The simplest fuzzy controller for po-
sition control of a mechanical system has two inputs, the first one being naturally
the position error and the second one being a velocity signal in order to stabilize
the closed-loop: either the actual velocity of the mechanical system or the velocity
error or the change in the controlled position or the change in the position error. If
the torque or force driving the mechanical system is proportional to the output of
the controller the latter can be compared either to a linear P-D controller or to a
P-D discontinuous controller. Then assuming no spring force and no external loading
force such a control system has no steady-state position error in its normal operating
range. However such a control system is unable to cope with disturbances, so that
there will a position error due to static loading. The only way to achieve disturbance
rejection is the inclusion of an integral action in the control system. This can be
obtained via a three-input controller which can be compared to a linear P-I-D con-
troller. Now three terms are necessary in the control system: the first of them is the
natural control action, the second one is required for stabilizing the control loop and
the third one aims at rejecting disturbances. However, when it is relatively simple
to design a two-input rule base or the equivalent decision table, by establishing some
correspondence between the decision table and the phase plane of the control sys-
tem, this becoines more difficult with a three-input rule base. There are some recent
results concerning that approach, they are reported in a following section. Another
way is the use of a model-based control structure which allows a neat separation of
the tracking and disturbance rejection functions of the control system [32, 28].
External Disturbances
Rule
base
Kp K,
(Id
<>
onto the universe(s) of discourse of the fuzzy variable(s), a sampler for sampling
data, and possibly also analog to digital converter(s). T h e sequence of these opera-
tions depends on the actual implementation of the interface. On the other side the
defuzzyfier contains a sample-and-hold device, in general together with a digital to
analog converter, and a scaler or amplifier for mapping the universe of discourse of
th<^ fuzzy controller output into the range of the real control variable. In position
control one of the input variables is the position error, the other one is a velocity
signal which can be the actual velocity of the mechanical system or the velocity error
or the change in the controlled position or the change in the position error. In the
two latter cases the position or error change must be computed in the input inter-
face of the control system. In the two first cases velocity measurement is required
in addition to position measurement; moreover the use of the velocity error requires
the knowledge of the desired velocity, which is the case for robotic manipulators and
some other servo systems where the desired trajectory is programmed in advance and
ilesired positions and velocities can be stored in the computer memory.
In some papers on fuzzy control of servo systems the universe of discourse of each
fuzzy variable may be divided in up to thirteen different classes, in most works this
number is limited to seven classes. Sometimes those classes are defined by numbers,
but usually the following linguistic labels are assigned to them:
467
For simplicity most authors use for the various fuzzy sets membership functions
having trapezoidal or very often triangular shapes. However it should be noted that
in some works on autonomous drive a n d / o r parking of vehicles the number of fuzzy
sets may be different and the shape of the membership functions may be tailored to
the type of variable, being different for acceleration, speed and position [50]. Often it
is desired to have a fine control for small position errors while control can be coarse
for big errors. Some authors deliberately use two different decision tables for that
purpose [63]. In many cases it is sufficient to introduce a nonlinear mapping of the
range of the real variable onto the universe of discourse of the corresponding fuzzy
variable. This can be achieved by means of a piecewise linear function or a sigmoidal
function like atan (.). Another way to obtain the same effect is the use of asymmetric
triangular membership functions like in Fig. 8 [29].
NB NM NSAZPS PM PB
Universe of discourse
In a two-input fuzzy control system the control strategy consists of a set of rules or
qualitative statements which can be expressed by linguistic or verbal expressions such
as:
If there are rn fuzzy sets for the first input and n for the second one the number of
such rules may go up to rn * n. For a two-term position controller with seven sets
for position and as many for velocity this gives 49 rules. T h e most concise way of
representing that rule base is a decision table like Table 1, where there are two entries
corresponding to the two input variables of the controller.If one thinks of that table
as a qualitative partition of the (e, e) phase-plane of the control loop one sees clearly
t h a t there are five "parallel" switching lines in that plane. This analogy between
fuzzy logic and multilevel relay control had been pointed by [92]. This set of rules
468
may not give the closed-loop system the best transient performances. The responses
in the neighbourhood of the setpoint can be improved by changing some control rules
or as in [63] by the use of two lookup tables, one for coarse control the other one for
fine control, the controller switching from one lookup table to the other depending
whether the error is within given limits or not.
PB AZ PS PM PB PB PB PB
PM NS AZ PS PM PB PB PB
PS NM NS AZ PS PM PB PB
Velocity AZ NB NM NS AZ PS PM PB
Error NS NB NB NM NS AZ PS PM
NM NB NB NB NM NS AZ PS
MB NB NB NB NB NM NS AZ
NB NM NS AZ PS PM PB
Position Error
It is well known that, in fuzzy control on the contrary of pure qualitative control,
several control rules can be active at the same time with various degrees of fulfilment
and contribute to the final control action to the extent of the latter. So all the active
rules must be combined in an appropriate way to produce a resulting fuzzy control
action. T h e degree of fulfilment of a given rule is obtained as the minimum or the
product of the membership grades of the values of the two input variables which
are able to activate the rule. Then the membership function of each rule control
action is either clipped to a maximum value equal to its degree of fulfilment or it
is simply multiplied by the latter. T h e resulting rule contributions are combined
in a final fuzzy control action through the union or the sum of their membership
functions, and the control action is defuzzyfied in the defuzzyfication part of the
control system. Several defuzzyfication methods can be used, such as the ''nieau-
of-maxima" method, the "centre-of-gravity" method and the "fuzzy-mean" method.
Analysis and experiments [45, 61] show that the first method reduces the "fuzziness"
of the control system and may lead to steady-state errors a n d / o r oscillations (limit-
cycles). So this method is not suitable for fuzzy control. The second method is that
which preserves the best the fuzziness of the control system, but its usual continuous
form is not suitable in terms of computations for practical applications. The "fuzzy-
mean" method can be viewed as a discrete version of the "centre-of-gravity" method,
using discrete representations of the fuzzy sets at the output of the controller and
rej)lacing integrals by sums for determining the centre of gravity [45, 80]. Then the
defuzzyfied value of the control variable is given by:
u— (16)
rules (that have a non null degree of fulfilment), A' being now the number of active
rules and 7^,. the degree of fulfilment of an individual rule.
The role of defuzzification methods in fuzzy control is analyzed in [45]. For simplic-
ity and speed of calculations in position control one often chooses the "fuzzy-mean"
method with crisp o u t p u t fuzzy sets, summing for all the active rule their contribu-
tions weighted by their degrees of fulfilment and dividing the weighted sum by the
sum of the degrees of fulfiment. W h a t may be more important for the control system
performances is the distribution of the fuzzy values in the universes of discourses of
the inputs and the output of the fuzzy controller. Fig. 9 and 10 show the transient
responses of the control loop for various distributions of the universes of discourses,
which may be different for the inputs and the output of the controller. Fig. 8 is an
example of classes that are inequally distributed over the universe of discourse. It
can be seen that distributions corresponding to equivalent controller gains which are
higher for small errors than for bigger errors result in better performances. Then
fuzzy logic appears as a clever way of introducing nonlinearities in the control sys-
tem in view of better performances. At the opposite if the choice of the universes of
discourses for the inputs and the o u t p u t of the controller leads to a fuzzy controller
which is equivalent to a linear controller the performances of both are similar as it
can be seen from Fig. 9 and 10 . Other ways of introducing nonlinearities and so
some intelligence in the fuzzy control system were proposed bv several authors, see
e.g. [:?6].
^
Fuzzy PI Fuzzy I
Figure 11: Fuzzy PID controller structures (fuzzyfier and defuzzyfier are included in
the two blocks]
/ / error is E and change, of error is C then assign value P to the performance index .
A possible modification of control rules can be based on the assumption that the
current value of the performance index is mainly due to the controller o u t p u t n sam-
pling intervals in the past. If the performance index has value zero no modification
is needed. On the contrary a non null value of the performance index is a measure
of the correction which is required at the system output. Then the linguistic output
value of a control rule which was active n sampling intervals before is changed, the
new value being equal to tlie previous one plus the current value of the control index.
Such a self-organisation of the control rule base was used in two- and three-term fuzzy
controllers for position control of three-link manipulators [88, 93]. Moreover in [88] a
third layer was introduced in view of automatic self-tuning of the scaling coefficients
related to the three inputs of the fuzzy controller (Fig. 13). Self-tuning is based on
a small number of tuning rules relating increments of the input gains to variations
of some performance ratios (integral squared error, rise-time, % of overshoot). The
gains are updated through a cyclic procedure allowing successive refinements of each
performance ratio. However an initial guess motivating a stable control loop is nec-
essary to start the tuning procedure. The tuning rules may be defined by analogy
with Ziegler-Nichols rules for tuning P-l-D parameters [74] taking into account the
peculiarities of the present control structure.
472
Rule
Modifier
Rule
base
Fuzzy
('ontroUer
Inference
Scaler RFI engine FRI Scaler —i
- R - Q - —^ Estimate
of PI
Learning
Module
Rule
Modifier
Estimate of
Performances
Inference
S Cider RFI engine FRI • Scaler
Self-organizing controllers (SOC) give very good performances for a given task but they
may be not so good for other tasks which would very different from the task performed in
the learning phase. Moreover it is obvious that learning requires a lot of successive ex])eri-
ments and involves many calculations. Then the use of such controllers is the most efficient
when the manipulator has to repeat many times the same task, the latter l)eing known a
priori or not.
x^fiq) (17)
where Jiq) is the .Jacobian matrix ,/•' = —-^ . Then, a princijile of virtual works yields
dqj
the following equalities:
T'^ -Sq = F'^ -fix ^ F'^ •J{q)-6q (19)
which imply:
T = r{q) • F (20)
where r and F are the vector of torques and forces respectively on the joints and in the
workspace of the manipulator. Therefore the changes in the setpoints of the joint position
control loops are:
tq^ = J-\q) • (/ - S) • 8x^ -H A'-' • Tj (21)
where K = J^{q) • lix " •^(9)1 A'^ is a stiffness matrix in the workspace and TJ is the vector
containing the outputs of the force controllers. As it can be seen on Fig. 14 there are
474
several nonlinearities and time-varying gains in the force control loops, moreover there are
uncertainties in the stiffness matrix related to the contact forces between the robot and its
environment. Then in [38] authors use P-I controllers in the force control lines and they
introduce a fuzzy tuning of the gains of these P-I force controllers on the basis of the errors
and the error changes in the force control loops. This control scheme vifas tested through
simulation on a two-link manipulator in [38].
forces Robot
environment
contact forces
])ositions
setpoints Force -|- Position in
Joint Robot
in Control System ControDers workspace
workspace
Adaptation
Mechanism
desired T/
PI
forces Controllers A'-
setpoints for
desired I - S joint position
positions J-
controllers
to vibrations of the links. Then it is necessary to take into account flexibility of the links
and compliance in the joints as well as the interactions between different axes in the design
of the robot control system. Many papers on the control of manipulators with flexible
arms and the suppression of vibratory motions were pubUshed in the last years; see e.g.
[5, 15, 22, 66, 100, 104]. Again it is only recently that the use of fuzzy controllers has
been considered for that purpose. A study of fuzzy control of a flexible arm is reported in
[!)'i]. Authors use a two-term fuzzy controller for position control like in Fig. 15, but they
introduce a second two-term fuzzy controller for suppression of structural vibrations on the
basis of data supplied by a strain gauge fixed on the arm. Then the outputs of the position
controller and the vibration controller are combined in a fuzzy synthesizer which produces
a single control variable according to the following decision table (Table 2). Experimen-
tal results reported in [53] show that if the strain gauge location is carefully selected this
control scheme is very effective for suppressing the effect of the first vibration modes the
higher modes vanishing quickly thanks to the internal damping in the arm. It seems that
this fuzzy control system may be more robust than conventional optimal control when there
are some parameter changes in the controlled system.
Strain + ChaLUge
'
Vibration
Suppressor
Fuzzy •
Robot
Synthesizer
Position
Controller
PB AZ PS PM PB PB PB PB
PM NS AZ PS PM PB PB PB
Vibration PS NM NS AZ PS PM PB PB
Controller AZ NB NM NS AZ PS PM PB
Output NS NB NB NM NS AZ PS PM
NM NB NB NB NM NS AZ PS
NB NB NB NB NB NM NS AZ
NB N M N S AZ P S P M PB
Position Controller Output
476
Various analog or digital control systems based on a model of the controlled plant have
been proposed for process control applications since the fifties, see e.g.[3, 24]. Roughly
speaking a model-based control scheme is composed by three basic parts: the system that
unist be controlled, a controller or compensator driving this system and a model of the con-
trolled system, the difference between the outputs of the controlled system and its model
being used for correcting either the output or the setpoint of the previous controller or com-
pensator. Obviously the plant to be controlled cannot be altered, but the control systein
itself can be set up independently of the particular implementations of the controller and
the model of the process. Therefore as pointed in [29] either the controller or the model
or both of them can be implemented in the process computer using fuzzy and qualitative
control concepts. Hybrid implementation including fuzzy or qualitative implementation of
some parts of the control system and numerical or quantitative implementation of others is
also possible. Besides different sampling rates may be used in different parts of the control
system, for instance in the basic controller and in the model, which results in a multi-rate
control system.
The analogy of such a control scheme with the human behaviour is straightforward. It
can be reasonably thought that the control strategy of a human operator is based on a
qualitative understanding of the process that he has to control: in other words the oi)erator
would have in his mind a qualitative model of the process response, the latter being either
a static model (gain ) or even a dynamic model ( the process i.<; stable or not, it.s traii.sient
response is oscillatory or damped, the operator has a rough estimate of the ])rocess set-
tling time, . . . ) . Then a human operator would correct the manipulated control variable
according to the deviation of the actual process output from the output which had been
forecast by this internal model. Furthermore the basic control action taken by the operator
and the correction of this action can be performed at different time scales. Therefore the
477
There may be different model based control schemes, depending on the way the control
system uses the "return-difference" signal (difference between output signals of the actual
controlled process and its model; note that the output of the actual process is not neces-
sarily the controlled output variable, it can be an auxiliary output variable, but obviously
the latter must be measurable). First, this signal can be used for correcting the setpoint
of the basic controller like in Fig. 16, the control input being the same for the controlled
process and its model. This control scheme is known as "Internal Model (Control" (I.M.('.)
and is probably the oldest and the best known model-based control system; see e.g. [71].
In two other schemes, the return signal is used for correcting only either the control input
of the model or that of the controlled process. In the first of these two schemes the return
signal is used for forcing the state of the model to track that of the actual system, in fact
it is nothing else than the well known Luenberger observer [67]. The second of these two
schemes (Fig. 17) is the dual of the first one and as such, it has the same nice properties as
shown in [32, 60]. It can be said that in this control scheme the return signal aims to force
the output of the actual system to track that of the model and the effect of unmodelled
disturbances acting upon the system is counteracted by the effect of the return signal. The
latter scheme is nothing else than "Model Reference Control" or "Model Following Control"
[60].
disturbance
1
process on tput
Process
reference
tn cs)
V' .
