Академический Документы
Профессиональный Документы
Культура Документы
Contents
1 Presburger's Pro
edure . . . . . . . . . .
2 Newell, Shaw & Simon, and H. Gelernter
3 First-Order Logi
. . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . .
Index . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
6
7
12
15
With the ready availability of serious
omputer power, dedu
tive reasoning, espe
ially as embodied in mathemati
s, presented an ideal target for those interested in
experimenting with
omputer programs that purported to implement the \higher"
human fa
ulties. This was be
ause mathemati
al reasoning
ombines obje
tivity
with
reativity in a way di
ult to nd in other domains. For this endeavor, two
paths presented themselves. One was to try to understand what people do when
they
reate proofs and to write programs emulating that pro
ess. The other was
to make use of the systemati
work of the logi
ians in redu
ing logi
al reasoning
to standard
anoni
al forms on whi
h algorithms
ould be based. Ea
h path
onfronted daunting obsta
les. The di
ulty with the rst approa
h was that available
information about how
reative mathemati
ians go about their business was and remains vague and ane
dotal. On the other hand, the well-known unsolvability results
of Chur
h and Turing showed that the kind of algorithm on whi
h a programmer
might want to base a theorem-proving program simply did not exist. Moreover, it
was all too obvious that an attempt to generate a proof of something non-trivial
by beginning with the axioms of some logi
al system and systemati
ally applying the rules of inferen
e in all possible dire
tions was sure to lead to a giganti
ombinatorial explosion.
Ea
h of these approa
hes has led to important and interesting work. Unfortunately, for many years the proponents of the two approa
hes saw themselves as
opponents and engaged in polemi
s in whi
h they largely spoke past ea
h other.
One problem was that whereas they appeared to be working on the same problems,
they tended to dier not only in their approa
hes, but also in their fundamental
goals. Those whose method was the emulation of the human mathemati
ian tended
to see their resear
h as part of an eort to help understand human thought. Those
who proposed to use the methods of mathemati
al logi
tended to see the goal as
the development of useful systems of automated dedu
tion. Ultimately, the most
su
essful developments in
orporated insights deriving from both approa
hes.
For a brief a
ount of the history of the developments in logi
that provided the
ba
kground for resear
h in this eld see [Davis 1983
. An interesting a
ount of
the two approa
hes and their mutual intera
tion
an be found in [Ma
Kenzie 1995.
The volume [Siekmann and Wrightson 1983 is a useful anthology of the prin
ipal
arti
les on automated dedu
tion to appear in the years through 1966.
Martin Davis
great triumph was to prove that the sum of two even numbers is even.
Although it would be di
ult to
laim that this work has helped very mu
h with
su
h an ambitious agenda, it did provide a paradigm employed by many theoremprovers developed later, and this was surely its lasting in
uen
e. Among the te
hniques made expli
it were forward and ba
kward
haining, the generation of useful
subproblems, and seeking substitutions that produ
e desired mat
hes.
The authors emphasize that their program is \heuristi
" rather than \algorithmi
," and this purported distin
tion has given rise to mu
h dissension and
onfusion. In this
ontext, \heuristi
" seems to mean little more than the la
k of a
guarantee that the pro
ess will always work (given su
ient spa
e and time). The
algorithm they
ontrast with their own pro
edure is the \British Museum algorithm" by whi
h all possible proofs are generated until one leading to the desired
result is rea
hed. Indeed, the authors seem to have been unaware that Post's proof
of the
ompleteness of the Prin
ipia propositional
al
ulus using truth tables had,
in ee
t supplied a simple algorithm by means of whi
h a demonstration by truth
tables
ould be
onverted into a proof in Prin
ipia [Post 1921.
Wang and Gao [1987 presented a Gentzen-style proof system for the propositional
al
ulus designed for e
ien
y. Unlike the program of Newell et al, Wang's
system is
omplete: for any input, pro
essing eventually halts, yielding either a
1 A
tually
Whitehead and Russell's tenden
y to
onfuse obje
t and meta-language led them to
state this
onfusingly as \Anything implied by a true proposition is true." But this lapse is not
important for the present dis
ussion.
proof or a disproof. The simple examples that are expli
itly listed in Prin
ipia,
in
luding those that stumped the Logi
Theory Ma
hine, were easily disposed of.
