Вы находитесь на странице: 1из 11

Int J Adv Manuf Technol (2011) 52:821–831

DOI 10.1007/s00170-010-2742-z

ORIGINAL ARTICLE

A hybrid chaotic quantum evolutionary algorithm for resource


combinatorial optimization in manufacturing grid system
Haijun Zhang & Yefa Hu

Received: 5 January 2010 / Accepted: 18 May 2010 / Published online: 21 September 2010
# Springer-Verlag London Limited 2010

Abstract Resource composition in manufacturing grid employed to work collaboratively toward an identical goal
(MGrid) system is one of the recent critical issues of MGrid of dealing with distributed manufacturing tasks.
researches. Especially, resource combinatorial optimization In the MGrid environment, on the one hand, a
(RCO) becomes more challenging when multiple optimal considerable number of enterprises would like to publish
criteria are considered in MGrid system. Based on the their resources to make profits; on the other hand, in many
quantum evolution theory, we propose a hybrid chaotic cases, a manufacturing task consists of multiple subtasks
quantum evolutionary algorithm (CQEA) for RCO problems. which run on multiple MGrid resource nodes. A single
We also propose a novel resource encoding method for resource may not execute all processes. Thus, the ability to
CQEA, which is dynamic and flexible. The experimental select and integrate various resources in MGrid system
results show that the proposed CQEA is effective, efficient, efficiently and effectively at runtime is important to the
and scalable for the RCO problem in MGrid system. development of MGrid applications. The selecting and
integrating resource is called resource combinatorial opti-
Keywords Manufacturing grid . Quantum evolutionary mization (RCO) in the paper. In the literatures, existing
algorithm . Chaotic search . Combinatorial optimization . researches on MGrid primarily concentrate on the concept
Resource encoding and the architecture of MGrid [3], resource modeling [4],
quality of service (QoS) [5], resource scheduling [6], and
co-reservation [7]. The paper focuses on the RCO problem
1 Introduction in MGrid system.
The RCO belongs to the class of NP-complete problems
Manufacturing grid (MGrid) [1, 2] is a kind of advanced [8] and meanwhile often requires multiple criteria to be
manufacturing technology, which has been extensively considered. However, in a multi-dimensional parameter
researched in recent years. Through the network, MGrid space, it is in general hard to find a solution that is “best”
can connect some geographically distributed enterprises with respect to all the metrics at the same time [9]. Many
and resources and then form virtual organizations (VOs) researchers have proposed various algorithms and methods
that are centralized in logic but distributed in physics. In for RCO problems. T’kindt and Billaut extend the
VOs, all resources can be shared, and all enterprises can be definition of optimality to Pareto optimality to find all
optimal solutions [10]. Yu et al. present a genetic algorithm
(GA) approach to address scheduling optimization problems
H. Zhang (*) : Y. Hu
in grid workflow applications, based on two QoS constraints
School of Mechanical and Electronic Engineering,
Wuhan University of Technology, (deadline and budget) [11]. Masutti et al. propose a self-
Wuhan 430070, China organizing neural network using ideas from the immune
e-mail: zhhaijun@umich.edu system to solve the combinatorial optimization problem,
H. Zhang : Y. Hu
traveling salesman problem, but the algorithm demands a
Hubei Digital Manufacturing Key Laboratory, greater time for convergence, compared with other neural
Wuhan 430070, China methods [12]. Shu et al. apply simulated annealing to
822 Int J Adv Manuf Technol (2011) 52:821–831

quantum-inspired GA (QGA) in order to avoid the premature be “0” and “1,” respectively. A Q-bit may be in the 0 state, in
convergence of QGA on the knapsack problem [13]. Tao et the 1 state,or in a liner superposition  of the two.
al. propose a MGrid resource service composition and QðtÞ ¼ qt1 ; qt2 ; qt3 ;   ; qtn is a population of Q-bit
optimal-selection method, based on the principles of particle individuals at generation t, where n is the size of
swarm optimization, and describe four different task models population, and qtj is a Q-bit individual defined as:
(also called workflow structures) for the algorithm [8]. " t  t  t   t #
a a a  a
In recent years, the quantum evolutionary algorithm j1  j2  j3   jm
qtj ¼  
t  t  t   . . . . . .  t ð1Þ
(QEA) has gradually become more popular and has been bj1 bj2 bj3  bjm
successfully applied to solve various combinatorial optimi-
zation problems. The QEA is inspired by the concepts and where
 j=1, 2, 3,⋅⋅⋅, n, m is the number of Q-bits and
 t 2  t 2
the principles of quantum computing [14], such as the Q-bit aji  þ bji  ¼ 1, (i=1, 2, 3..., m).
 