Model
model out put
Correction return
Mechanism signal
As shown in [32] all these control schemes have the same basic properties: first, the nominal
tracking dynamics is the same whether the return signal is used or not in the control system;
second, the regulation dynamics can be tailored to the designer's objectives, thanks to a
suitable filter processing the return signal; third, the three schemes are equally robust.
The first pro])erty is obvious, as in case of exact model match and no disturbance the
return signal is null and therefore it cannot change the tracking dynamics of the controlled
system. As a consequence of that property control schemes of Fig. 16 and 17 can be used
as such if and only if the system to be controlled is open-loop stable. If not there must bo
another feedback loop for stabilizing the system and giving it the desired dynamics. This
loop may be either a secondary loop inside the model-based control scheme or an external
loop around it. In the first case the model must be that of the resulting stable closed-loop
478
system. On the contrary, in the second case, the model is that of the system which must
be controlled and it allows a correction of the control input to this system. Under some
general conditions this correction can be seen as an estimate of an equivalent disturbance
signal acting upon the system input. Then the latter model-based control scheme is also
called a "disturbance observer". Mechanical systems are generally unstable or neutrally
stable. Then several model-based control schemes are allowable for position control. The
case of a one-DOF servo system is considered in the followings.
disturbance
process output
Process
Model
model output
disturbance
Model
model output
Figure 18: Control with disturbance observer inside the main control loop
0.5
0.45 -
0.4 - fl D = 0.5
p
O 0-351
s
i 0.3-- D = 0.25
ill
t
n
E 0-2 ^
X^ \j\
r D = 0.125 \ 1 i1 \
r
o 0.15 -•
r
0.1 -
0.05 (;/
ij
'r '•''• '1
|/A /i^l'"'--,—. 1 . ""••--
10 15 20 25 30 35 40 45
Time [s]
Figure 19: Responses to constant disturbances: D = 0.125, 0.25, 0.5
Full lines: fuzzy PI) control
Dashed lines: fuzzy PD control + model-based control
480
disturbance
o Stabilizing
CJontroUer
Process
process output
6=
Model
model output
Correction return
Mechanism signal
disturbance
model output
which, at each sampling time, predicts the next value that the actual system output should
reach given the current data, and the difference between predicted and measured values is
used for correcting the control variable at the next sampling time. Due to the robustness
of model-based control schemes modeling must not be accurate: an approximate model
representing roughly the static gain and the settling-time of the controlled system should
be satisfactory in most cases. If the sampling interval for the model-based control scheme
is at least equal to the settling-time of the closed-loop controlled system the simplest model
is even a one-step ahead shift operator, (it can be seen as a "wait and see" strategy leading
to small updates of the control variable). This model can be embedded in any of the two
control schemes of Fig. 16 and 17. It was shown in [11, 29] that the disturbance rejection
capabibties are the same in the two cases (Fig. 22), but obviously the errors when tracking
time-varying signals are different; in particular for a ramp-wise varying desired position
the internal model control scheme leads to a constant error which is nothing else than the
so-called "velocity error" defined in textbooks on control systems (see e.g. [74]); this error
could decreased or even cancelled at the price of an increasing complexity of the model.
Another model which can be used in the control schemes of Fig. 16 and 17 is a first-order
predictor with a sampling interval approximately equal to half the rise-time of the closed-
loop controlled system [11, 29]. Obviously decreasing the sampling interval allows faster
reactions of the model-based control system as it can be seen on Fig. 22 (curve in dotted
Une).
0.5
0.4
D = 0.25
D = 0.125
-0.2
-0.3
0 10 20 30 40 50 60 70
Time [s]
Figure 22: Responses to constant disturbances: D — 0.125, 0.25, 0.5
Fxdl lines: model following control
Dashed lines: internal model control with one-step ahead shift operator
Dotted lines: internal model control with first-order predictor
482
PB AZ PS PM PB PB PB PB
PM NS AZ FS PM PB PB PB
PS NM NS AZ PS PM PB PB
Input 2 AZ NB NM NS AZ PS PM PB
NS NB NB NM NS AZ PS PM
NM NB NB NB NM NS AZ PS
NB NB NB NB NB NM NS AZ
NB NM NS AZ Ps PM PB
Input 1
Vu = y + M (22)
where u and y are respectively the input and the output variables and subscript u means an
u])dated value. (Uearly the accumulator is nothing else than an adder with a regenerative
feedback loop as shown in Fig. 23, and its qualitative implementation can be represented
by the same rule-base as in Table 3 just replacing one of the inputs by the last output
(Table 4).
1
^ Backward
Time shift ^
F A^ ¥•
PB AZ PS PM PB PB PB PB
PM NS AZ PS PM PB PB PB
PS NM NS AZ PS PM PB PB
Last Qniput Z NB NM NS AZ PS PM PB
NS NB NB NM NS A'Z PS PM
NM NB NB NB NM NS AZ Pb
NB NB NB NB NB NM NS AZ
NB NM NS AZ PS PM PB
Input
483
However pure fuzzy implementation of the neutrally stable closed-loop subsystem shown in
Fig. 23 would lead to an explosion of fuzziness, i.e. the support set of the output variable
would increase at each sampling time. Explosion of fuzziness is due to the type of fuzzy
rules "If X is A then Y is fl" used in the implementation, namely possibility riiles. Seman-
tic behind possibility rules is : the more X is A , the more possible Y is B. If several rules
are triggered, i.e. several control actions are possible, the union of these actions is possible.
More precisely in the case of the fuzzy accumulator, if the input and the last output of the
accumulator have respectively two and q simultaneous values the use of a decision table like
Table 4 may lead to ^ -|- 1 simultaneous values of the updated output. Then the number of
possible simultaneous values of the output could increase at each sampling time! Obviously
this number has an upper bound which is the number of classes in the universe of discourse
of the output variable, but it means that ultimately the accumulator output would have
simultaneously all the allowable qualitative values! Therefore the information would be
scattered over all the universe of discourse, which means a complete loss of information.
This is what has been called "explosion of fuzziness" [9]. Such situations could be avoided
by using another type of rule like certainty or gradual rules [21]. Implementation of the
operations related to these types of rules being harder, it was chosen to roughly limit the
fuzziness of the output: in the feedback loop of the accumulator the qualitative variable is
defuzzyfied at first, refuzzyfied immediately after and then fed back to the adder (Fig. 24).
Then the quaUtative values of the accumulator output which are fed back to the adder keep
some membership grades, and this prevents the explosion of fuzziness.
Vu
PB PS PM PM PM PM PM PB
PM AZ PS PS PS PS PM PB
PS NS AZ AZ AZ PS PM PM
Last Output AZ NS NS NS AZ PS PS PS
NS NM NM NS AZ AZ AZ PS
NM NB NM NS NS NS NS AZ
NB NB NM NM NM NM NM NS
NB NM NS AZ PS PM PtJ
Input
which results in coniplenientary attenuations of the input and feedback variables. If A has
a value around 1/2, in other words for values of the predictor time-constant approximately
equal to the sampling interval, the rule-base corresponding to qualitative implementation of
the predictor can be represented by Table 5. Here the attenuation of the feedback variable
lowers the regenerative effect and therefore the risk of explosion of fuzziness.
6 Conclusions
The numerous works which have been quoted in this chapter prove the effectiveness of fuzzy
control for position and/or velocity control of some mechanical systems such as servomech-
anisms, manipulators, robots, mobile robots, unmanned vehicles, . . . Most often the fuzzy
controllers used in those applications are position -|- velocity or P-D control systems and
their rule bases result from a straightforward qualitative interpretation of multilevel relay
control rules. Then the question is: as P-D hnear controllers are so well known and so
simi)le is it worthwhile to introduce fuzzy controllers in such appUcationsV There is no
unique answer to that question. For one-DOF servo systems with known parameters linear
controllers and fuzzy controllers with uniform distributions of the values in the universes
of discourses perform equally well. However even in that case modifications of those dis-
tributions allow the introduction of nonlinearities in the control system, which may lead
to better performances. Moreover if there are uncertainties or variations over time of some
parameters of the controlled servo system or if there are some nonUnear feedback terms such
as friction or gravitational forces a fuzzy control loop may perform better than a numerical
control loop, due to the fact that fuzziness implies by itself some robustness against the
influence of such pertnrbations. This is still more true for perturbations due to unmodeled
dynamics such as compUance in a mechanical transmission or flexibility of an arm. In the
latter case the work reported in [53] seems to be very promising as the control is based on
two separate two-term fuzzy controllers, one for usual position-|-velocity control and the
other one for vibration suppression. This idea of separating different tasks of the control
system is basically the same as that which can be found in the works on separation of
tracking dynamics and disturbance rejection in [9, 10, 11, 33, 27, 29, 28]. After all those
ideas are nothing else than apphcations of the ancient separation principle: ''Divide ut ini-
peraa" which should be the first of the basic C'omiuandments for any control engineer who is
facing tlie problem of designing a control system for a complex plant. This is the same idea
which is the basis of most control systems for multi-DOF manipulators where each joint is
actuated by its own controller the interactions between the motions of different jiarts being
considered as external disturbances. In that case also fuzziness in the control system may
imi)rove the overall performance.
Eventually there is another case where fuzzy controllers are more attractive than conven-
tional numerical controllers, it is the case where the control objectives and/or the control
strategy are only defined in loose terms; examples are not uncommon: driving and/or
parking a vehicle, path planning and control of unmanned vehicles or mobile robots, force
control, collision avoidance and/or cooperative tasks for robotic manipulators. Sucti a])])li-
cations were considered in several papers, see e.g. [47, 78, 102] and it is likely that further
developments of fuzzy controllers are to be expected in those areas. In most of the works
presented in the previous paragraph fuzzy control was applied at the lower level of acttia-
tion of a mechanical system that is position and/or velocity control. On the contrary, in
many of the a]>plications mentioned here fuzzy control is often applied at a higher level in
a liierarchical control system.
In some apphcations also fuzzy concepts are used for scheduling or tuning the parame-
ters of numerical controllers or for combining the outputs of several parallel controllers.
485
For instance in [89] a robot with a compliant joint is controlled by four controllers, each of
them having been calculated for a given area of the robot workspace, and the final control
variable is a fuzzy convex combination of the outputs of those four controllers. In [38] the
setpoints of some of the conventional position controllers of a manipulator are supphed by
P I force controllers the gains of which are tuned by a kind of fuzzy adaptation mechanism.
Sometimes also the controUer includes a numerical controller and a fuzzy controller acting
in parallel. This basic idea of such a hybrid control scheme is that numerical control and/or
compensation should be used for the known part of the controlled system and fuzzy control
introduced in order to cope with the additional unknown or uncertain parts. Such a hybrid
controller was used in [49] for position control of a servo system. In some sense all these
schemes can be related to the fuzzy supervision of control systems, see e.g. [75]; this is also
an area of active research.
References
[1] R.A. Aliev, G.M. Oulanov, and A.E. Tserkovnyi. Decision making in intelligent
robots. Soviet Physics Doklady, 31(10):780-782, 1986.
[2] C. An and J. Hollerbach. Dynamic stability issues in force control of manipulators.
In IEEE Conference on Robotics and Automation, pages 1890-896, 1987.
[3] C.B. Brosilow and B. Joseph. Inferential control of processes. A. I. Ch. E. J., 24:48,'3
500, 1978.
[4] M. Brown, R. Fraser, C.J. Harris, and C.G. Moore. Intelligent self-organizing control
for autonomous guided vehicle: comparative aspects of fuzzy logic and neural nets.
In '91 FEE International Conference on Control, pages 134-139, 1991.
[5] R.H. Cannon and E. Schmitz. Initial experiments on the end-point control of a flexible
one-link robot. The International Journal of Robotics Research, 3(3):62-75, 1984.
[6] R. Cassinis, E. Biorh, A. MeregaUi, and F. Scalise. Behavioral model architecture: a
new way of doing real-time planning in intelligent robots. International Society for
Optical Engineering, 852:275-280, 1988.
[7] Hai Quan Dai, G.R. Dalton, and J. Tulenko. Fuzzy control system for a mobile robot.
Trans, of American Nuclear Society, 65:464-465, 1992.
[8] M. De Neyer and R. Gorez. Fuzzy and quantitative model-based control systems for
robotic manipulators. International Journal of Systems Science, to appear, 93.
[9] M. De Neyer, R. Gorez, and J. Barreto. Disturbance rejection based on fuzzy models.
In M.G. Sing and L. Trave-Massuyes, editors. Qualitative Reasoning and Decision
Support Systems, pages 215-220. North-Holland, Amsterdam, The Netherlands, 1991.
[10] M. De Neyer, R. Gorez, and J. Barreto. Fuzzy controllers using internal models. In
IMA CS - IFA C Symposium on Modelling and Control of Technological Systems, pages
726-731, LiUe, France, 1991.
[11] M. De Neyer, R. Gorez, and J. Barreto. Simulation analysis of control systems with
internal fuzzy models. In ESS '91: Intelligent process control design and scheduling.
pages 145-150, Gent, Belgium, 1991.
[12] J. De Schutter. A study of active compliant motion control methods for rigid manip-
ulators based on a generic scheme. In IEEE Conference on Robotics and Automation,
pages 1060-1065, 1987.
486
[13] C'.W. De Silva and A.G.J. McFarlaiie. Knowledge-based control approach for robotic
manipulators. International Journal of Control, 50(l):249-273, 1989.
[14] S. Deepak, M.C. Repko, N.C. Mood, and R.B, KeUey. A fuzzy approach to the
interpretation of robot assembly forces. International Society for Optical Engineering.
1002:426-433, 1989.
[15] K. Desoyer, P. Kopacek, P. Lugner, and I. Troch. Flexible robots: modeling and
control. In S.G. Tzafestas, editor, Intelligent Robotic Systems, pages 421-453. M.
Dekker, New York, 1991.
[16] R. Deyong, J. Poison, R. Moore, C. Weng, and J. Lara. Fuzzy and adaptive control
simulation for a walking machine. IEEE Coiitrol Systems Magazine, r2(3):43-50.
1992.
[17] D.R. Dodds. Fractals, fuzzy sets and image representation. International Society for
Optical Engineering, 1001:87-94, 1988.
[18] D.R. Dodds. Fuzziness in knowledge-based robotic systems. Fuzzy Sets and Systems.
26(2): 179-193, 1988.
[19] E.O. Doebelin. Control System Principles and Design. J. Wiley and Sons, New York,
1985.
[20] Y. Dote, A. Suyitno, and M. Strefezza. Fuzzy learning grasping force controller
for manipulator hand. In lECON '90: 16th Annual Conference of IEEE Indu,<itrial
Electronics Society, pages 1259-1265, 1990.
[21] D. Dubois and H. Prade. Fuzzy rules in knowledge-based systems - modelling grade-
ness, uncertainty and preference. In R. Yager and L. Zadeh, editors, An introduction
to fuzzy logic applications in intelligent systems, pages 45 68. Kluwer Academic Pub-
lishers, Dordrecht, The Netherlands, 1992.
[22] A.R. Fraser and R.W. Daniel. Perturbation techniques for flexible mantpulaiors.
Kluwer Academic Publishers, Dordrecht, The Netherlands, 1991.