Although Wang seems not to have quite understood that produ
ing an e
ient
generator of proofs in the propositional
al
ulus was not what the Logi
Theory
Ma
hine designers were after, they did leave themselves open to Wang's
riti
ism
by giving the impression that the absurd British Museum algorithm was the only
possible \non-heuristi
" proof-generating system for the propositional
al
ulus.
Like the propositional
al
ulus, the elementary geometry of the plane
an be
spe
ied by a formal system for whi
h an algorithmi
de
ision pro
edure is available.
This is seen by introdu
ing a
oordinate system and relying on the redu
tion of
geometry to algebra and Tarski's de
ision pro
edure for the algebra of the real
numbers. However, unlike the
ase of truth table methods for the propositional
al
ulus, this method is utterly unfeasible. Although theoreti
al
onrmation of this
did not
ome until mu
h later, it was already apparent from Davis's experien
e with
the mu
h simpler Presburger pro
edure. Herbert Gelernter's Geometry Ma
hine is
very mu
h in the spirit of Newell at al. A
lue to Gelernter's orientation is provided
by his statement [Gelernter 1959:
. . . geometry provides illustrative material in treatises and experiments in human problem-solving. It was felt that we
ould ex
hange valuable insights with
behavioral s
ientists . . .
Te
hni
ally, in addition to the repertoire of The Logi
Ma
hine (ba
kward
haining, the use of subproblems), the geometry ma
hine introdu
ed two interesting
innovations: the systemati
use of symmetries to abbreviate proofs and the use of a
oordinate system to simulate the
arefully drawn diagram a student of geometry
might employ. This last was used to tip o the prover to the fa
t that
ertain pairs
of line segments and of angles \appeared" to be equal to one another, and thereby
to guide the sear
h for a proof.
3. First-Order Logi
Unlike the
ases of propositional logi
and elementary geometry, there is no general
de
ision pro
edure for rst-order logi
. On the other hand, given appropriate axioms
as premises, all mathemati
al reasoning
an be expressed in rst- order logi
, and
that is why so mu
h attention has been paid to proof pro
edures for this domain.
Investigations by Skolem and Herbrand in the 1920s and early 1930s provided the
basi
tools needed for theorem-proving programs for rst-order logi
[Davis 1983
.
In 1957 a ve week Summer Institute for Symboli
Logi
held at Cornell University was attended by almost every logi
ian working in the United States. Many
of the more theoreti
ally in
lined resear
hers from the nearby IBM fa
ilities were
also present; FORTRAN, a brand-new innovation in programming pra
ti
e was unveiled. After dis
ussions with Gelernter, the logi
ian Abraham Robinson was led to
give a short talk [Robinson 1957 in whi
h he pointed to Skolem fun
tions and \Herbrand's theorem" as useful tools for general purpose theorem-provers. He also made
the provo
ative remark that the auxilliary points, lines, or
ir
les \
onstru
ted" as
Martin Davis
the program they found that the rule for eliminating atomi
formulas (later
alled
ground resolution) whi
h repla
ed a formula
(p _ A) ^ (:p _ B ) ^ C
by
(A _ B ) ^ C
used too mu
h RAM. So it was proposed to use instead the splitting rule whi
h
generates the pair of formulas
A
The idea was that a sta
k for formulas to be tested
ould be kept in external
storage (in fa
t a tape drive) so that formulas in RAM never be
ame too large.2
Although testing for satisability was performed very e
iently, it soon be
ame
lear that no very interesting results
ould be obtained without rst devising a
method for avoiding the generation of spurious elements of the Herbrand universe
[Davis, Logemann and Loveland 1962.