and the superposition of states. In 2002, Han and Kim PðtÞ ¼ X1t ; X2t ; X3t ;    ; Xnt is a population
 of binary
introduced a quantum-inspired evolutionary algorithm to solutions, where one binary solution, Xjt ¼ xtj1 ; xtj2 ; xtj3 ;    ;
solve a class of combinatorial problems [15]. Xiao et al. xtjm Þ, can be obtained by observing the state of qtj . For each
present a quantum chaotic swarm evolutionary algorithm to xtji (i=1, 2, 3..., m) in the binary solution X t , a random ζ is
 2j  2
solve the DNA sequence optimization problem [16]. Li et generated from the interval [0, 1]; if z < bjt  or z < 1  atjt  ,
 t

al. propose a quantum-inspired GA for multi-objective flow then set xtji ¼ 1, otherwise set xtji ¼ 0. Thus, the binary
shop scheduling problem [17]. Li et al. propose a novel solution Xjt , which actually is an m-length binary string of
QEA based on the Bloch coordinates of Q-bits, in which length m, can be constructed.
three coordinates of a point on the Bloch sphere are
regarded as three paratactic genes, each chromosome 2.2 Quantum gate
contains three gene chains, and each gene chain represents
an optimization solution [18]. Vlachogiannis and Lee In the QEA, the state of Q-bit individual can be changed by
present a QEA for bid-based optimal real and reactive a quantum gate. There are several kinds of quantum gates,
power (P-Q) dispatch [19]. such as NOT gate, controlled NOT gate, rotation gate,
A hybrid chaotic quantum evolutionary algorithm (CQEA) Hadamard gate, etc. [15]. A quantum rotation gate is
is presented in the paper and is first used to solve the resources employed to update each Q-bit in a Q-bit individual. The
combinatorial optimization problem in MGrid system. By operation of the rotation gate Ujit on each Q-bit in the jth Q-
merging the QEA and the chaotic search, the CQEA cannot bit individual qtj is presented as:
only avoid the disadvantage of easily getting to the local " # " #
atþ1
ji atji
optional solution but also keeps the rapid convergence ¼ Ujit ; ði ¼ 1; 2; 3 . . . ; mÞ ð2Þ
performance. Furthermore, we propose a novel resource btþ1
ji btji
encoding method for CQEA. We argue that a valid and simple
resource encoding method is necessary and important for the where m is the number of Q-bits in qtj and the rotation gate
QEA, especially for the CQEA. This is because there are is defined by
frequent encoding and decoding in the proposed CQEA. 2    3
cos q tji  sin q tji  
The rest of the paper is organized as follows. In Section 2, Ujit ¼ 4     5; qtji ¼ r atji ; btji  $q ð3Þ
the original QEA is briefly introduced. A hybrid CQEA is sin qtji cos q tji
proposed for the RCO problem in MGrid system in Section 3.
Section 4 proposes a novel resource encoding method for the Here, qtji is a rotation angle of each Q-bit toward either 0
CQEA. Section 5 presents two fitness functions that are or 1 state. The value of Δθ is set to 0.09π in the paper.
employed in the CQEA. Section 6 gives and analyzes the The rotation angles in the QEA are given in Table 1,
experimental results. Section 7 outlines the conclusion and where bi is the ith bit of the current best solution B, and ri is
future works followed by acknowledgements. the ith bit of the jth Q-bit individual at the generation t. The
original QEA algorithm is shown in Fig. 1.

2 Quantum evolutionary algorithm


3 Chaotic quantum evolutionary algorithm
2.1 Representation of QEA
3.1 Chaotic search
In QEA, a Q-bit is defined as the smallest information unit,
which is defined as [a, β]′, where |a|2 +|β|2 =1. The modulus Chaos is a kind of universal nonlinear phenomenon, whose
|a|2 and |β|2 give the probabilities that the state of Q-bit will action is complex and similar to random behavior. Owing to
Int J Adv Manuf Technol (2011) 52:821–831 823