[23] T. Fukuda, Fujiyoshi, F. Aral, and H. Matsuura. Design and dextrous control of
micromanipulator with 6 dof. In IEEE International conference on Robotics and
Automation, pages 1628-1633, 1991.
[24] C.E. Garcia and M.Morari. Internal model control - a unifying review and some new
results. Ind. Eng. Chem. Process Des. Dev., 21, 1982.
[25] A. Garcia-Cerezo, A. Barreiro, and A. OUero. Design of robust intelligent control
of manipulators. In IEEE International Conference on Systems Engineering, pages
225-228, 1991.
[26] .1. Gasos, M.C. Garcia-AUegre, and R. Garcia Rosa. Fuzzy local navigation of a
simulated real robot in unknown environment. In IEEE International Worksliop on
Intelligent Motion Control, pages 445-449, 1990.
[27] R. Gorez, M. De Neyer, and J. Barreto. Fuzzy internal model control. In S. G. Tzafes-
tas, editor, Engineering Systems with Intelligence - Concepts, Tools and Applications,
pages 41.5 422. Kluwer Ac. PubL, Dordrecht, The Netherlands, 1991.
487
[43] ('. Isik. Inferfiicp engines for fuzzy rule-based control. International Journal of
Approximate flca.foru'ng, 2(2): 178-188, 1988.
[44] (', Isik and A.M. Meystel. Pilot level of a hierarchical controller for an unmanned
mobile robot. IEEE Journal of Robotics and Automation, 4(3):241-25.'j, 1988.
[45] R. Jager, H.B Verbruggen, and P.M. Bruijn. The role of defuzzification methods in
the application of fuzzy control. In Symp. on Intelligent Components and Instruments
for Control Applications, pages 111-116, Malaga, 1992.
[46] CI.S. Kalley and K.J. McDermott. Using expert systems to incorporate the principles
of motion economy as a means of improving robot manipulator path generation.
Computers and Industrial Engineering, 16(32):207-213, 1989.
[47] H. Rang and G. Vachtsevano.s. An intelligent strategy to robot coordination and
control. In CDC '90: 29th IEEE Conference on Decision ajid Control, pages 2208-
2213, 1990.
[48] C. Kemal. Fuzzy rule-based motion controller for an autonomous mobile robot. RobcA-
ica, 7(l).-37 42, 1989.
[49] ('hen Ken, Lin ,Jinn-Ya, and Lu Yong Xiang. Fuzzy control of robot manipulators.
In 1988 IEEE International Conference on Systems, Man and Cybernetics, pages
1210-1212,1988.
[•)0] K. Kevin. Self - designing with fuzzy logic. IEEE Spectrum,, pages 42-44. November
1990.
[.51] 0 . Khatib and J. Burdick. Motion and force control of robot manipulators. In IEEE
Conference on Robotics and Automation, pages 1381-1386, Malaga, 1986.
[52] W.,J.M. Kickert and H.R. van Nauta Lemke. Application of a fuzzy controller in a
warm water plant. Automatica, 12:301-308, 1976.
[53] J.W. Kim, Y.K. Ha, and Y.P. Park. Vibration control of a flexible arm using fuzzy
control theory. In Eirst International Conference on Motion and Vibration Control,
|)ages 517-522, Yokohama, 1992.
[54] P.,J. King and E.H. Manidani. The application of fuzzy control systems to industrial
processes. Automatica, 13:235-242, 1977.
[55] B.D. Klafter, Th.A. ("hmielewski, and M. Negin. Robotic Engineering - An integrated
approach. Prentice-Hall International Editions, Englewood (Jliffs, N.J., 1989.
[56] K.C. Koh, A.H.S. Cho, S.K. Kim, and I.S. Jeong. AppUcation of self-organizing fuzzzy
control to the joint control of a puma 760 robot. In IROS '90 IEEE Intertiational
Workshop on Intelligent Robots and Systems, pages 537-542, 1990.
[57] T.K. Kubota and H. Hashimoto. A strategy for collision avoidance among moving
obstacles for a mobile robot. In 1 Ith IFAC World Congress on Automatic Control in
the Ser-vice of Mankind, pages 105-110, 1990.
[.58] D.P. Kwok, P. Tam, C.K. Li, and P. Wang. Linguistic PID controllers. In 11th IFAC
World Congress on Automatic Control in the Ser-vice of the Mankind, pages 192- 197,
Tallin, Estonia, 1990.
489
[59] 1). Lakov. Adaptive robot under fuzzy rontrol. Fuzzy Sets and Systems, 17(1):1 S,
1985.
[60] Y. Landau. Adaptive control: The model reference approach. M. Dekker Inr., New-
York, 1979.
[61] Ch. Lee. Fuzzy logic in control systems: Fuzzy logic controller, part II. IEEE Trans,
on Systems, Man mid Cybernetics, SMC-20('2):419—135, 1990.
[62] Y. Li. The automatic recognition of mechanical parts on membership's principles of
fuzzy sets. International Society for Optical Engineering, 697:175-180, 1986.
[63] Y.F. Li and C.(!. Lau. Development of fuzzy algorithms for servo systems. IEEE
Control Systc7ns Magazine, 9(3):65-71, 1989.
[64] ('.M. Lim and T. I Hiyama. Application of fuzzy logic control to a manipulator. IEEE
Trans. o7i Robotics and Automation, 7(5):688-691, 1991.
[65] C.E. Lin and Y.R. Sheu. A hybrid control-approach for pendulum-car control. IEEE
Trans, on Industrial Electronics, 39(3):208-214, 1992.
[66] K.H. Low and M. Vidyasagar. A Lagrangian formulation of the dynamic model for
flexible manipulator systems. ASME Journal of Dynamics, Measurement and Control,
110:17.5-181, 1988.
[77] R. Palm. Fuzzy controller for a sensor guided robot manipulator. Fuzzy Sets and
Systems, 31(2): 133-149, 1989.
[78] R. Palm. Control of a redundant manipulator using fuzzy rules. Fuzzy Sets and
Systems, 45(3):279"298, 1992.
[79] R.C;. Paul. Robot Manipulators: Mathematics, Programming and Control. The MIT
Press, Cambridge, Mass., 1981.
[80] B. Postlethwaite. Basic theory and algorithms for fuzzy sets and logic. In M. Criniblp
J. McGhee and P. Mowforth, editors. Knowledge-based systems for industrial control.
Peter Peregrinus Ltd, Stevenage, UK, 1990.
[81] M. Raibert and J. Craig. Hybrid posit ion/force control of manipulators. Journal of
Dynamic Systems, Measurement and Control, 1Q3(2):26-133, 1981.
[82] B. Robyns, D. CJalardini, R. Gorez, F. Labrique, and H. Buyse. Asservissement de
Vitesse d'un actionneur asynchrone par modele de I'actionneur ct de sa cominande
vectorielle. In Journees d'etudcs S.E.E. sur les asscrxnssements electromecaniques
rapides: modelisation et regulation avancees, pages 3-51, Metz, France, 1992.
[83] E. Scharf. f^uzzy logic could redefine robot control. Automation, 21(2);11-14, 1985.
[84] E.M. Scharf and N.J. Mandic. The application of a fuzzy controller to the control of
a multi-degree of freedom robot arm. In M. Sugeno, editor. Industrial Applicatioiis
of Fuzzy Control. North-Holland, Amsterdam, The Netherlands, 1985.
[85] K. Seung-Woo and P. Mignon. Fuzzy compliance robot control. In IROS' 91 Inter-
national Workshop on Intelligent Robots and .Systems, pages 1628 1631, 1991.
[86] L. Shiyong, H. Hengzhang, and S. Andong. A three dimensional fuzzy intelligent
controller. In 1988' International Conference on Systems, Man and Cybernetics.
pages 1060-1061, Malaga, 1988.
[87] K.M. Stellakis and K.P. Valavanis. Fuzzy logic-based fonnvdation of the organizer of
intelligent robotic systems. Journal of Intelligent and Robotic Systems: Tlieoi-y and
Application.% 4(1):1 -24, 1991.
[88] I). Stipanicev, M. De Neyer, and R. Gorez. Self-tuning self-organizing fuzzy robot
control. In I. Troch, K. Desoyer, and P. Kopacek, editors, Robot Control, pages
171 176. Pergamon Press, Oxford, U.K., 1992.
[S9] ,). Swevers. Linear identificatio7i and contwl of flexible robots. PhD thesis, K.U.
Leiiven, 1992.
[90] T. Takeuchi. An autonomous fuzzy mobile robot. Advanced Robotics, 5(2):215-230.
1991.
[91] T. Takeuchi, Y. Nagai, and N. Enonioto. Fuzzy control of a mobile robot for obstacle
avoidance. Information Sciences, 45(2):231-248, 1988.
[92] K.L. Tang and R.J. Mulholland. Comparing fuzzy logic with classical controller
designs. IEEE Trans, on Systejns, Man and Cybernetics, SMC-17(6):1085-1087,1987.
[93] R. Tanscheit and E.M. Scharf. Experiments with the use of a rule-based self-organizing
controller for robotics appUcations. Fuzzy Sets and Systems, 26:195-214, 1988.
This page intentionally blank
491
[94] S.G. Tzafestas. Adaptive, robust and fuzzy rule-based control of robotic nianiprilators.
In S.G. Tzafestas, editor, Intelligent Robotic Syfitems, pages 313-419. M. Dekker, New
York, 1991.
[95] M. Uragami, M. Mizumoto, and K. Tanaka. Fuzzy robot controls. Journal of Cyber-
neticse, 6(l-2):39~64, 1976.
[96] G.J. Vachtsevanos, K. Davey, and K.M. Lee. Developenient of a novel intelligent
robotic manipulator. IEEE Control Systems Magazine, 7(3):9-15, 1987.
[97] K.P. Valavanis and K.M Stellakis. A general organizer model for robotic assemblies
and intelligent robotic systems. IEEE Trans, on Systems, Man and Cybernetics,
SMC-21(2):302-317, 1991.
[98] .]. Van Amerongen, H.R. van Nauta Lemke, and J.C.T. Van der Veen. An autopilot
for ships designed with fuzzy sets. In H.R. van Nauta Lemke, editor, Digital Computer
Applications to Process Control. North-Holland, Amsterdam, The Netherlands, 1977.
[99] B.A.N. Wakileh and K.S. Gille. Use of fuzzy logic in robotics. Computers in induMry,
10(l):35-46, 1988.
[100] D. Wang and M. Vidyasagar. Transfer functions for a single flexible link. International
Journal of Robotics and Automation, 10(5):540-549, 1991.
[101] D. Whitney. Historical perspective and state of the art in robot force control. In
IEEE Conference on Robotics and Automation, pages 262-268, Malaga, 1985.
[102] Hyun Woong Keun, II Hong Suh, and Kyong Gi Kim. Intelligent control of redundant
manipulators in an environment with obstacle. In IROS' 91 International Workshop
on Intelligent Robots and Systems, jiages 161-166, 1991.
[103] Y. Xu, R.P. Paul, and H.Y. Shun. Fuzzy control of robot and comphance wrist
systems. In IEEE Indu.ftry Applications Society Annual Meeting, pages 1431 1437,
1991.
[104] W. Yim, J. Zuang, and S. Singh. Experimental dual-mode control of aflexible robotic
arm. International Journal of Robotics and Autoination, 10:135-145, 1992.
[105] T. Yoshikawa. Foundations of Robotics: Analysis and Control. The MIT Press,
Cambridge, Mass., 1990.
[106] L.A. Zadeh. Fuzzy sets. Information and Control, 8:338-353, 1965.
[107] L.A. Zadeh. A rationale for fuzzy control. Trans.A.S.M.E., ser.G, 94:3-4, 1974.
[108] H.J. Zimmermann. Approximate reasoning in manufacturing. In A. Kusiak, editor.
Intelligent Design and Manufacturing, pages 701-722. Wiley-Interscience, New York,
1992.
CHAPTER 18
FUZZY CONTROL FOR ROBOT MANIPULATORS
WITH ARTIFICIAL RUBBER MUSCLES
2. C O N T R O L OBJECT
[Compressor J
Regulator
'
Personal D/A V/A Servo
computer -»- converter converter
—»- valve
—^- Manipulator
PC-9801
1
Encorder
-
Reducing pressure
Air
Air - ^
Adding pressure
3. C O M P U T E D T O R Q U E C O N T R O L L E R
As a mathematical model of the rubber muscle manipulator, consider the usual
SCARA-type robot manipulator model described by
where 0 e TZ" is the joint angle vector,^ T e W is the driving input, M(ff) £ K"''^"
is the positive inertia matrix and V(0,0) <E K" is a iionHnear term which represents
the Coriolis force, centrifugal force and gravitation, etc. It is assumed that we set
the desired reference joint angle 64, angular velocity 0i and angular acceleration 64.
H we have the output joint angle 9, the angular velocity 0, and the model matrices
M(l?) and V(ff, 0), then the computed torque can be constructed by
Fuzzy
compensator
e
Torque Robot
computation e
II a
4. D E S I G N O F A F U Z Z Y L O G I C C O N T R O L L E R
for 2n-input variables ( e i , . . . , e„, e j , . . . , e„) and n-output variables ( u j , . . . , Un). Here.
R, is the i-th control rule, A,j is the fuzzy set of antecedent part corresponding to
497
the j-th input variable in the i-th control rule, and B,j is the constant value corre-
sponding to the j-th o u t p u t variable in the i-th control rule. Using 2n confidences
fiAniei),..., fj.An„^{en), fJ-An^^i^iei),..; tJ.An2„)(^n), we have t h e confidence of the
antecedent part in the i-th control rule, hi, as follows:
h, = /i^,i(ei)A/J^,.,(e2)...A/iA,„(e„)A/i^.,„^;j(ei)A/iA,(„+j,(e2).-A^A,(,,,,(en) (7)
where r denotes the total number of control rules; generally r = f^" if the n u m b e r
of labels is i.
In order to transform the input variables to the values on the support set,
for example, on [—6, 6], we need the scalers for the i n p u t - o u t p u t d a t a of the fuzzy
reasoning. Here, we define t h a t GIN(i),? = l , . . . , 2 n are the input scalers for the
input d a t a e i , . . . , e„, ii,..., e„ and G O U ( i ) , i = 1,..., n are the o u t p u t scalers for the
inferred consequents u*, ...,u*.
GIN(l)
Calculation GOU(l)
of Defuzzi-
Ae minimum fication
grades
GIN(3)
GIN(2)
G0U(2)
G(
Calculation ^ \
of Defuzzi- U ^ 2f
Ae minimum fication
^ grades
h^
GIN(4)
0.91
if |GIN(i)oM X e,| > I
GIN(One (9)
GIN(i)oia otherwise
This means t h a t if the scaled error of the joint angle is scaled out from the support
set [ ~ i , X], then the input d a t a is rescaled to fall on the 90% range of the support
set. Similarly, the scalers GIN(i + n ) , . . . , GIN(2n) for the input d a t a (i.e., the error
of the joint angular velocity) e j , . . . , e „ can also be adaptively determined by the
similar condition.