During the same years, Hao Wang was attempting to apply some of the more
sophisti
ated work that had been done in proof theory and on solvable
ases of
Hilbert's Ents
heidungsproblem to automati
dedu
tion programs. He announ
ed
a
omputer program that proved all of the theorems (about 400) of Whitehead
and Russell's Prin
ipia Mathemati
a of rst-order logi
with equality [Wang and
Zhi 1998, Wang and Zhi 1998. However, this apparently momentous a
hievement
in automating dedu
tion was (as Wang himself pointed out) possible only be
ause all of these theorems
an be brought into prenex form with the simple prex
8 : : : 89 : : : 9. Wang
on
luded that:
The most interesting lesson from these results is perhaps that even in a fairly
ri
h domain, the theorems a
tually proved are mostly ones whi
h
all on a very
small portion of the available resour
es of the domain. ([Wang 1963
p. 32)
Prawitz's [1960 in
uential paper taught the growing automated dedu
tion
ommunity that unne
essary terms in the Herbrand expansion
ould be avoided by using
algorithms that did not generate elements of the Herbrand universe until needed.
Most later progress was based on this key insight. Prawitz's pro
edure worked by
obtaining expansions into disjun
tive normal form before repla
ing variables by
2 Unfortunately
both pro
edures using rules 1,2, and 3 and pro
edures using rules 1,2, and 4
have been
alled the \Davis-Putnam pro
edure" in the literature; the rst is generally
onsidered
for worst
ase analysis while it is the se
ond that is ordinarily implemented.
Wolfgang Bibel has kindly pointed out to me that the \rule for eliminating atomi
formulas"
otherwise known as \ground resolution" was rst proposed in A. Blake's dissertation in 1937 and
(in its dual form) was also mentioned by W.V. Quine in 1955 under the name \
onsensus rule".
For further information, see [Bibel 1993. Otherwise, as far as I know, the other rules mentioned
o
urred for the rst time in [Davis and Putnam 1958.
It should also be mentioned that rules 2 and 4 were found independently by Dunham, Fridsal
and Sward [1959. They emphasized that a program based on these rules performs very ee
tively
without using \heuristi
" devi
es.
10
Martin Davis
elements of the Herbrand universe. The algorithm thus generated disjun
tive normal forms of in
reasing length seeking one with the property that some substitution
from the Herbrand universe would yield a truth-fun
tionally unsatisable formula.3
Sin
e this
ondition amounts to ea
h disjun
tive
lause in
luding a pair of literals
of the form `; :`, it
an be formulated as the need to satisfy a system of equations
in the parameters of the expansion.4
Prawitz's pro
edure was a great improvement over what had been done previously
be
ause no spurious elements of the Herbrand universe were generated. Unfortunately, the huge expansions into disjun
tive normal form that would be generated
by all but the simplest problems made it
lear that, at least as presented, this
was still an unsatisfa
tory pro
edure. However, it
ontained the seminal idea of
sear
hing for substitutions that would transform pairs of literals into negations of
one another. Moreover if existential quantiers are eliminated in favor of Skolem
fun
tions at the outset, instead of systems of equations, one has the simple problem
of unifying pairs of terms.
In his survey paper, Davis [1963 proposed
. . . a new kind of pro
edure whi
h seeks to
ombine the virtues of the Prawitz
pro
edure and those of the Davis-Putnam pro
edure.
The idea, also noted by Dunham and North [1962, was that by the \pure literal
rule" from the Davis-Putnam pro
edure (Rule 2, above), substitutions
an help to
render a
onjun
tive set of disjun
tive
lauses unsatisable only if they su
eed in
transforming a literal from one of these
lauses into the negation of a literal in
another
lause. A theorem-proving program based on these ideas was written by
D. M
Ilroy at Bell Laboratories and was improved and
orre
ted by Peter Hinman.
The program in
luded an implementation of the ordinary uni
ation algorithm
[Chinlund, Davis, Hinman and M
Ilroy 1964.
Merely the existen
e of this volume makes it abundantly
lear that automated
reasoning is a thriving eld with a huge literature. The bimonthly publi
ation The
Journal of Automated Reasoning is devoted entirely to this eld. If one event
an be
pinpointed as marking its emergen
e as a mature subje
t, it would be the publi
ation [Robinson 1965b in whi
h J.A. Robinson announ
ed the resolution prin
iple.