Table 1 The table of rotation    


angle in the QEA xtji bi f Xjt  f ðBÞ Δθ r atji ; btji

atji btji 0 atji ¼ 0 atji btji 0 btji ¼ 0

0 0 False 0.09π −1 ±1 +1 ∓1
0 0 True 0.09π −1 ±1 +1 ±1
0 1 False 0.09π +1 ∓1 −1 ±1
0 1 True 0.09π −1 ±1 +1 ∓1
1 0 False 0.09π −1 ±1 +1 ∓1
1 0 True 0.09π +1 ∓1 −1 ±1
1 1 False 0.09π +1 ∓1 −1 ±1
1 1 True 0.09π +1 ±1 −1 ±1

the ergodicity, randomicity, irregularity, and special ability CQEA is developed for the RCO problem in MGrid
of avoiding being trapped in a local optimal solution, the system. The procedures of CQEA are described as follows:
chaotic search is a novel optimization technology and has
Step 1: Given a manufacturing task T={ST1, ST2, ST3,⋅⋅⋅,
attracted considerable attention for application in various
STN}, Mj is the number of candidate resources for
fields. The paper introduces the chaotic search into the
the jth subtask STj (j=1, 2, ⋅⋅⋅, N). As discussed in
original QEA, which is used to solve the RCO problem in
Section 2.1, let m refer to the sum of bit numbers
MGrid system.
of M1, M2,⋅⋅⋅, MN (for how to encode resources for
The well-known logistic map is employed for generating
CQEA, please go to Section 4).
chaotic variables. The equation of logistic map is defined as
Step 2: Set t=0 and initialize  a population of Q-bit
follows:
individuals, QðtÞ ¼ qt1 ; qt2 ; qt3 ;    ; qtn , where
xcnþ1 ¼ m  xcn ð1  xcn Þ; ð0  x0  1Þ ð4Þ each Q-bit individual, qti , (i=1, 2,⋅⋅⋅, n), is a group
of Q-bits with length m. 
where μ is the control parameter, xcn is a chaotic variable Step 3: Make a population of binary solutions PðtÞ ¼ X1t ;
and cn=0, 1, 2, ⋅⋅⋅. When μ=4 and x0 ∉{0, 0.25, 0.5, 0.75, X2t ; X3t ;    ; Xnt g by observing the states of QðtÞ ¼
1}, slight variations in the initial chaotic variables yield 
qt1 ; qt2 ; qt3 ;    ; qtn , as discussed in Section 2.1.
dramatically different results over time. In addition, chaotic Step 4: Repair P(t).
variables can traverse all states in a space without repetition Step 5: Evaluate P(t) using the fitness functions and store
to find the global optimal solution. the best solution B.
The simple chaotic search algorithm is shown in Fig. 2. Step 6: Update Q(t) with the rotation gate (3).
Step 7: Repeat step 3 to 6 until B does not change within a
3.2 Chaotic quantum evolutionary algorithm certain number of iterations K1.

In order to avoid the local optimum caused by the QEA, the


chaotic search is introduced into the QEA. A novel hybrid
Step 1: Set the maximum iteration K 2 , the initial best fitness
Step 1: Set t = 0 ;
value f * and the initial chaotic variable x0 ;
Step 2: Initialize a population of Q-bit individuals Q ( t ) ;
Make P (t ) by observing the states of Q ( t ) ; Step 2: Change chaotic variables xcn into the decision

Evaluate P (t ) using the fitness function f ( X );


t variables of practical optimal problems;
j Step 3: Evaluate the new optimal solution with decision
Store the best solution among P (t ) into B ; variables using the fitness function f (.) ;
Step 3: While the termination criterion K1 is not reached, do Step 4: If the fitness value of new optimal solution is greater
t = t +1; than that of the stored best solution, then store the
Make P (t ) by observing the states of Q ( t − 1) ; new optimal solution and fitness value as the current
Evaluate P (t ) ; best solution and best fitness value f * , respectively;
Update Q ( t ) with a quantum gate; Step 5: While the maximum iteration is reached, output the
best solution and fiteness values; otherwise, let
Store the best solution among B and P (t ) into
cn = cn + 1 , calculate xcn using the equation (4)
B;
End. and go to Step 2.

Fig. 1 The original QEA Fig. 2 The simple chaotic search algorithm
824 Int J Adv Manuf Technol (2011) 52:821–831