J{k)^\Y.\\e{kTW (14)
and construct the learning algorithm based on minimizing this function. Here, N
denotes the total number of samplings. Figure 5 shows the block diagram of a
multiple fuzzy control system. From equations (12)-(14), the gradient of J with
respect to w, becomes
dJ{k) dJ{k) de{kT)du{kT)
dwi " de{kT) du(kT) dw,
*:=0 ^ '
500
Learning
mechanism
+ e
Fuzzy «i
[^Le controller
1
Plant
H Fuzzy
Uu
controUer
M
Linear NN
Therefore, using the delta rule (Rumelhart et al., 1986) gives the following u p d a t e
equation for w,:
N
de{kT)
o,{i +1) = «,.(/) + , Y. ^^(^^)^;7Pf("'(^-^) ;i6)
dv{kT)
fc=0
where / denotes the l-ih. u p d a t e time and r; is a small positive constant. Since in the
application of a fuzzy control the plant structure or p a r a m e t e r s may be unknown,
we will evaluate the Jacobian 89 jdu in equation (16) as
de,{kT) AOiikT)
l,...,m,j = l,...,p (IT)
dujikT) Auj{kT)
where Auj{-) and A^i(-) are generated from the input a n d o u t p u t data. If the plant
is originally a discrete-time system with no any time-delay of the input, then we
must evaluate d6i[kT)/duj[(k — 1)T] instead of (17). Note however t h a t if the plant
s t r u c t u r e and parameters are completely known, then we can obtain the Jacobian in
an accurate form. Note furthermore t h a t we can readily apply the s t a n d a r d linear
identification algorithm, for example, by the K a l m a n filter ( W a t a n a b e f.t al., 1991)
t o estimate toi, because equation (13) is linear with respect to iv,.
6. E Q U A T I O N O F M O T I O N F O R T H E MANIPULATOR
We consider a two-link manipulator model as shown in Fig.6. T h e equation of
motion is assumed to be given by (1), but with
joint 1
Fig.6 Two-link m a n i p u l a t o r model
7. S I M U L A T I O N EXAMPLES
To simulate the motion for the above r u b b e r muscle manipulator, we use the
4th-order Runge-Kutta-Gill m e t h o d . It is assumed t h a t the control sampling period
is T = 25 [ms], the step width of t h e integration is 0.5 [ms], and the initial states are
01 = 0.7304 [rad], 02 = 0.4858 [rad], ^i = ^ 2 = 0 . Further, assume t h a t the reference
A
trajectories Xd{t) x{t) y{t)] in the work-coordinate systems are given by
nominal actual
Link 1 1.39 0.02734 0.03281 0.385 0.160
Link 2 0.65 0.00117 0.00189 0.110 0.057
NB NM ZO PM PB
0
-12
Fig.7 Membership functions for 5 labels
which are the circular trajectory with 25 [mm] radius, and the corresponding cycle
is 10 [s]. From solving the inverse kinematics problem, the reference angles Oj{t) =
[Odi{t) ^d^]^ in the joint-coordinate systems are also given by
Ae,
NB NM ZO PM PB
NB -12 -12 -12 -6 -6
NM -6 -6 -6 -6 6
ZO -12 -6 0 6 12
PM -6 6 6 6 12
PB 6 6 12 12 12
Link 2 40 6 *
Fig.ll Link 1 30 8 150
T 1 I 1 1 1 1 1 1 r
Set point
55 0.8 Response
0 5 10
1 1 1 1 1 1 1 1 1 1 1
^1.2
T3
(0 - ,f^
// ^
^0.8 - // ^ V
CD
- //// \
\ -
3 //
// -
3-04 " \ //
• \
//
O 1 1 I l l 1 ""
5 10
Time (s)
~1 1 1 1 r "1 I V
Set point
"i 0.8
Response
J I ] L
0 5 10
1 1 1 1 1 1 1 1 1 1 1
1?
T3
km.
^^-^
(M
OR \
O)
+-•
3 \ -
Q. -
3
0.4
o " 1 1 1 1 1 1"
5 10
Time (s)
Fig.9 Control results due to the computed torque inethod
with fuzzy compensation
306
1—I—I—I—r
f 0.4
5
Time (s)
Fig.10 Control results due to the computed torque method
with tuned feedback gains
1 1 1 1 1 1 1
I 380 //
//
// \
i
I
\\
' 360 1
>- ll
1
//
//
//
//
340- >>•
1 1 1 1 1 1 1
ei Ae,-
NB ZO PB
NB -12 -12 -12
ZO -12 0 12
PB 12 12 12
1 1 1 1 r
l i 0.005
CD
o
I™
CD
g-O.OOSh
W 0.02
CD
m
"3 -0.02h
f
Time (s)
Fig. 12 Control results due to the elemental fuzzy
controllers FCl and FC2
1 I I—I I—I—I—I—I—1—r
2 0.004 0 th
f: -0,004
zi
a.
3 -O.OOBi -J t J I I L I .J I __1__iZ
0 5 10
g 0.02
CM
CD
B- ^0.02
J i L I I i I I I I I I
Time |s)
"E* 380
.2
T 360
340:
8. C O N C L U S I O N S
We have applied the fuzzy logic controller for controlling a irmlti-link robot
manipulator with rubber muscles. We usuallj can not obtain the perfect physical
parameters to mathematically model the manipulator and we also require some
trials and errors to tune the feedback gains of the computed torque controller, even
though we have a considerably accurate model. Using the independent fuzzy rea-
soning, we have first shown that the simple fuzzy controller is useful for designing a
robust computed torque control system. Kext, we have also shown that an iterative
learning fuzzy controller is eiiective for controlling the same manipulator without
any modehbased controller.
9. R E F E R E N C E S
C h e n , Y.H., R o b u s t Control System Design: K o n - A d a p t i v e Versus A d a p t i v e , Int. J. Contro!, 51-6,
pp.1457^1477 (1990)
Corless, M . J . , Tracking Coiitrolleis for Uncertain Systems: Application to a Mantitec R3 R o b o t , ASME
J. of Dt/namic Syaiems, Meas. and ConiToI, 1 1 1 , pp.609^618 (1,989)
Corless, M.,1. and L e i t m a n n , G., Coiitiiiuoiis S t i t e Feedbaclt G u a r a n t e e i n g Uniform U l t i m a t e Bounded-
ness for Uncertain Dynamical Systems, IEEE Trans. Aut. Control, A C - 2 0 , p p . H S 9 - H 4 4 (1981)
Daley, S. and Gill, K . F . , A design s t u d y of a self-organising fuzjy logic controller, Proc. htsi. Meek.
Eng., 2 0 0 - C l , pp.59^69 (1986)
510
DeCarlo, R.A., Zak, S.H., and M a t t h e w s , G.P., Variable S t r u c t u r e Control of Nonlinear Multivariable
Systems: A Tutorial, Proc. IEEE, 7 6 , pp.212-232 (1989)
Linkens, D.A. and Hasnain, S.B., Self-organising fuzzy logic control and application to muscle relaxant
anaesthesia, lEE Proceedtngs-D, 1 3 8 - 3 , pp.274-284 (1991)
Hayashi, I., Nomura, H., and Wakami, N., Acquisition of Inference Rules by Neural Network Driven
Fuzzy Reasoning, J. of Japan Society for Fuzzy Theory and Systems, 2-4, pp.585-597 (1990) (in
Japanese)
Horikawa, S., Furuhashi, T., O k u m a , S., and Uchikawa, Y., A Learning Fuzzy Controller Using a Neural
Network, Trans, of the Society of Instrument and Control Engine€rs,27-2, pp.208-215 (1991)
(in J a p a n e s e )
Hui, S. and Zak, S.H., R o b u s t Synthesis for U n c e r t a i n / N o n l i n e a r Dynamical Systems, Automatica, 28-2,
pp.289-298 (1992)
Imura, J., Sugie, T., and Yoshikawa, T., A d a p t i v e R o b u s t Control for R o b o t M a n i p u l a t o r s , Trans, of
the Society of Instrument and Control Engineers, 2 7 - 3 , pp.314-319 (1991) (in J a p a n e s e )
Maeda, M. and M u r a k a m i , S., Self-Tuning Fuzzy Controller, Trans, of the Society of Instrument and
Control Engineers, 24-2, pp.191-197 (1988) (in J a p a n e s e )
M a m d a n i , E.H., Application of fuzzy algorithms for control of simple d y n a m i c plant, Proc. lEE, 121-12,
pp.1585-1588 (1974)
M a m d a n i , E.H., Advances in t h e Linguistic Synthesis of Fuzzy Controller, Int. J. Man-Machine Studies,
8-6, pp.669-679 (1976)
M i z u m o t o , M., Fuzzy Reasoning M e t h o d s for Fuzzy Control, J. of the Society of Instrument and
Control Engineers, 2 8 - 1 1 , pp.959-963 (1989) (in J a p a n e s e )
Procyk, T . J . and M a m d a n i , E.H., A Linguistic Self-Organizing Controller, Automatica, 1 5 , pp.15-30
(1979)
Psaltis, D., Sideris, A. and Y a m a m u r a , A.A., A Multilayercd Neural Network Controller, IEEE Control
System Mag.,, 8, pp.17-21 (1988)
R u m e l h a r t , D.E., McClelland, J.L., and the P D F Research G r o u p , Parallel Distributed Processing, 1,
M I T Press (1986)
Tanji, J. and Kinoshita, M., A Fuzzy Controller with a R o b u s t Learning Function, T r a n s , of ikt Society
of Instrument and Control Engineers, 23-12, pp.1296-1303 (1987) (in J a p a n e s e )
Utkin, V.L, Variable S t r u c t u r e Systems with Sliding Modes, IEEE Trans. Aut. Control. AC-22,
pp.212-222 (1977)
W a t a n a b e , K., Fukuda, T. and Tzafestas, S.G., Learning algorithms of layered neural networks via
extended K a l m a n filters, Int. J. Systems Set., 22-4, pp.753-768 (1991)
W a t a n a b e , K. and Tang, J., Learning Controller based on Fuzzy Gaussian Neural Network, Proc. of
2nd Symposium, on Intelligent Systems, Nagoya, 920-87, pp.255-260 (1992) (in J a p a n e s e )
W a t a n a b e , K., Shiramizu, K., and Fukuda, T., Multiple Fuzzy Controls, Trans. JSME, Series G,
Vol.58, No.554, pp.2970-2976 (1992) (in J a p a n e s e )
W a t a n a b e , T. and Ichihashi, H., Fuzzy Control of a Robotic M a n i p u l a t o r by the Feedback Error Learn-
ing, Trans, of the Institute of Systems, Control and Information Engineers, 3-7, pp.212-217
(1990) (in J a p a n e s e )
Yamazaki, T. and Sugeno, M., Self-Organizing Fuzzy Controller, Trans, of the Society of Instrument
and Control Engineers, 20-8, pp.720-726 (1984) (in J a p a n e s e )
CHAPTER 19
Carl G. Looney
I. INTRODUCTION
Networks for information flow were invented in the dissertation of C. A. Petri (1961) [21]
at the University of Bonn. A. W. Holt and F. Commoner (1970, [9]) extended them and
called them Petri nets. Peterson (1981, [20]) and Reisig (1985, [23]) provide treatises on
the theory of Petri nets, that are used nowadays to model flows such as parts and
materials in manufacturing processes (Zhou and Leu, 1991 [32]), and logical truths in
machine reasoning systems (Looney, 1988, [15], and Murata et al, 1991, [19]).
For our purposes, a logic Petri net P consists of i) a net architecture; and ii) an
operational procedure. Architecturally, it is a directed graph that consists of two kinds
of nodes: a) conditions, designated by circles; and b) events, denoted by bars. The
conditions and events are connected by arrows according to two rules: 1) an arrow may
connect from a condition to an event, or from an event to a condition; and 2) an arrow
may never connect two nodes of the same kind. Figure 1 is an example of such a scheme.
Procedurally, the net uses tokens, represented graphically by dots inside condition
nodes, to activate the conditions where they appear. A token denotes a Boolean truth
value of 1 for the condition, while the absence of a token indicates a value of 0. An event
is enabled when every arrow entering it comes from a condition that contains a token. An
enabled event fires to activate (i.e., make true) all conditions to which its departing
arrows directly connect, by sending them tokens. In Figure 1, Condition C2 is activated,
i.e., made true, with a token. A clock is implicit in the net in that it connects to each
event so that upon being enabled, an event fires on the next clock pulse. A clock is
necessary to maintain sequential order.
A logic Petri net that interacts with the external world has boundary nodes that
connect to the external environment. In Figure 1, Condition Cj, and Events Ej, and E^,
are boundary nodes. Arrows enter boundary conditions from the environment, and arrows
depart boundary events to the environment. External events that activate boundary
conditions are called sources, while external conditions that are activated by boundary
events are called sinks. The operation of a logic Petri net is initiated when certain source
events fire to activate boundary conditions. These enable events that fire to activate other
conditions, which may in turn enable other events to fire to activate other conditions, etc.
511
S.G. Tzafestas and A.N. Venetsanopoulos (eds.)
Fuzzy Reasoning in Information, Decision and Control Systems, 511-527.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
512
•oF' /^^2 ^
~1
\^)—
1 EG'N. fs '
I )K}__ ^C^
1 ^r
1 '^ET^" "^'1
1 C g ^ CyM 1
1 "cP
I -^6 J
——. / — — /
SINK/ BOUNDARY SINK/
\oJ—^
1
1
•^10
Ep ^ " ^ ^ ^
~ ^ 1
1
1
1
CPN
" ^
^H-^1
L J
SINKi^7 BOUNDARY SINK/
Conditions are propositions, i.e., statements that must be either true or false. In Figure
1, conditions Cj, Cj and C^Q are true initially, as designated by the tokens. This causes
Event Ej to be enabled by Cj and Cj, so that Ej fires on the next clock pulse to activate
513
conditions C3 and C4, as shown in Figure 2. The truth of C4 enables Event E3, which fires
to activate Condition C^, and C3 enables Ej, which fires to activate both conditions C5
and Cg. Further enabling, firing, and activation cause Event E5 to fire to propagate a truth
token to a sink. Note that there is no conservation of tokens, which may split or combine.
The firing of events corresponds to the modus ponens syllogism. Figures 1 and 2
show that the net architecture contains the built-in rule [(Cj AND Cj) - • (C3 AND C4)],
which is equivalent to the two rules [(C^ AND Cj) -» C3] and [(C^ AND Cj) "* C4].
When the rule antecedent (C^ AND C2) is activated (made true), then the rule events
(implications) are enabled to fire and activate the rule consequents C3 and C4. Thus AND
logic can be modeled in both antecedents and consequents of rules. Condition Cg can be
activated by the firing of either Event E3 or Ey, so OR logic is implemented. We
incorporate NOT into the net by putting a small circle on the end of an arrow to denote
negation, as shown at Event E7 in Figures 1 and 2, to signify the rule [NOT CJQ - • Cg].
Nonspecific rules of the form [C^ - • (C^ OR C„)] are not modeled because they do not
specify which of the consequent conditions is activated. Rules of the form (C OR C^) -*
Cn, may be considered to be two separate rules: [Cj -» C^], and [C^ -» C^].
Now let C = {Cj,...,C„} be the conditions of a logic Petri net P. A marldng of P is
a vector M = (mi,...,mj that distributes truth values over the n conditions, where raj =
1 denotes a token (truth value of 1) for Condition Cj, else m; = 0 indicates that C; is false.
We also call M the truth state of P. In Figure 2 the truth state is M =
(1,1,1,1,0,0,0,0,0,1). Because a Petri net is dynamic, there is particular interest in any
sequence {Mo,Mj,...,Mf} of markings that ensues from firings of events due to an initial
marking MQ. We say that a marking Mj is reachable from MQ if and only if there is a
sequence of such markings MQ,...,Mf. When an initial marking M^ starts a sequence of
markings (truth states), the net may eventually reach a marking Mj that does not change
further. In such a case, Mj = (mj^^,...,m„^^ is the final truth state and indicates all of the
true conditions that can be deduced from the initial truths in MQ. Deadlock is a situation
where the logic Petri net has no enabled events, in which case no rules can fire to
propagate truths. A deadlocked Petri net remains in a fixed state (marking).
used fuzzy truth state operated on by fuzzy rule matrices with MIN/MAX logic to reason
forward, and also used a fuzzy Petri net (1988) [15] to model fuzzy reasoning for control.
Lipp (1983) [13] also used a Petri net for modeling a complex controller, but it did not
implement logic flow. Postlethwaite (1990) [22] discusses fuzzy expert systems.
A fiazy set F (with respect to a universal domain set X) is the set of all elements x
e X, along with a membership Junction fif(x), where 0 « ^p(x) s 1, that gives the extent
to which each x in X belongs to F. For fuzzy sets F and G on the domain X, the fiazy
intersection is the set FDG of all elements in X with the fuzzy membership function
defined via: fifpfjOi) = min{Mp(x),/Zo(x)}. Thus x belongs to FflG only to the (least)
extent that it is a member of both F and G. The fiizzy union is the set FUG of all
elements x e X with fuzzy membership % U G W ~ ^^^{f^^)^(^)}- Then x belongs to
FUG to the (greatest) extent that it belongs either to F or G. The fuzzy complement of
F is ~F with fuzzy membership function yM_p(x) = 1 - fif(x), so that x belongs to ~F to
the extent that it does not belong to F.
When let X and Y be two different universes of discourse, and F is a fuzzy set on X
and G is a fuzzy set on Y, then what is the meaning of FflG and FUG? Because / ^ x )
and ^ ( y ) each take on values between 0 and 1, they are comparable, so we define E =
FOG to be the fuzzy set on XxY that has set membership function fij:fya{^,y) =
m[n{/if(x),fiQ(y)} for any (x,y) E XXY. This is the extent to which it is true that x belongs
to F and y belongs to Y. Similarly, H = FUG is a fuzzy set with /HpjQix,y) =
Implication is a combinational operation that assigns a fuzzy truth to the result, and
is not the same as the modus ponens law of logic [A AND [A -»(f) B] that means: the
implication fuzzy truth f of [A -* B] and the fuzzy truth |A| of A imply the fuzzy truth
|B| of B. The fuzzy truth f is a proportion of openness of a fuzzy switch, so an
implication [A -»(f) B] with truth |A -»(f) B| = f > 0, and the fuzzy truth |A| of A, imbue
B with the truth of A up to the truth (openness) f of the implication switch. On the other
hand, the implication truth f can not pass a truth to B that is greater than the source truth
of A. Therefore, the model for the truth of B is |B| = min{|A|,f} = |A| MIN (f). This
model also holds for implication in Boolean logic. Of course, B may have a truth value
due to some other implication or due to fuzzification. We may restrict the implied truth
of B to the A-context, i.e., the truth contribution to B only from A, designated by jB^^^J.
Contextual fuzzy implication holds for Boolean logic also.
Fuzzy modus ponens: This is the basic implication law: (A AND [A -»(f) B]) - • B.
The transferred fuzzy truth due to A and [A - • B] is defined to be |B/^J = |A| MIN f,
according to the above discussion. A must be initialized with a positive truth value, i.e.,
become a fUzzy fact, to fire and activate B with a fuzzy truth value. Fuzzy modus ponens
and contextual implication allow the building up of evidential truth for B from the various
contextual implications, and then combining all truths logically for a final truth value, say
|B| = max{|B,y^-,|,|B,D-,|}, where A and D each implies B. For example, let (A ^•(0.3) B)
and (D -»(0.8) B) be propositions. For |A| = 0.6 and |D| = 0.7, B^^) is activated (only
from A) with the truth min{0.6,0.3} = 0.3, but B-^s is activated (only from D) with the
truth min{0.7,0.8} = 0.7. Thus the truth of B fi-om A OR D is |B| = max{|B, J.lD,^)!} =
max{0.3,0.7} = 0.7.
Not all laws of Boolean logic extend to fuzzy logic (see Sugeno and Takagi, 1983
[27], Dubois and Prade, 1987 [2], and S. Weber, 1985 [29]). Kosko (1986) [10]) questions
the truth of the law of contraposition (modus tollens) in the world of partial (fuzzy)
implication (i.e., (A - • B) is true if and only if (~B - • ~A) is true). His example is that
(smoking -» lungcancer) is true to some extent, but that does not imply the extent to
which (NOT lung_cancer -* NOT smoking) is true. We demonstrate that fuzzy
contraposition holds under constraints.
Fuzzy Contraposition: Let |A| s f and 1-|A| s f. Then X = (A -»(f) B) has fuzzy truth
f if and only Y = (~B -•(f) ~A) has fuzzy truth f. For the Boolean case where f = 1, this
always holds. We note that this is not the fuzzy modus tollens logic law: (NOT B AND
[A - • B]) -^ NOT A).
516
Proof: Let |A| s f and |~A| = 1-|A| s f. Then X = (A -»(f) B) means that |B| =
min{|A|,f} = |A|. Thus 1-|B| = 1-|A| so that |~B| = |~A|. Then we can write (~B -•(f) ~A),
because |~A| = |~B| = min{|~B|,f}. By symmetry due to ~(-F) = F, the argument in the
reverse direction holds.'
For a counterexample, let |A| = 0.2 and f = 0.6, so (A -*[0.6] B) means that |B| =
min{0.2,0.6} = 0.2. But |~B| = 0.8, so that (~B -*[0.6] ~A) means that |~A| =
min{0.8,0.6} = 0.6 ^ 0.8 = 1 - 0.2 = 1 - |A|. Thus contraposition fails here, and f < 1 -
|A|. For another counterexample, let |A| = 0.75 and f = 0.5, so that |B| = min{0.75,0.5}
= 0.5. Then contraposition (modus tollens) means that |~A| = min{|~B|,f} = min{0.5,0.5}
= 0.5 K = 0.25 = 1 - 0.75 = 1 - |A|. For an example that satisfies the hypothesis of fuzzy
contraposition, let |A| = 0.4 and f = 0.8. Then |B| = min{0.4,0.8} = 0.4. Contraposition
then implies that |~A| = min{|~B|,f} = min{ 1-0.4,0.8} = 0.6, so that |A| = 1 - |-A| = 0.4,
which is consistent.
The fuzzy operations NOT, MIN, MAX, contextual implication, and equivalence are
sufficient for our needs, and they imply the truths of certain laws such as the De Morgan
laws, which are needed to convert to conjunctive normal form (CNF) or disjunctive
normal form (DNF) (see Schalkoff, 1990 [25], or Rich and Knight, 1991 [24]). The
former form is the conjunction (ANDing, or MINing) of disjunctions (ORing, or MAXing,
of variables), e.g., (A MAX B) MIN (C MAX D MAX E) MIN (F). The latter form is
the disjunction (ORing, or MAXing) of conjunctions (ANDed, or MINing, of variables),
e.g., (A MIN B MIN C) MAX (D) MAX (E MIN F). These are also known as product-
of-sums and sums-of-products, respectively.
Fuzzy De Morgan Laws: i) ~(X MIN Y) = (~X MAX -Y); ii) -(X MAX Y) = (~X
MIN ~Y).
Proof: The first one follows from the fact that ~(X MIN Y) has the truth value 1 -
min{X,Y} = 1 - (1 - max{l-X,l-Y}) = max{l-X,l-Y} = (~X MAX ~Y). The second
argument is similar.'
fuzzy implication truths. Conditions may have zero truth values until initialized with fuzzy
truths for a process of fuzzy deduction, at which time they become fuzzy facts. Others
may obtain nonzero fuzzy truths upon being activated by the firing of one or more rules.
Aft4zzyrule-based system contains: i) a fuzzy knowledge base; ii) a condition initializing
function to post fuzzy facts; iii) an inference function for scheduling and interpreting
rules (to fire each rule or not); iv) a user interactive procedure; and v) an editor.
-^02
Cs
E X ^ ^^^A^
1—
z
'y^"uT^ LJ
z-
•
^3/
>
C*\ •-10 / Z
UJ
V^Cg Q r^s K
A3 %
^02
^03 E^,Ir><C_
^^01
The procedures for enabling events to fire on the next clock pulse, and for the
activation of fuzzy truths of conditions by event firings, are more complicated than in the
Boolean case. We describe them with two new definitions. For any node Nj of either kind
(condition or event), the set of all nodes from which arrows enter Nj is called the preset
of Nj, denoted Nj. The set of all nodes that are entered by arrows departing Nj is called
518
the postset N; of Nj. For example, in Figure 3, Ej = {€^,€5} is the preset of E^, Ej
{Cg} is the postset of Ej, and Cj^* = {Ej.Es} is the postset of CJQ.
nj = 0,6
e, = 0 . 6
Qj = 0.6
fg= 0.76
r:^= 0.76
The Event Ej of Figure 4(a) is enabled only if each condition in its preset Ej
contains a token that represents a nonzero truth value (a nonzero fuzzy marking). These
fiizzy values from the preset conditions are MINed to obtain a fuzzy enabling value Cj
= min{mi:Cj£*Ei} at Ej. Because an Event Ej represents an implication, it must have a
fiizzy implication value fj. The activating value propagated across the implication event
is then a^ = min{fj,ej} = min{0.65,0.6}, and this activates each of the conditions in E^'
with the fuzzy truth aj = 0.6. In Figure 4(a), the Event E^ enabling value was obtained
via Bj = min{mi,m2,m3} = min{0.6,0.7,0.8} = 0.6.
The Condition C^ of Figure 4(b) is activated by more than one event in its preset C^.
Each Event E- e C4 activates C4 with an activating value a: that provides evidence for
the truth of C4 only from Ej (contextual implication). Another Event E^ e *C4 also
activates C4 and provides its contextual implication (evidence) a,^ only from E^^. But the
existing truth (fuzzy marking) of C4 also represents evidence obtained previously from
some other origin. Then the updated activating truth, or evidence, for C4 is 34 =
max{|C4|,a^aj} = max{m4,a,j,aj}. On the clock pulse where a4 becomes the fuzzy truth of
Cj, then a4 becomes the new marking (state) m4 for C4, where j = 1 and k = 2 in Figure
4(b) that shows the computation of a new marking m4 for Condition C4, where the present
fuzzy truth of C4 is m4 = 0.7. The new marking is assigned a new fuzzy value via m4 «-
max{aj,a2,m4} = max{0.6,0.76,0.7} = 0.76.
A thresholded fuzzy Petri net uses a fuzzy activation threshold value X- at each jth
Event Ej. When Event Ej with threshold t has a fuzzy implication value f, and an
519
enabling value e,, then the enabled Event E: fires only if the activating value aj =
min{£,ej} satisfies aj s tj. Thresholds guarantee that decisions are not made based on
weak evidence.
Procedure Update_Enabiings:
FOR j = 1 to m DO /* m events to be enabled */
emin = 1.0;
FOR i = 1 to n DO /* minimize up to q preset conditions */
IF (EP[j,i] s emin) THEN emin = EP[j,i];
e[j] = emin; /* enabling value is minimum value */
Procedure Update_Activations:
FOR j = 1 To m DO
IF (fjj] < elj] THEN aU] = f|j] /* min. result with fuzzy implication */
ELSE a[j] = e[j] /* to obtain fuzzy activation */
Procedure Update_Markings:
FOR i = 1 to n DO /* n conditions to be updated */
amax = 0.0;
FOR j = 1 to m DO
IF (CP[i,j] a amax) THEN amax = CP[i,j];
IF (amax i m[i]) THEN m[i] = amax; /* maximum of old marking*/
/* and amax */
We consider here the case of supervised training of a fuzzy Petri net for control
decisions. A time trajectory of desired plant outputs must be known. Figure 5 shows a
general trainable fuzzy controller. For output variables, e.g., y and z, a given desired
output {target) trajectory over time k is denoted by {(yd(k),Zj(k))}|j, while the actual plant
outputs due to fuzzy Petri net decisions are designated by {(ya(k).ZaO'))}k- ^ ^ ^^"'^'
squared error (SSE) is then computed over time k = 1,...,K as
E = 2(k=i.K) lya(k) - yd(k)P + k(k) - zM"-
FIGURE 5 - A GENERAL TRAINABLE FUZZY RULE-BASED CONTROLLER
Target Plant
State
MMSE
I TRAINER
^ 1
• da
\
< - ^onnanc
^ FUZZY ^
"
1 EXPERT PLANT •
CONTROLLER
Plan-t
State
The process of training entails an initial computation of the SSE using the coarse
fiizzy implication values. Then each single fuzzy implication value fj in turn is perturbed
and the SSE is computed to determine whether or not the SSE decreases. If so, the new
£ is kept, else it is discarded. This continues over all events Ej, 1 s j s m. When the SSE
stops decreasing for a particular fj, then the process is repeated for the fuzzy implication
fj+j, etc. When all fuzzy implications iy,...,ij^ have been adjusted to approximately
minimize the SSE E, the training is done.
Step 0: input initial fuzzy implication vector FQ; input target trajectory {t(k)}; input initial
marking MQ; input EP(j',i]; input CP[i,j];
Step 1: input stop criteria s; input Max_Loop_Count; input fuzzy increment delta 6Q
Step 2: compute the SSE E = S^^^i K) ^(i=\^) I tj(k) - s^Ot) I ^
Step 3: REPEAT /* repeat minimization until change is small */
LoopCount = 0
If the rules are partitioned into a binary tree of disjoint cases with the rules that
involve the single fuzzy implications f as leaf nodes, then we say that the rules are
orthogonal. In this case each f need be minimized over a single turn (adjustments of
other fuzzy implications f^ (k 4= j) do not affect it). Under feedback noise, the net deals
with behavior in an average sense, and is equivalent to the minimum mean square error
methodology, invented independently by Gauss (1809) [6] and Legendre (1810) [12], and
implemented frequently nowadays (see any textbook on least mean-squares filtering, e.g.,
Haykin, 1991 [8]). The inner loop adjusts each fuzzy weight to lower the overall SSE. To
obtain near-optimal fuzzy implications, the target trajectory must be sufficiently varied
to invoke all rules so each can be adjusted. This process can be extended to also adjust
the fuzzy values given to the conditions, based on plant feedback over k = 1,...K.