Step 8: Calculate each index of allocated resource rj (j=1, Table 2 An example of resource encoding method
2,⋅⋅⋅, N) using the chaotic variables
 
in the Subtasks ST1 ST2 ST3 ST4 ST5
following equation: rj ¼ 1 þ xcn Mj  1 .
Step 9: Evaluate the new resource combination R={r1, r2, Mj Decimal 5 6 3 8 4
r3,⋅⋅⋅, rN} using the fitness functions. Binary 101 110 11 1000 100
Step 10: If R is better than B, then R is transformed into rj Decimal 3 2 3 2 1
the binary solution RX, as discussed in Section 4. Binary 011 010 11 0010 001
Store RX into B and set t=0 and initialize each The binary solution X 011 010 11 0010 001
Q-bit individual in Q(t) with RX and next go to The length of Q-bit individual m=3+3+2+4+3=15
step 6; otherwise, go to step 8 until the
predefined maximum iteration K2 is reached.
From the above example, it is concluded that the
When any of binary solution X in P(t) violates the proposed method is dynamic and flexible. According to
number of candidate resources, the following repair process both of the numbers of subtasks and candidate resources for
is employed (as shown in Fig. 3). each subtask, the method can obtain a binary solution for a
resource composition easily, namely resource encoding.
Meanwhile, the binary solution can also be transferred to
4 Resource encoding method for CQEA the index of allocated resource easily, namely resource
decoding. Most important of all, the binary solution
According to the statement of the RCO problem in [13], an obtained by the proposed method uses a Q-bit individual
MGrid task is defined as T={ST1, ST2, ST3,⋅⋅⋅, STN}, with the shortest length to represent a resource composition.
where N is the number of subtasks, and STj (j=1, 2,⋅⋅⋅, N)
denotes the jth subtask of T. Mj is the number of candidate
resources for STj. 5 The fitness functions
rj is defined as the index of allocated resource for
subtask STj, (1≤rj ≤Mj). Hence, R={r1, r2, r3,⋅⋅⋅, rN} refers The paper employs six QoS properties of MGrid resource as
to a resource combination for task T. We transfer each Mj the evaluation criteria of RCO. They are time (T), cost (C),
and rj (decimal numbers) into binary numbers, respectively. reliability (Rel), maintainability (Ma), trust-QoS (Trust), and
It is noted that the bit number of binary rj must be identical function similarity (FS). Interested readers may refer to [13]
to that of the corresponding Mj. If the bit number of binary for more details. The higher the value of T and C, the lower
rj was less than that of Mj, then 0 should be added before the quality; while the higher the value of Rel, Ma, Trust, and
the binary number of rj. Thus, these binary numbers of rj FS, the higher the quality. Motivated by [20], T(rj) and C(rj)
arranged in sequence can constitute a binary string (namely, will be scaled according to Eq. 5, and Rel(rj), Ma(rj), Trust
a Q-bit individual). For example, as shown in Table 2, a (rj), and FS(rj) will be scaled according to Eq. 6.
resource combination, R={r1, r2, r3, r4, r5}={3, 2, 3, 2, 1}, 8 Δ Δ
can be represented by the binary solution X={01101011 < Valuemax  Valuej ; if ValueΔ ≠ValueΔ
>
Δ Δ
Scaled valuej ¼ Valuemax  Valuemin
Δ max min
0010001}. The binary length for each of the five candidate >
:
resources is 3, 3, 2, 4, and 3, respectively. Therefore, the 1; otherwise
length of each Q-bit individual is 3+3+2+4+3=15. In ð5Þ
other words, it consists of 15 Q-bits and can represent 215 8 Δ Δ
binary states simultaneously. Conversely, given {Mj|j=1, 2, < Valuej  Valuemin ; if ValueΔ ≠ValueΔ
>
Δ
⋅⋅⋅, N.}, we can obtain a resource combination from a binary Scaled valuej ¼ Valuemax  ValueΔ
Δ max min
>
:
min

solution, which is important for evaluating P(t) in Section 5. 1; otherwise


ð6Þ

Repair (X ) where Δ={T, C, Rel, Ma, Trust, FS}, ValueΔ j is the value
of a certain kind of property (e.g., T(rj), C(rj), Rel(rj), Ma
Step 1: According to Section 4, the binary solution X is
transformed into the decimal number string Table 3 Two scenarios for resource composition R1 better than R2
R = {r1 , r2 , r3 ,…, rN } .
Step 2: If rj = 0 then set rj = 1 ; else if rj > M j Index Violation(⋅) f(⋅)

( j = 1, 2, 3, …, N ) then set rj = M j . 1 Violation(R2)>Violation(R1) N/A


2 Violation(R2)=Violation(R1) f(R1)>f(R2)
Fig. 3 The process of repair
Int J Adv Manuf Technol (2011) 52:821–831 825

Table 4 Three groups of parameter values

Index Figure The number of candidate resources The length of Q-bit individual Others