- = l / 2 - V a y Sensor
V Dne-Woy Rail Line
=\ /~ 1 / 4 - V a y Sensor
-\ J Direction o f Trains
PlQtforn Q Station S
SOUTH CENTER
METRDCITY
1. Both trains are to start at the same time when the Doors_Secure,
OtherTrainReady, Emergency, and CIock_Start signals are appropriate, which causes
the Start signal to enable and activate the application of power increments until the train
reaches the 1/4-way point, provided that the Emergency signal is 0.0.
2. At each of the half-way and 3/4-way points, there are to be two sensor strips
a short distance apart for the purposes of sensing the positions and ascertaining accurate
velocities. The speed of the trains will then be adjusted up or down toward the desired
maximum speed.
3. The Emergency signal is to take two values: i) 0.0 means that there is no
emergency; and ii) 1.0 means that an emergency exists. In the first case, all applications
of power, and other functions, are enabled. An emergency automatically applies the
brakes at the full nonlocking level. An emergency disables all applications of power.
4. At the 7/8-way points, power is reduced to approximately half-power.
5. At the Approach points, power is reduced to zero level (the trains coast).
6. At the Station-Entry points, brakes are applied at 1/2 level.
7. At the Platform-Edge points, brakes are applied at full nonlocking level.
8. When stopped, all fuzzy logic values are reset for the next trip, and the clock
is started for a two-minute wait.
Figure 7 shows a fuzzy Petri net that is a simplified high level model. When the
Emergency signal at Node C is set to 0.0, it outputs a fuzzy value of 1.0 to enable power
conditions. It and Nodes B, D, and E enable the event between B and I so that power
incrementation is activated at Node I to enable a sink event. As long as the 1/4-way point
has not been reached. Node F will be 0.0, and will output a 1.0 value to enable further
increases in power. After the 1/4-way point is reached, the power increase by Node I is
524
Enei-g. /AA
CContinued in
BOUNDARY 2) Next Frane]
disabled per the event between Nodes I and J. When the half-way point is reached, L is
activated with fuzzy truth value 1.0, which makes it possible for the power to be
increased at N or decreased at Q, depending upon what fuzzy values the Speed_Too_High
or Speed_Too_Low signals take from the sensing of the speed at the 1/2-way sensing
525
Strips. The fuzzy values provided to one of these two signals determines the value to
which the power will be changed, because each is MlNed with values of 1.0 of the 1/2-
way signal.
[Continued]
Power
Dff
The fuzzy sensor values come in from the external world, while the fuzzy power level
controls are sent out to the external world when a decision is made. At the 3/4-way point,
the power level is adjusted again by a fuzzy value from Node CC, depending upon
whether the speed is too high or too low, provided that there is no emergency. At the 7/8-
way point, the speed is adjusted to half speed in the absence of the Emergency signal.
When the Approach signal is activated (and the Power signal is present from Node GG,
then the PowerOff signal is activated at Node JJ. When the Station point is reached, the
brakes are applied, at the fiizzy value 1/4, and then applied at fuzzy value 1/2 at the
Platform signal. When stopped, the logic is cleared and the two-minute clock is started.
When an enabled event fires, its enabling value is consumed, and that event must be
enable again if it is to fire again. When an event is enabled by fuzzy tokens in its preset
conditions, those tokens are consumed in the enabling. The control variables are usually
fuzzy in order for proper adjustments to be made. It is obvious that a fuzzy Petri net
model is not unique. For example, we have adjusted the power for speed change via two
different structures: i) by Nodes N or M, one of which decreases and the other which
increases. However, Node CC sets the power level through a new fuzzy value, determined
either by DD or EE, that is either a decrease or increase from what it was. The designer
may develop multiple versions before one is accepted as being advantageous in some
manner, e.g., by being simpler.
526
REFERENCES
[1] B. D'Ambrosio, Qualitative Process Theory Using Linguistic Variables, Springer-
Verlag, NY, 1990.
[2] D. Dubois, and H. Prade, The management of uncertainty in fuzzy expert systems and
some applications, in Analysis of Fuzzy Information, Vol. II, ed. by J. Bezdek, CRC
Press, Boca Raton, 1987, 39-58.
[3] J. Efstathiou, Knowledge-based systems for industrial control, Computer-Aided
Engineering Journal, 4, 7-28, 1987.
[4] R. Forsyth, Fuzzy reasoning systems, appeared in Expert Systems (Edited by
Forsyth), Chapman and Hall, London, 1984.
[5] M. L. Garg, S. I. Ahson, and P. V. Gupta, Fuzzy Petri net for knowledge
representation and reasoning. Info. Proc. Ltrs., vol. 39, no. 3, 1991, 165-171.
[6] K. F. Gauss, Theoria Motus Corporum Coelestium in Sectionibus Conicus Solem
Ambientum, Hamburg (translation: Dover, NY, 1963), 1809.
[7] F. Hayes-Roth, Rule based systems, ACM Comm. 28, 921-952, 1985.
[8] S. Haykin, Adaptive Filter Theory, Prentice-Hall, Englewood Cliffs, 1991.
[9] A. W. Holt, and F. Commoner, Events and Conditions, New York: Applied Data
Research, also in Rec. Project MAC Conf. Concurrent Systems and Parallel
Computation, New York: ACM, 1970, 1-72.
[10] Bart Kosko, Fuzzy cognitive maps, Int. J. Man-Machine Studies 24, 1986, 65-75.
[11] R. C. T. Lee, Fuzzy logic and resolution principle, J. ACM, vol. 19, no. 1, 1972,
109-119.
[12] A. M. Legendre, Methodes des mondres quarres, pour trouver le milieu le plus
probable entre les resultats de differences observations, Mem. Inst. France, 149-154,
1810.
[13] H. P. Lipp, The application of a fuzzy Petri net for controlling complex industrial
processes, IFAC Conf. Fuzzy Info. Contr., Marseille, Jul. 1983, 459-465.
[14] C. G. Looney, Expert control design with fuzzy rule matrices, Int'l J. Expert
Systems, vol. 1, No. 2, 159-168, 1988.
[15] C. G. Looney, Fuzzy Petri nets for rule based decisionmaking, IEEE Trans. Systems,
Man, and Cybernetics, Vol. 18, No. 1, 178-183, 1988.
[16] E. H. Mamdani, Applications of fuzzy algorithms for control of simple dynamic
plant, Proc. lEE, vol. 121, no. 12, 669-678, 1974.
[17] E. Mamdani, and S. Assilian, An experiment in linguistic synthesis with a fuzzy
logic controller, Int'l J. Man Machine Stud., 7, 1-13, 1975.
[18] E. Mamdani, Advances in the linguistic synthesis of fuzzy controllers, Int'l J. Man
Machine Stud., 8, 669-678, 1976.
[19] T. Murata, V. S. Subrahmanian, and T. Wakayama, A Petri net model for reasoning
in the presence of inconsistency, IEEE Trans. Knowledge and Data Engineering, vol.
3, no. 3, 1991, 281-292.
[20] J. L. Peterson, Petri Net Theory and the Modeling of Systems, Prentice-Hall,
Englewood Cliffs, 1981.
This page intentionally blank
527
IFCTR-CNR
Via Ampere 56
Milano, Italy
Abstract
This paper illustrates the use of fuzzy logic-based tools, including a hybrid system based
on neural networks and fuzzy set representation techniques, in building medical expert
systems. These tools have been employed in several medical diagnostic situations
presenting different complexities. The first set of applications concerns diagnostic
problems in the field of gynecology; the second includes biomedical image
interpretations using radiological and colposcopic images.
1. Introduction
Fuzzy set techniques are widely used in the solution of various practical problems in
medicine: the current literature provides various examples of medical expert systems
with fuzzy inference models and of medical applications using fuzzy expert system shells
[2].
A major concern of these systems has been the development of fuzzy logic based
methods for handling uncertainty which are both theoretically sound and computationally
tractable. However the solutions adopted by a medical expert system will be of little
value, unless it can guarantee the reliable acquisition of expert information. A knowledge
acquisition strategy must be provided which can actively support systems by translating
medical knowledge into a machine readable expression.
The neural network approach has received renewed attention in the last few years. This
approach offers the advantage that, unlike the rule-based approach which relies on expert
input, the knowledge base is derived directly from accumulated data [3]. But, as in
medicine both of these source of information are extremely important, it seems a good
idea to incorporate both approaches into a single comprehensive framework. Cohen and
Hudson propose the integration of neural network techniques with approximate
reasoning in a knowledge system for the treatment of carcinoma of the lung [4].
Machado and Rocha present a fuzzy connectionist expert system for diagnosing renal
syndromes [5].
This paper illustrates the use of fuzzy logic-based tools, including a hybrid system based
on neural networks and fuzzy set representation techniques, in building medical expert
systems. These tools have been employed in several medical diagnostic situations
presenting different complexities. The first set of applications concerns diagnostic
problems in the field of gynecology; the second includes biomedical image
interpretations using radiological and colposcopic images.
The fuzzy representation model interprets qualitative concepts involved in the diagnostic
problem as fuzzy sets. In particular, a clinical sign may be interpreted as a linguistic
variable and its description represented as a fuzzy declarative proposition of the form X
is A, where X is the linguistic variable representing a given sign and A is a term
belonging to the term set of the linguistic variable and denotes a fuzzy set in a given
universe of discourse f/ = (M,,....,M„) characterized by the membership function |i^(M).
The set U contains all the possible numerical or, more generally, crisp values assumed
532
In some cases experts could refer the distribution of possibility associated with a given
fuzzy label to a standard function and the system then elicits membership functions by
adjusting suitable parameters in standard piece wise functions [11]. The parameters are
cardinality measures of intuitive information about the distribution of possibility and
may be elicited directly from the experts without the risk of creating artificial
information.
This above technique has been amply applied to build a forwardchaining fuzzy rule-
based system in the domain of postmenopausal osteoporosis [11]. The application
533
has had to cope with the huge number of clinical signs involved and the lack of
standardization in defining borderline conditions. A l clinical signs and diagnostic classes
have been encoded as linguistic variables and their descriptions fo^rmalized in trams of
fuzzy declarative propositions. Diagnostic rules have been formalized and stored in terms
of a set of fuzzy conditional statements. The elements considered were 29, including
hematic and chemical parameters, mineralometric parameters, generic risk factors aad
specific risk factors. Most of the clinical signs involved in the diagnosis had terms
represented by standard functions. Figure 1 shows the standard membership functions
of the linguistic tenms used by physicians to characterize the Bone Mineral Content of
the Radius.
Figure 1 -Membership functions for tlie clinical sign bone mineral - content
Tliese functions result from an elicitation process, in which preliminary data are acquired
(in the example in figure 1, the numerical range and the type of instrument with which
the numerical values are obtained).
For each function, the experts partition the numerical range, identifying the intervals in
which they judge the compatibility of numerical values with the linguistic label
concerned high, medium or low. Parameters for the specification of standard piecewisc
functions are computed directly from this information, taking the values at the extremes
of the intervals estabhshed by the experts, and applying the "mean" operator to
average the different values obtained.
534
This procedure has been applied to elicit membership functions for 62 terms. The
physicians interacted directly with an intelligent interface, providing data and validating
results.
A fuzzy matching procedure has then applied for a quantitative evaluation of the
membership functions elicited. The degrees of compatibility between the numerical
values of clinical signs and the linguistic terms were collected firom the experts, the
corresponding degrees of membership computed using the analytical formulation of
membership functions. The comparison took into account the intrinsic variability in
experts' answers. Exact matching would not provide a reliable evaluation of the
agreement between experts and automated procedure. The problem has been addressed
by estimating, for each degree of similarity resulting from the implemented
membership functions, a range of acceptance R = [\xia)-t,[i(a) + t] where / is a value
related to the variabiUty in experts' answers [12]. They were assessed to agree if the
human degree of compatibility lies within the corresponding R
A test set of 700 samples from the universes of discourse of all the memberships (62
terms) showed an agreement of 82%. Disagreement was confined to membership
functions representing borderhne conditions.
There are complex situations in which it is difficult to predict the trend of the
membership functions. In these cases data elicited from experts serve to determine not
only representative values for parameters of known functions, but the overall
distribution of an unknown trend as well. This technique has been used in a medical
application aimed at formaUzing and automating the clinical assessment of body hair
growth in women for use of hormonal disorders. This chnical problem has usually been
approached as if there were a clear dividing line between the normal and the abnormal
state and assessment has been based on a quantitative, crisp method providing a
numerical score to represent the clinical evaluation [13]. The resulting discrepancies
among experts and unpredictable subjectivity in the diagnosis have suggested that a
fuzzy medical expert system could better guarantee an objective, expUcit and
deterministic assessment of the diagnostic process. The clinical signs involved in the
evaluation of hair growth are density, consistency and area. The difficulties encountered
in creating fuzzy sets for these signs are due principally to the global visual perception
underlying the clinical evaluation.. Visual data, in the form of scenes and pictures, are
often processed in visual terms alone, without any corresponding translation or
recording as verbal labels or representations [14]. This was a crucial aspect in the
elicitation of fuzzy sets for the area sign. Experts were unable to qualify linguistically
it and required of contextual information to provide reliable evaluations. We addressed
the problem by implementing a graphical interface displaying all the elements of the
body as parts of an integrated whole, and not as separate and independent items. Figure
2 shows the interface window for the acquisition of membership functions related to the
sign area evaluated in the chin site.
535
firade
nwr
m
YES 2
J. jpm-
_-J HO 1
NO 4
NO 5
NO S
» S
The global region of ioterest was divided into 5 items. Progressive combinations of
these items were displayed: they constituted the range on which the membersMp
functions for the linguistic terms significant, borderline, unsignificant are defined.
Because it was impossible to procure explicit descriptions of the sign area, these terms
were semantically related to a meta^diagnosis in which, together with the area,
consistency and density parameters are evaluated by the experts. Following the
membership function exemplification method [15], the experts were asked to express the
compatibility of each term with each combination of items by answering yes or no and
assigning a number from 0 to 5 to kidicatc their degree of confidence in the answer.
The collected data were then processed to produce the representative value |i,(x)of the
membership function for the term t and the combination of area items x.. The following
formula were used:
\i,{x) = - + dx(—)
2 10
where d-l if the answer was yes and d-Q if the answer was no. The degrees n,(x),
quantifying the compatibihty between the linguistic term t and the crisp value assumed by
the cliflical sign concerned for each expert, were aggregated. Several measures were
536
used to group the different answers; among these were the mean, median, expected value
and the fuzzy expected value. Of all these, only the fuzzy expected value seemed to
give consistent results: extreme values did not greatly affect it, as they did the other
measures, and, despite the great variability in answers, it provided a measure indicative of
some sort of a central tendency [16],
Figure 3 shows membeiship functions for the sign area obtained by applying a spline-
based fitting operation on the representative values computed with the fuzzy
expected value.
The fuzzy matching procedure was applied to compare membership functions values
and expert judgements. Global agreement, of 75%, evaluated on a training set of 50
sample values, was found. A partial agreement of 90% was found for clinical signs for
which the fuzzy consistency was fully respected.
Knowledge-based systems for biomedical appMcations are often called upon to formalize
and store strongly structured descriptive knowledge regarding medical descriptions of the
anatomical parts concerned, feature descriptions and static relations among objects. In
particular, when dealing with biomedical image interpretations, the underlying knowledge
is characterized by a large number of items and complex hierarchical relations among
them.