Group 1 Figs. 1 and 2 M1 =M2 =M3 =M4 =M5 =10 m=42×5=20 N ¼5


Group 2 Figs. 3 and 4 M1 =M2 =M3 =M4 =M5 =50 m=6×5=30 100  K1  1; 050
Group 3 Figs. 5 and 6 M1 =M2 =M3 =M4 =M5 =100 m=7×5=35 100  K2  1; 050

(rj), Trust(rj), FS(rj)); ValueΔ


max is the maximal value of a similarity, reliability, and maintainability, respectively. In
certain kind of property of all candidate resources, while the paper, the fitness functions focus on three QoS
ValueΔ min is the minimal value of a certain kind of property properties of MGrid resources: time, cost, and trust-QoS.
of all candidate resources. After scaling the six properties of In order to deal with these inequality constraints in
MGrid resources, the RCO problem in MGrid system can Eq. 8, it is a popular approach to use penalty functions.
be converted into the following constraint combinatorial However, they generally need to be set up a series of
optimization problem: penalty factors. In order to avoid setting these annoying
factors, we separate the fitness functions into the objective
max f ðRÞ ¼ a  Scaled T ðRÞ þ b  Scaled CðRÞ function as represented in Eq. 7 and the constraint violation
þ g  Scaled TrustðRÞ ð7Þ 9. The constraint violation value of a resource combination
R is calculated as follows:
8  
< FS rj  FS0 ; 8j ¼ 1; 2; 3; . . . ; N
s:t: Rel r  Rel0 ; 8j ¼ 1; 2; 3; . . . ; N ð8Þ P
N FS0 FSðrj Þ
:  j ViolationðRÞ ¼ max 0;
Ma rj  Ma0 : 8j ¼ 1; 2; 3; . . . ; N j¼1
FS0

P
N Rel0 Relðrj Þ
where
8 þ max 0; Rel0 ð9Þ
> PN j¼1

>
> Scaled T ðRÞ ¼ Scaled valueTj =N ; P
>
>
N Ma0 Maðrj Þ
>
> j¼1 þ max 0;
< PN j¼1
Ma0
Scaled CðRÞ ¼ Scaled valueCj =N ;
>
>
>
>
j¼1
Hence, we argue that R1 is better than R2 in any of the
>
> PN
>
: Scaled TrustðRÞ ¼ Scaled valueTrust =N : following two scenarios (as shown in Table 3).
j
j¼1 The fitness functions 7 and 9 will be critical to steps 5
R={r1, r2, r3,⋅⋅⋅, rN} represents a resource combination; and 9 of the CQEA in Section 3.2. In the fitness functions,
a, β, and γ are the corresponding weighted factors; FS0, there are only three weighted factors that MGrid users need
Rel0, and Ma0 are the minimum requirements of function to set, according to the users’ requirement.

Table 5 The property values (Time) in group 1

Index Time

Subtasks

ST1 ST2 ST3 ST4 ST5

1 57.61 (0.43) 33.04 (0.68) 1.79 (1.00) 67.82 (0.33) 43.48 (0.58)
2 50.73 (0.50) 25.94 (0.75) 96.65 (0.03) 80.54 (0.20) 33.08 (0.68)
3 32.58 (0.69) 39.99 (0.61) 85.45 (0.15) 32.87 (0.68) 43.99 (0.57)
4 90.80 (0.09) 38.28 (0.63) 24.80 (0.77) 82.53 (0.18) 96.08 (0.04)
5 68.11 (0.32) 89.54 (0.11) 95.06 (0.05) 99.98 (0.00) 56.26 (0.45)
6 53.70 (0.47) 2.18 (0.99) 75.15 (0.25) 21.73 (0.80) 19.12 (0.82)
7 82.00 (0.18) 38.61 (0.63) 91.69 (0.08) 77.21 (0.23) 17.63 (0.84)
8 19.03 (0.82) 68.48 (0.32) 92.14 (0.80) 5.67 (0.96) 68.51 (0.32)
9 33.06 (0.68) 21.82 (0.79) 29.30 (0.72) 89.40 (0.11) 22.90 (0.79)
10 78.20 (0.22) 72.04 (0.28) 33.82 (0.67) 60.76 (0.40) 79.95 (0.20)
ValueTmax ¼ 99:98ð0:00Þ ValueTmin ¼ 1:79ð1:00Þ

The numbers in brackets represent the values of scaled properties, as discussed in Section 5
826 Int J Adv Manuf Technol (2011) 52:821–831