537
a-priori anatomical knowledge about intrinsic properties of human organs and image-
processing knowledge about properties of the image involved in the application. The
fuzzy frames are used as a generalization of the knowledge structures introduced in AI
with the aim of representing in a unified framework multi-typed objects characterized
by multiple uncertainty. The frames represent objects in terms of a set of attributes
which may be value-attributes, specifying by means of assignment or procedure
computation the value or range of values the attribute of a given object should have,
or relational attributes denoting links with other objects. In real applications, where
attributes are very often fuzzy in nature, independent levels of uncertainty may be
integrated in the frame structure to represent the degree of certainty with which a value
is assigned to a given object attribute, or to define gradual multivalued predicate links
among objects. Several attempts to formalize the concept of fuzzy frames are described
in the literature [17]. We have focused our attention on fuzzy frame representation for
biomedical image recognition purposes, introducing two kinds of frame: model frames,
representing models of anatomical parts and object frames, representing the description
of specific objects in the current image and created when an object is recognized as
an istance of a given anatomical model. Frames representing models are predefmed on
the basis of the knowledge provided by the experts and describe classes of objects. In
each model-frame, own-slots list characterizing attributes of the class, and member slots
describe those attributes that characteriaze an element of the class and assume a
significant role in the development of the interpretation process of the final image. Object
frames are instances of model frames and contain a member-of link to the class to
which the object represented by the frames belongs. Slots of the object frames are own
slots and represent the attributes of the object; these slots are coincident with the
member-slots of the corresponding model frame. In order to deal with situations in which
knowledge related to a given object is qualitatively or intrinsically imprecise, fuzzy slots
are introduced in model and object frames, and a fuzzy inheritance mechanism is defined
to represent partial belongingness of an object to a class represented by a model frame.
4.1 The use of fuzzy frames for the representation of knowledge in the analysis of
CT images of the vertebrae
Object frames are related to those objects in the current images that are recognized or
created in accordance with the goals set. In particular, object frames for cortical bone
and spongy region are created as a result of the recognition procedure. An example
of the model and object frames involved in the analysis of CT images of Vertebrae is
given in table 1. The o&jert/rame SP-RE is a fuzzy frame in two respects. First, the
degree of property inheritance from the model frame in the member-of slot may be
specified as a number between [0,1] (in the case of no values specified, the default
value 1 is assumed). Second, the other slots may contain fuzzy sets as values.
The degree of membership specified in the member-of slot documents the results of the
fuzzy reasoning process underlying the frame structure and dedicated to the
recognition task. The reasoning activity for recognizing as Spongy Region a candidate
Obj-i , resulting from a previous segmentation procedure, is based on the evaluation of
the degree of matching between attributes in own slots of model frame of Spongy
Region and attributes of the candidate object. The matching evaluation impMes the
execution of procedures contained in the own slots of Spongy Region model frames.
The procedures Below and Surrounded require that object frames for Spinal Canal and
Cortical Bone respectively be already recognized. A metarule in the Knowledge-Base
subordinates the activation of the recognition rule to this condition. Spinal Canal is
recognized by a thresholding procedure, and Cortical Bone is recognized as the region
539
expresses the discriminant condition that the shape of the spongy region is circular,
conceiving as circular a fuzzy concept with an intrinsic variability in its definition.
Difficulties may arise in ehciting the overall set of diagnostic rules directiy from the
experts of a given medical domain. The interview technique, although it may be more
successful than other methods in revealing the detailed structure of a concept, presents
the risk that the knowledge elicited may be an artificial creation produced under the
pressure of questioning. Leaming-by-example strategies offer domain experts tools
for the direct transfer of their specific, exemplified diagnostic knowledge and provide
inductive algorithms to induce knowledge in a general abstract form. Various methods for
generating fuzzy production rules or, fuzzy relations in general, have been proposed in the
540
literature [19]. Our approach focuses on an empirical learning strategy which provides a
reasonable explanation and monitoring of intermediate results and is consistent with the
possibilistic, rather than probabilistic, nature of uncertainty [20].
In the cervicography images, the scene observed is the uterine cervix with its different
types of epithelial tissues. The region of interest is the squamocolumnar junction,
that is, the transformation zone in which possible alteration processes are located. This
is the locus where metaplastic or neoplastic lesions generally appear.
Figure 5 presents details of cervocographic images showing the region of interest of
several subjects.
541
In his visual inspection , the expert bases his judgement on botli variations in the color of
tlie different tissues, and on featurc-s of the sqoamocoliimnar junction .
Several types of lesions can be differentiated by the expert. We restrict our analysis to
three classes, to reduce the complexity and in view of the fact that even this first,
simplified level of classification may have a significant and immediate impact in the
screening activity. Tlie classes identified are:
• ECTOPIA, grouping normal subjects, which present under colposcopy only normal
malpighian and colnmnar epithelia and no abnormal lesions. The junction zone
has a sharp and imegular boundary and the color varies from the internal to the
external region, from dark to light.
• NTZ, where an immature squamus metaplasia is present. The junction is fuzzy and
dark to light.
• ATZ, identifying subjects for which colposcopically diiBcted biopsy is prescribed.
The junction is well defined and regular and the color difference is from light to dark.
Classification is achieved with a complex interpretation strategy which consists mainly in
selecting and performing those ¥isual procedures which first focus analysis on
significant regions in the scene and subsequently allow the extraction of discriminant
features. For the sake of brevity, only classification tasks are described here. A
542
In this application, classes were assigned on the basis on the selected features: sharpness,
morphology, color tranmsition . In the fuzzy classification scheme each feature is
considered a linguistic variable with the following associated term sets:
• Sharpness = ( well defined, medium, fuzzy)
• Morphology = ( regular, medium, irregular)
• Color Transition = ( light-dark, dark-light)
Terms in the above term sets are fuzzy sets having as their universe of discourse the
numerical ranges of the corresponding features , provided by the described above
measuring procedures.
The classification rules were modeled in terms of fuzzy production rules having the
following general form [8]:
t'=lu,J^]u'eT'
where u, is the vector of measurements obtained in correspondence with the selected
set of features, ;* is the expert's judgement expressing the strength or degree of
satisfaction with which the image structure represented by u, may be assigned to class
The experts use predefined linguistic labels, such as very high, medium, low, very low to
express their mental judgements; these labels are automatically translated into numbers in
the interval [0,1].
Given the diagnostic class D^, let:
• /, be the cardinality of the set of all the possible antecedents ( -4, )
• m, the cardinality of the training set T^
• p, the cardinality of the term set W containing the terms with which the strength of
implication may be expressed
\^A>^) ^i.,(«J
A/r = M,
H^,(«l) \^A,(UJ ^i^.0'^) ^.,0•»)
G = A/*oM*
to obtain matrix G with dimension Ixm. The element g^ expresses the induced degree of
certainty with which the corresponding rule may be generated.
This learning procedure has been applied to the classifcation of cervicographic images
using a training set of 120 images. Twenty-six classification rules have been generated.
When these were applied on a test set of 50 subjects, they gave an accuracy of 79%.
Table 2 shows a mle describing a typical case of ATZ. Figure 6 shows the main phases
of the overall image interpretation procedure, including segmentation, contour detection
and classification, applied to three different subjects.
544
'colposcopic
The application has been introduced in section 3.1.2, detailing the strategy adopted to
elicit and representing the fuzzy sets characterizing the linguistic descriptions of visual
clinical sign concerned.
In this application, a fuzzy neural franiework has been designed and implemented to
model diagnostic relationships between clinical signs and diagnostic judgements.
The clinical assessment of body hair growth in women used in the study of hormonal
disorders, proceeds by evaluating clinical signs, such as area, density and consistency in
9 different body sites.
A metadiagnosis is made for each body sites and partial diagnostic judgements are then
aggregated to perform the final diagnosis.
The overall diagnostic process has been implemented by an hybrid intelligent system
which uses fuzzy trained NN and fuzzy production rules in a unified framework.
Thus each sub-diagnosis (of a body part) is characterized by a set of input events, a set
of output hypotheses, and a NN implementing a fuzzy relationship between input and
output.
The NN assumes a feed-forward topology with three layers. Each input node represents
the value of an attribute of the object (or event) to be classified; in our approach each
term linguistically qualifying a clinical sign is represented by a different neuron .
The degree of confidence in each of these terms (obtained by evaluating its membership)
is represented by the state of activation of the neuron. Hidden neurons represent
intermediate abstractions.
The output layer represents the hypothesis. In particular the state of activation of each
output neuron represents the degree of possibility of the event under analysis to the
hypothesis represented by the neuron.
In our application there is one diagnostic class, corresponding to the clinical assessement
of body hair growth in women; the different hypotheses are linguistically qualified
degrees of severity, ranging from very low to very high.
The sigmoid function is used to determine the output state. The output value V, from the
i-th node is:
1
V.=-
l + e '"
where m is the number of nodes contributing and V. is the output value of the j-th node.
Back-propagation, which is also known as the generalized delta rule, is one of the most
popular and widely investigated learning methods used to train feddforward neural
networks.
The back-propagation learning algorithm is based on nonlinear optimization technique of
gradient descent on the sum of the squared differences between the activation of the
nodes in the output layer and the desired output. Application of the gradient descent
method yields iterative weight update rules [24].
The application of neural network learning mechanisms presents some limitations, despite
their ability to leam in an uncertain or unknown environment.
A comparison of experimental results in classification application shows advantages, but
also disadvantages, with respect to other methods employing knowledge-based
approaches [25].
The main drawback of the NN method lies in the fact that the generalized delta rule is
computationally complex: when the training data are numerous the learning time can be
very long. Moreover, the generalized delta rule can not guarentee that the global error
minimum will be found, and the solution strategy depends on the initial network
configuration (often assigned at random).
547
Several attempts have been made to improve the performance of back propagation
algorithms; fast learning algorithms heve been proposed to accelerate the training phase
[26].
Our approach integrates problem-dependent knowledge in the analytical formulation of
the back propagation algorithm in order to control and subsequently improve the weight
update procedure.
TTie central concept is that of subordinating the activation level of one or more hidden
neurons to the presentation of certain input pattern configurations to input layer.
The fuzzy logic framework is used to support the representation of this knowledge, which
may be uncertain and/or linguistically qualified. The flexibility of the fuzzy reasoning
approach makes it easy to define and update rules until a reasonable performance Ls
achieved, and to introduce new intuitive knowledge at different levels of abstraction in a
natural and controlled way.
Fuzzy production rules, structured as evaluation-decision pairs, are used as knoweldge
structures to represent all the ingredients of the fuzzy reasoning process. In the
antecedent of the rules, conditions related to input patterns are expressed in terms of
compound fuzzy declarative proposistions. The decision part of the rules contains fuzzy
declarative propositions which describe linguistically the values that variables related to
hidden neuron activation assume when the corresponding evaluation part of the rules is
satisfied.
We have defined a forward chaining fuzzy inference mechanism to interpret these rules: it
assigns a degree of confidence to each rule in the knowledge base, using a fuzzy matching
procedure, and deduces the best hidden neuron activation through a defuzzification
process applied to the variables in the consequent part of rules.
We use the fuzzy logic based learning algorithm to train the networks devoted to each
diagnostic subproblem involved in the medical application concerned.
At the present stage of our experiments, preliminary versions of the NN classifiers have
been implemented for each of the 11 body sites examined.
Figure 7 shows the topology of the neural network used. The number of input nodes is
determined by the dimension of the input patterns, which are degrees of evidence
specified for each term of the clinical signs.
Input neuron activation lies to the range [0,1]. Each output neuron represents a degree of
possibility of assigning the current event to a certain degree of severity and also has
activations in the range [0,1].
548
o
density consistency
Each of the NNs implemented have been trained using training sets constituted by
diagnosed clinical cases.
A total number of 400 examples have been considered.
The fuzzy logic-based NN has been implemented using different knowledge bases in
function of the site.
For the chin for example, 1 hidden neuron has been constrained to the presentation to the
input layer of evidence confuming a low degree of severity.
Table 2 lists a subset of rules in the knowledge base for the chin site.
Suitable membership functions have been elicited to characterize linguistic terms in the
rules.
549
6.5.Preliminary results
To assess the diagnostic accuracy of the implemented neural networks and their
classification capability, a test set of examples, consisting of already diagnosed clinical
cases referred to specific body sites, has been used. The testing data considered were 200.
The diagnosis was confirmed in 82% of the cases.
6.6.Work in progress
We are currently engaged in experimental work to conclude the application and the
model the overall diagnostic process.
The cognitive problem underlying this application has been broken down into a hierarchy
of classification tasks:
• a procedure using an higher level NN; in this case the outputs of the first level are the
input for the higer NN;
• a procedure using a fuzzy rule-based system; in this case the symbolic learning
procedure presented in previous section is adopted to the problem under
investigation.
The two solutions present unique advantages and disadvantages and further
experimentation has been designed to exploit and evaluate each.
References
1. Szolovits P., Patil R.S., Schwartz W.B., Artificial Intelligence in medical diagnosis.
Annals of internal medicine, 108, pp.80-87, (1988).
2. Adlassnig K.P., Kolarz G., SheithauerW., Present State of the Medical Expert System
Cadiag-2, Meth. Inform. Med., 24, 1, pp.13-20, (1985).
3. Kandel A., Langholz G. (Eds.),"Hybrid Architectures for Intelligent Systems", CRC
Press, Boca Raton (Florida), 1992.
4. Hudson D.L., Cohen M.E., Banda P.W., Blois M.S., Medical Diagnosis and
Treatment Plans Derived from a Hybrid Expert System. In "Hybrid Architecture for
Intelligent Systems", (Kandel A., Langholz G. Eds.), CRC Press, Boca Raton
(Florida), 1992.
550
5. Machado R.J., Rocha A.F., A hybrid Architecture for a Fuzzy Connectionist Expert
System. In "Hybrid Architecture for Intelligent Systems", (Kandel A., Langholz G.
Eds.), CRC Press, Boca Raton (Florida), 1992.
6. Zadeh L.A., Fuzzy Sets, Information and Control, 8, pp.338-353, 1965.
7. Zadeh L.A., The concept of Linguistic variable and its Application to Approaximate
Reasoning. In "Fuzzy Sets and Applications", (Yager R.R., Ovchinnikov S., Tong
R.M., Nguyen H.T. Eds.), pp. 193-329. John Wiley & Sons, 1987.
8. Zadeh L.A., PRUF - a Meaning Representation Language for natural Languages. In
"Fuzzy Reasoning and its Applications", (Mamdani E.H. and Gaines B.R. Eds.),
Academic Press, 1981.
9. Hall L., Szabo S., Kandel A., On the Derivation of Memberships for Fuzzy Sets in
Expert Systems, Information Science, 40, pp.39-52, 1986.
10. Civanlar M.R. and Trussel H.J., "Constructing Membership Functions using Statistical
Data", Fuzzy Sets and Systems, 18, pp 1-13, 1986.
11. Binaghi E., A fuzzy Logic Inference Model for a Rule-based System in Medical
Diagnosis, Expert System, 7, 3, pp. 134-141, (1990).
12. Binaghi E., Delia Ventura A., Rampini A., Schettini R., A fuzzy Reasoning Approach
to Similarity Evaluation in Image Analysis, Int. J. of Intelligent Systems, 1993, in
press.
13. Ferriman D.M., Gallwey J.D., Clinical Assessment of Body hair Growth in Women,
J. Clin. Endocr. Metab., 21, 1440-1447, 1961.
14. Freedman J., Haber R.N., One Reason Why We Rarely Forget a Face, Bull. Psychon.
Soc.,3,pp.l07-109, 1974.
15. Chameau J., Santamarina J., Membership Functions I: Comparing Methods of
measurements, //;/. J. of Approaximate Reasoning, pp.288-317, 1987.
16. Pal S.K., Fuzzy Tools for the Management of Uncertainty in Pattern Recognition,
Image Analysis Vision and Expert Systems, Int.J. System Science, 22, 3, pp.511-
549, 1991.
17. Yager R. R., Linguistic representation of default values in Frames, IEEE Tran. on
Systems Man and Cybernetics, SMC-14, No.4, 1984.
18. Binaghi E., Delia ventura A., Rampini A., Schettini R., A Fuzzy Knowledge-Based
System for Biomedical Image Interpretation. In "Uncertainty in Knowledge Bases",
(Bouchon-Meunier B., Yager R.R. Eds.), Springer-Verlag, Beriin, 1991.
19. Delgado M. Gonzalez A., The frequency of Fuzzy Domains and its Application to the
System Identification, in Proceedings of the 2nd International Conference on Fuzzy
logic and neural networks, Ilzuka (Japan), 1992.
20. Binaghi E., Empirical Learning for Fuzzy Knowledge Acquisition, in Proceedings of
2nd International Conference on Fuzzy Logic and Neural networks, Ilzuka (Japan),
1992.
21. Darnell Jones D.E., Creasman W.T., Dombroski R.A., Lentz S.S., Waeltz J.L.,
Evaluation of the atypical Pap Smear, Am. J. Obstet. Gynecol., pp. 157-544, 1987.
22. Schettini R., Low-Level Segmentation of Complex Color Images. In "Signal
Processing VI: Theories and Applications", (Vandewalle J., Boite R., Moonen M.,
Oosteriinck Eds.), pp.535-538, Elsevier Science Publishers, 1992.
This page intentionally blank
551
23. Binaghi E., Mazzetti A., Orlando R., Rampini A,, Integration of Fuzzy Reasoning
Techniques in the Error Back-Propagation Learning Algorithm, in Proceedings of
6th Italian Workshop on Neural Nets, Vietri sul Mare (Salerno), 1993.
24. Rumelhart D.E., Hinton G.E., Williams R.J., Learning Internal Representation by
Error Propagation. In "Parallel Distributed Processing", (Rumelhart D.E., Mc
Clelland J.L. Eds.), MIT Press, Cambridge MA, 1986.
25. Binaghi E., Rampini A., Fuzzy Decision-Making in Classification of Multisource
Remote Sensing Data, Optical Engineering, 1993, in press.
26. Vogl T.P., Mangis J.K., Rigler A.K., Zink W.T., Alkon D.L., Accelerating the
convergence of the back-propagation method, Biological Cybernetics, 59, pp.257-
263, 1988.
CHAPTER 21
1.INTRODUCTION
After many years of research, the "repetitive movement robot" seems to have reached its
peak of evolution and has become nowadays a routine in the product line. Recently,
research was directed towards the development of machines that are able to perform more
complicated tasks. Examples are robots equipped with vision, built-in intelligence, double-
arms and so on [1,2]. In this paper an attempt is made to create a higher level robot
equipped with vision and artificial intelligence. An interesting application of artificial
intelligence in robotics can be found in the Hirota - Aral - Hachisu project [5] where a
fuzzy algorithm of twenty four rules is used to estimate the distance that the gripper must
traverse according to the speed and the present distance between the gripper and the object
on a belt conveyor. Fuzzy logic theory was first introduced by Zadeh [4] and suggests a
mathematical formulation for the linguistic sets and human reasoning. In this way the
human style of reasoning can be simulated by a computer. The fuzzy logic has been
successfully applied to industrial and chemical process control [6,8] and recently attempts
have been made to apply its methods to the control of robot arms.
In this paper, by "non destructive collection of fruits" we mean the process of collecting a
fruit without destroying the actual tree. Two main issues that must be considered in any
attempt to create a machine capable of non-destructive collection of fruits are the following:
a) The fruits of most types of trees, and particularly the apple tree, are found in
bunches. The fruits of a certain bunch are connected to some point of the tree and thus, as
a result of this, each fruit of a certain bunch must be detached towards the side at which it
is connected with a bunch. The robot, through the picture it receives, must determine the
side on which it must be detached.
b) It is possible that a fruit of the bunch may not be directly visible to the robot, as its
line of sight may be blocked from another fruit of the same bunch. This must be taken
under consideration.
From the above it is seen that the task of non-destructive collection, demands a robot
with vision (in order to recognise the fruits) as well as with artificial intelligence (in order
to infer a direction in which each fruit must be grabbed and detached). The fact that a fruit
must be detached from a certain side and not in a specific direction, gives an error tolerance
which is very helpful as far as the final solution of the problem is concerned. The direction
553
S.G. Tzafestas and A./^. Venetsanopoulos (eds.)
Fuzzy Reasoning in Informalion, Decision and Control Systems, 553-561.
© 1994 Kluwer Academic Publishers. Printed in the Netherlands.
554
of approach and detachment of each fruit is inferred through a fuzzy algorithm giving an
approximate result. The advantages of this method are:
a) It gives an approximate result by simulating the subconscious human way of
thinking.
b) It successfully combines the vague information taken from a picture with the error
tolerance result.
c) It gives a quick response, since, in discrete analysis, the results of all input
combinations can be pre-calculated and stored as a matrix.
2.IMAGE PROCESSING
The purpose of the image processing unit is to identify the fruits in the picture taken from
the camera. Specifically the position, the size, and the boundary of each fruit that appears
in the picture is determined.
It was found that image analysis of 64x64 pixels with four bits (16 grey levels) was
adequate for our experiment. A powerful source of light is used from the point of
observance in order to eliminate shadows on the fruits that may be created by natural light,
and to make the system able to function during the night. The location of the fruits is
determined in two stages using a modification of the region growing method [3,7].
Initially, certain "seed areas" are chosen (using appropriate criteria) which in turn are
enlarged step-by-step. At each step, if a pixel neighbours an area and fulfils certain
requirements, it is embodied and the procedure is then repeated. In this case the image
processing unit locates the areas of high density. This is because, due to the way of
lighting, the intensity is high at the central area of the fruits, which gradually decreases
towards the boundary surface. After this, at every stage of enlargement, the neighbouring
pixels of each area are checked. If the intensity of a certain pixel is less or equal to the
intensity of the neighbouring border pixel, the pixel is embodied in the main area. This
process is repeated until the enlargement of the seed area is completed.
It is noted that with this procedure, all areas of high intensity are located (not only the
fruits), i.e. parts of the sky (during the day) as well as the reflection of light on leaves.
Based on the smooth changes of intensity of the areas which represent the fmits in contrast
with a sudden change of other areas, the program distinguishes which of the located areas
are actually fruits and which are not.
i=l
where E stands for union.
For example if U = 0+1+2+..+9 then a fuzzy subset could be ;
A = small digits = 1|0 + 1|1 + 1|2 + 0.7|3 + 0.3|4 + 0|5 + ... + 0|9
ii) The intersection of the fuzzy subsets A and B is denoted by A^B with a
membership function defined by
iii) The complement of the fuzzy subset A is denoted by -,A with a membership
function defined by:
The relation R is a fuzzy subset of the cross product U X V with a membership function
defined by [6]
Given a fuzzy relation R and a new fuzzy subset A' (the antencent) the consequent B is
inferred from the compositional rule of inference [5]:
B' = A' 0 R
556
Finally two or more rules R; can be combined using the connective ELSE to give a fuzzy
algorithm R which describes the relation between linguistic variables leading to an
approximate result. The algorithm R is given by the relation:
R = R l u R2 U R 3 u . . u R j
3.3. Defuzzification
The result which is given by the a fuzzy algorithm is a fuzzy subset B'. Various
procedures may be used to select a certain value for the output. Here, the element with the
maximum membership value is chosen. In case there are more than one elements with
maximum membership value in B' the average of the maximum and minimum of these
elements is chosen.
considered as the section connecting the centre of the fruit's image (Kx,Ky)
with the considered point of connection. Thus, the angle q) is determined by :
Ky
tf= arctan
Kx
If the centre of the fruit's image is to the right of the point of connection, then the solution
in the interval (-n/2, n/2) is accepted, otherwise the solution is sought in the
interval (jt/2, 3jt/2).
The angle 8 is determined through the fuzzy algorithm. The input variables here are the
following:
a) The visible percentage of the fruit from the observing point.
b) The distance of the images centre from the point of connection with the branch.
For example if the fruit is fully visible and the point of connection lies behind it then the
approach should be attempted from the front side (6=0 ). On the other hand if the point of
connection lies towards the fruits outermost points then the approach should be attempted
from the side (6=90 ). The entire fuzzy algorithm consists of seven rules.
The Universe of Discourse of the variable "Area" (A) is:
U = 1 + 2 + ... + 10
where the value 10 specifies a fully visible fruit. Three labels were chosen: SmalUS),
Medium (M) and Large(L). The appropriate sub-sets are given in Table 1.
V = 0 + 1 + 2 + ... + 10
The "Distance" may be: Short(S), Medium(M), Long(L). The support subsets are
shown in table 2.
The variable "Angle" (8) is quantised into 13 levels ranging from 0 up to 180° (every
15°). The Universe of Discourse is
W = 0 + 1 + 2 + ... + 12
Seven normalised labels where chosen: ©1 (front) 92, 0 3 , 0 4 (sidewise), 05, 06,
0 7 (Back). The corresponding support subsets are given in table 3.
Table 3: Support sets for the labels of "Angle"
e 0 1 2 3 4 5 6 7 8 9 10 11 12
»i 10 7 3 0 0 0 0 0 0 0 0 0 0
02 3 7 10 7 3 0 0 0 0 0 0 0 0
®3 0 0 3 7 10 7 3 0 0 0 0 0 0
©4 0 0 0 0 3 7 10 7 3 0 0 0 0
es 0 0 0 0 0 0 3 7 10 7 3 0 0
©6 0 0 0 0 0 0 0 0 3 7 10 7 3
07 0 0 0 0 0 0 0 0 0 0 3 7 10
q)=270O , 8=450
taking into account the possibility that the second fruit may be hidden behind the visible
fruit.
559
5.RESULTS
The values of 9 for all combinaiions of the input variables are given in the next look-up
table:
^'A 1 2 3 4 5 6 7 8 9 1 0
0 173 165 135 120 120 120 75 8 8 8
1 173 165 135 120 120 120 75 8 8 8
2 165 165 135 120 120 120 75 8 8 8
3 143 143 135 120 120 120 75 38 38 38
4 135 135 128 120 120 120 83 45 45 45
5 135 135 128 120 120 120 83 45 45 45
6 135 135 128 120 120 120 83 45 45 45
7 128 128 128 120 120 120 83 68 68 68
8 120 120 120 120 120 120 105 90 90 90
9 120 120 120 120 120 120 105 90 90 90
10 120 120 120 120 120 120 105 90 90 90
Two examples of the program's execution are given bellow. The first picture (a) is the
image taken from the camera, the second (b) depicts the values of intensity of every pixel,
the third (c) gives the "seed areas" and the fourth (d) providesthe areas which have been
located as fruits. Values of intensity greater than nine have been replaced with the value of
nine for reasons that are connected to the representation of the picture. For the same
reason, in pictures (c) and (d) intensities of zero value are not typed.
1) First example:
(a) (b)
(c) (d)
Fig.2: First Example
560
The program locates a single fruit on the left side and a branch consisting of three fruits on
the right.-
Results; Single fruits: f=27l* , 8=45®
Bunch; Fruit No 1: f=-f« , §=90®
Fruit Mo 2; f=189« , i=l§5<»
Fruit No 3: f^-SS® , i=12iO
2) Second example:
la} (b)
. " j i - J , - ----J - • ] .
ic) (di
Mg.?- Secotid E.«iump!e
The program locaies live iruiis in two bunches;
Tlie upper bunch c^.n.si^ts of ibrcc frjits and the lower one consists of two fiiiits:
Results:
Upper hunch- first fruit. ip^fo , i=f§o
second fruit : f=172'' , e=fi*
third fruit: f=2630 , i=12§o
It is noted, that after the removal of a fruit from a bunch, the positions of the other fruits
change and the procedure must be repeated.
6.CONCLUSIONS
This paper, belongs to the area of designing robots that function in the natural
environment. Such robots are equipped with vision and artificial intelligence. The
problems that are faced in the attempt of locating the fruits are many, since in each case the
position of the fruits in the picture are random and the background different. Through the
study of ways to surpass these problems, useful conclusions of the importance of
characteristics, such as colour, shape and intensity, in order to locate the fruits, are drawn.
The principal issue of the present paper is the fuzzy algorithm. The central problem in
constructing this algorithm was the description in linguistic terms of the way the human
infers in which direction the fruit must be detached from the bunch. Since this problem
was surpassed the fuzzy algorithm was easy to construct. This is actually the main
advantage of fuzzy logic, i.e. from the moment the control strategy can be expressed
linguistically as a set of fuzzy conditional statements, the fuzzy logic system is readily
synthesized.
Here, the fuzzy algorithm was shown to be of a greater success than expected,
encouraging future studies of the fuzzy logic in image processing too. In conclusion fuzzy
logic theory proves to be a helpful tool in the human attempt to create robots with "human"
abilities such as fruit's collection, etc..
REFERENCES
[1] S.G.Tzafestas .: Expert Systems in Engineering Applications, Springer -
Verlag, Berlin, 1993.
[2] S.G. Tzafestas.: Intelligent Robotic Systems, Marcel Dekker, New York, 1991.
[3] R.G Gonzalez and R.E. Woods. : Digital Image Processing, Addison Wesley,
pp.574-5,I992.
[4] LA. Zadeh: Outline of a new approach to the analysis of complex systems
and decision processes. IEEE Trans. Syst., Man, Cybern.,vol.SMC-3,no.l,pp28-
44,Jan.l973.
[5] K.Hlrota, Y.Arai and S.Hachisu: Moving mark recognition and moving
object manipulation in fuzzy controlled robot. Control: Theory and Adv.
Techn.,vol 2, no.3, pp 399-418, Nov.1986.
[6] E.H.Mamdani: Application of fuzzy logic to approximate reasoning using
linguistic synthesis. IEEE Trans. Computers, vol c-26,no. 12,Dec.l977.
[7] K.S.Fu, R.C.Gonzalez, C.S.G.Lee: Robotics: Control, Sensing, Vision and
intelligence. McGraw-Hill 1987.
[8] R.M.Tong: A control engineering review of fuzzy systems.
Automatica,vol.l3,pp.559-569.1977.