Table 6 The property values


(Cost) in group 1 Index Cost

Subtasks

ST1 ST2 ST3 ST4 ST5

1 79.11 (0.18) 92.58 (0.04) 31.69 (0.68) 56.65 (0.42) 29.87 (0.69)
2 77.65 (0.20) 4.14 (0.96) 31.61 (0.68) 9.01 (0.91) 0.41 (1.00)
3 33.79 (0.65) 15.53 (0.84) 92.86 (0.04) 44.13 (0.55) 76.79 (0.21)
4 46.09 (0.53) 68.62 (0.29) 72.34 (0.25) 0.57 (0.99) 11.59 (0.88)
5 93.02 (0.04) 56.57 (0.42) 84.91 (0.12) 84.49 (0.13) 37.02 (0.62)
6 96.90 (0.00) 2.78 (0.98) 71.77 (0.26) 88.09 (0.09) 2.72 (0.98)
7 94.53 (0.02) 46.80 (0.52) 67.01 (0.31) 79.97 (0.18) 8.52 (0.92)
8 65.11 (0.33) 3.88 (0.96) 65.83 (0.32) 58.07 (0.40) 27.76 (0.72)
9 33.48 (0.66) 14.57 (0.85) 5.95 (0.94) 24.85 (0.75) 76.83 (0.21)
10 52.61 (0.46) 35.59 (0.64) 20.99 (0.79) 66.13 (0.32) 38.99 (0.60)
ValueCmax ¼ 96:90ð0:00Þ ValueTmin ¼ 0:41ð1:00Þ

Table 7 The property values


(Trust-QoS) in group 1 Index Trust-QoS

Subtasks

ST1 ST2 ST3 ST4 ST5

1 85.03 (0.88) 11.19 (0.09) 65.40 (0.67) 94.45 (0.98) 49.47 (0.50)
2 50.38 (0.51) 37.98 (0.38) 42.28 (0.42) 95.55 (0.99) 95.45 (0.99)
3 88.99 (0.92) 45.54 (0.46) 96.24 (1.00) 63.00 (0.65) 15.72 (0.14)
4 89.33 (0.93) 71.38 (0.74) 29.47 (0.29) 18.35 (0.17) 81.36 (0.84)
5 43.56 (0.44) 72.27 (0.74) 20.97 (0.20) 67.83 (0.70) 34.80 (0.35)
6 66.26 (0.68) 59.97 (0.61) 27.41 (0.27) 2.99 (0.01) 9.35 (0.07)
7 74.52 (0.76) 74.74 (0.77) 37.99 (0.40) 22.78 (0.22) 6.10 (0.04)
8 18.18 (0.17) 83.53 (0.86) 86.37 (0.89) 2.41 (0.00) 10.94 (0.09)
9 13.46 (0.12) 7.13 (0.05) 83.47 (0.86) 19.84 (0.18) 62.68 (0.64)
10 15.94 (0.14) 79.82 (0.83) 52.16 (0.53) 75.83 (0.78) 24.49 (0.24)
max ¼ 96:24ð1:00Þ
ValueTrust min ¼ 2:41ð0:00Þ
ValueTrust

Fig. 4 Fitness (×100) for two


increasing iterations K1, K2 in
group 1
Int J Adv Manuf Technol (2011) 52:821–831 827

Fig. 5 Fitness (×100) for two


increasing iterations K1, K2 in
group 2

Fig. 6 Fitness (×100) for two


increasing iterations K1, K2 in
group 3

6 Experimental results

In this section, we perform a number of experiments to


examine the performance of the CQEA, in terms of
effectiveness (the convergence characteristics) and compu-
tational efficiency (the computation time to solve RCO
problems). Besides effectiveness and efficiency, we also
investigate the scalability of CQEA with respect to
increasing the number of candidate resources. Furthermore,
we study the effect of the size of the population of Q-bit
individuals. At last, we compare the performances of the
original QEA and the CQEA. The QEA and CQEA are
implemented with MATLAB 7.6.0 and run on a personal
computer with CPU 2.67GHz and RAM 6.00GB.
The number of Q-bit individuals n in Section 2.1 is set to
20. The parameters of the fitness functions in Section 5 are Fig. 7 Fitness (×100) for the increasing iteration K1 at the iteration
a = 0.4, β = γ = 0.3. The initial chaotic variable x0 in K2 =1,000 in groups 1, 2, and 3
828 Int J Adv Manuf Technol (2011) 52:821–831

Fig. 8 Computation time in


seconds for two increasing
iterations K1, K2 in group 1

Fig. 9 Computation time in


seconds for two increasing
iterations K1, K2 in group 2

Fig. 10 Computation time in


seconds for two increasing iter-
ations K1, K2 in group 3
Int J Adv Manuf Technol (2011) 52:821–831 829

Fig. 11 Computation time in seconds for the increasing iteration K1 at Fig. 13 Fitness (×100) for two increasing iterations K1, K3 in group 2
the iteration K2 =1,000 in groups 1, 2, and 3

Section 3.1 is set to 0.6. Three groups of the parameter 50 times, and the average value (e.g., fitness and compu-
values, such as the number of candidate resources Mj, the tation time) is recorded as a final result.
number of subtasks N, the iteration of QEA K1, and the In the first experiment, we focus on measuring the
iteration of chaotic search K2, are used in the experiments. effectiveness of CQEA. We start with candidate resources
The different parameter values for each specific group are of group 1. We gradually increase the number of iteration of
given in Table 4. Based on these parameters, the values of QEA K1 from 100 to 1,050, while increasing the number of
six QoS properties of MGrid resources are randomly iteration of chaotic search K2 from 100 to 1,050. The fitness
generated within the range of (0, 100]. The values of three result is shown in Fig. 4. Apparently, the CQEA achieves
properties (T, C, and Trust) in group 1 are given in Table 5, the best solution from the very beginning. The convergence
6, and 7. Note that for convenience of comparison results, characteristics of CQEA for groups 2 and 3 are shown in
the number of subtasks N for each group is set to 5, and we Figs. 5 and 6, respectively. In order to examine the
assume these three properties (Rel, Ma, and FS) would not effectiveness of CQEA more clearly, we cut through the
violate their respective minimum requirements set by 3D-plot along the axis K2 =1,000 in Figs. 4, 5, and 6,
MGrid users. In order to account for the random nature of respectively, and then the three cutting curves are shown
the quantum evolution, each experimental run is repeated together in Fig. 7. From Fig. 7, it is concluded that the

Fig. 12 Fitness for increasing the number of Q-bit individuals at the


iterations K1 =K2 =1,000 in groups 1, 2, and 3 Fig. 14 Fitness (×100) for two increasing iterations K1, K3 in group 3
830 Int J Adv Manuf Technol (2011) 52:821–831

CQEA achieves the best solutions at the number of iteration Reducing the number of iteration is beneficial for improving
of QEA K1 =800, 950 for groups 2 and 3, respectively. the performance of CQEA.
In the second experiment, we are interested in measuring
the computational efficiency of CQEA. The method and
parameters are the same with the first experiment. The Acknowledgments This paper is supported by the National Natural
Science Foundation Key Project of China: Digit Manufacturing Basic
computation time for groups 1, 2, and 3 are shown in Theories and Key Techniques under Network Environment (no.
Figs. 8, 9, and 10, respectively. We compare three groups of 50335020) and the Hubei Digital Manufacturing Key Laboratory
computation time in Fig. 11. Apparently, both of the Opening Fund Project: Research on Resource Service Search and
computation time for groups 2 and 3 is greater than that for Optimal-Selection Theories and Experiments in Manufacturing Grid
System (no. SZ0621). The authors would like to thank the anonymous
group 1. Especially, the computation time for group 2 is reviewers for their comments and suggestions.
almost identical to that for group 3. The phenomenon is
explained by the fact that both of the lengths of Q-bit
individual of groups 2 and 3 are much larger than that of References
group 1. Therefore, the proposed CQEA is sensitive to the
length of Q-bit individual. The shorter length of Q-bit 1. Fan Y, Zhao D, Zhang L, Huang S and Liu B (2004)
individual is benefit for improving the performance of CQEA. Manufacturing grid needs, concept, and architecture. Lecture
In the third experiment, we examine the influence of the Notes in Computer Science. Springer, Heidelberg, 3032:653–
size of the population of Q-bit individuals on the 654
2. Qiu RG (2004) Manufacturing grid: a next generation manufac-
performance of CQEA. Figure 12 shows the average turing. IEEE International Conference System, Man and Cybe-
achieved solution of Eq. 7 achieved as a function of the metics. The Hague, The Netherlands, Oct. 10–13, pp. 4667–4672
number of Q-bit individuals in the population ranging from 3. Zhang HJ, Hu YF, Tao F, Zhou ZD (2008) Study on semantic-
five to 25 Q-bit individuals at the iterations K1 =K2 =1,000. aware manufacturing grid architecture. Fuzzy Systems and
Knowledge Discovery, Jinan, China, Oct. 18–20, pp. 626–630
From Fig. 12, it is concluded that the CQEA with 20 Q-bit 4. Ding YF, Tao F, Sheng BY, Zhou ZD (2008) Modelling and
individuals starts to achieve the best solutions of the three application of optimal-selection evaluation for manufacturing grid
groups. resource. Int J Comput Integr Manuf 21(1):62–72
In the fourth experiment, we compare the performances 5. Tao F, Hu YF, Zhao DM, Zhou ZD, Zhang HJ, Lei Z (2009)
Study on manufacturing grid resource service QoS modeling and
of the traditional QEA and the CQEA. The convergence of evaluation. Int J Adv Manuf Technol 41(9–10):1034–1042
QEA for groups 2 and 3 are shown in Figs. 13 and 14, 6. Tao F, Hu YF, Zhao DM, Zhou ZD (2009) An approach to
respectively, where K3 refers to the order of experiments, manufacturing grid resource service scheduling based on trust-
and the values of fitness employ the average fitness value QoS. Int J Comput Integr Manuf 22(2):100–111
7. Zhang HJ, Hu YF, Zhou ZD (2009) Research on Co-reservation in
of 50 experiments. It is obvious that there are fluctuations the manufacturing grid system. Int J Adv Manuf Technol 47(5–
in Figs. 13 and 14, especially in Fig. 13. This is because the 8):699–717
traditional QEA is easy to fall into the local optimism, in 8. Tao F, Hu YF, Zhao DM, Zhou ZD (2008) Resource service
which the value of fitness is up and down. In contrast, the composition and its optimal-selection based on particle swarm
optimization in manufacturing grid system. IEEE Transactions on
surface is smoothing in Figs. 5 and 6. It demonstrates that Industrial Informatics 4(4):315–327
the chaotic search can help the QEA escape from the local 9. Wieczorek M, Hoheisel A, Prodan R (2009) Towards a general
optimism. model of the multi-criteria workflow scheduling on the grid.
Future Gener Comput Syst 25(3):237–256
10. T’kindt V, Billaut J (2002) Multicriteria scheduling. Springer,
Berlin
7 Conclusions and future works 11. Yu J, Buyya R (2006) Scheduling scientific workflow applications
with deadline and budget constraints using genetic algorithms. Sci
Executing complex manufacturing tasks is a major goal of Program 14(3–4):217–230
12. Masutti TAS, de Castro LN (2009) A self-organizing neural
the development of MGrid technologies. Thus, RCO is network using ideas from the immune system to solve the
critical to MGrid system. In this paper, we propose a hybrid traveling salesman problem. Inf Sci 179(10):1454–1468
CQEA for the RCO problem in MGrid system. Meanwhile, 13. Shu W (2009) Quantum-inspired genetic algorithm based on
we also present a dynamic and simple resource encoding simulated annealing for combinatorial optimization problem.
International Journal of Distributed Sensor Networks 5(1):64–65
method for CQEA. The simulation experiments are imple- 14. DiVincenzo DP (1995) Quantum computation. Science 270
mented with the goal of minimizing time, cost, and (5234):255–261
maximizing trust-QoS. The experimental results show that 15. Han KH, Kim JH (2002) Quantum-inspired evolutionary algo-
the proposed CQEA is efficient, effective, and scalable to rithm for a class of combinatorial optimization. IEEE Trans Evol
Comput 6(6):580–593
solve the RCO problem in MGrid system. A recommended 16. Xiao J, Xu J, Chen Z, Zhang K, Pan L (2009) A hybrid quantum
future work includes how to design a proper iteration chaotic swarm evolutionary algorithm for DNA encoding.
criterion, not a fixed maximum number of generations. Comput Math Appl 57(11–12):1949–1958
Int J Adv Manuf Technol (2011) 52:821–831 831

17. Li BB, Ling W (2006) A hybrid quantum-inspired genetic 19. Vlachogiannis JG, Lee KY (2008) Quantum-inspired evolutionary
algorithm for multi-objective scheduling. Lecture Notes in algorithm for real and reactive power dispatch. IEEE Trans Power
Computer Science. : Springer, Berlin, 4113:511–522 Syst 23(4):1627–1636
18. Li P, Li S (2008) Quantum-inspired evolutionary algorithm for 20. Zeng L, Benatallah B, Ngu AHH, Dumas M, Kalagnanam J,
continuous space optimization based on Bloch coordinates of Chang H (2004) QoS-aware middleware for Web services
qubits. Neurocomputing 72(1–3):581–591 composition. IEEE Trans Softw Eng 30(5):311–327

Вам также может понравиться