[Robinson 1965b was Robinson's se
ond paper in the area, and it is helpful in tra
ing his thought to begin with the rst [Robinson 1963. He began with the basi
framework of Davis-Putnam: existential quantiers eliminated in favor of Skolem
fun
tions and
onjun
tive normal form. He noted Prawitz's te
hnique for avoiding
spurious elements of the Herbrand universe and Davis's survey paper. Evidently
Davis's sket
h of his proposed pro
edure was insu
iently
lear, and Robinson
wrote:5
3 This
a
ount is not quite a
urate be
ause in Prawitz's paper matters are expressed in terms
of nding a proof rather than a refutation. So what he a
tually did is pre
isely the dual of what
is stated above.
4 As pointed out to the author by G
erard Huet, this same use of equations o
urs already in
Herbrand's [1930, p. 145 thesis.
5 In the interest of
larity, the referen
e numbers in this quote were repla
ed by the numbers in
the present bibliography
orresponding to the same papers.
11
Davis [1963 has therefore proposed a way of exploiting Prawitz' powerful idea
while avoiding Prawitz' disasterous use of normal forms{in mu
h the same way
that the te
hniques of Davis and Putnam [1960 avoid the use of normal forms
whi
h
aused Gilmore's [1960 program to be unable even to solve [an easy problem. From the few remarks at the end of [Davis 1963 it does not yet seem
lear just how Davis will pro
eed, and one waits with great interest his further
resear
hes along these lines.
The rest of the paper has a number of interesting
omputer proofs generated by
using the Davis-Putnam \one literal
lause" rule, and, when that fails, requiring
the user to pre-spe
ify the elements of the Herbrand universe needed to obtain a
proof. Finding these elements was
onje
tured to be \the really `
reative' part of
the art of proof-
onstru
tion."
Robinson's method of resolution introdu
ed in his highly in
uential [1965b revolutionized the subje
t. Robinson found a single rule of inferen
e, easily performable
by
omputer, that was
omplete for rst- order logi
. Using resolution required no
separate pro
edure for dealing with propositional
al
ulus. Starting with the usual
pre-pro
essed
onjun
tive set of disjun
tive
lauses, Robinson's te
hnique was to
seek all possible \uni
ations" that would make it possible to express the set of
lauses as
(` _ A) ^ (:` _ B ) ^ C
where ` is a literal that doesn't o
ur in C . This yields the \resolvent"
(A _ B ) ^ C
whi
h after (A _ B ) is \multiplied out" yields a new set of
lauses that is unsatisable just in
ase the original set was. This was similar to Davis's proposal
[Davis 1963, Chinlund et al. 1964, in seeking uni
ations that generate
omplementary literals. It diers not only in not requiring separate truth fun
tional testing,
but also in not requiring, as part of the input, spe
i
ation of the number of instan
es of ea
h
lause to parti
ipate in the nal proof. [Robinson 1965b is striking
for its
ombinatorial simpli
ity, as well as for the sheer mathemati
al elegan
e of the
presentation. Unfortunately, as soon be
ame apparent, the bare resolution method
ould easily produ
e many thousands of
lauses without rea
hing a proof. Finding
a proof using resolution be
omes the problem of providing
riteria for the order in
whi
h resolutions are to be sought. Early attempts to
ut down the sear
h spa
e
were Robinson's own elegant hyperresolution [Robinson 1965a, and the strategies
of unit preferen
e [Wos, Carson and Robinson 1964 and set of support [Wos, Robinson and Carson 1965.
The three de
ades sin
e the rst implementations of resolution have seen an
outpouring of resear
h devoted to automated reasoning systems. While some of
the most su
essful are based on resolution, others have pro
eeded in dierent
dire
tions. For further information, the reader is referred to the other arti
les in
this volume.
12
Martin Davis
Bibliography
[1993, Dedu
tion. Automated Logi
., A
ademi
Press. with the assistan
e of Steen
Holldobler.
Chinlund T. J., Davis M., Hinman P. and M
Ilroy M. [1964, Theorem proving by mat
hing.
Bell Laboratories.
Davis M. [1957, A
omputer program for presburger's algorithm, in `Summaries of Talks Presented at the Summer Institute for Symboli
Logi
', Institute for Defense Analysis. 2nd
edition, published in 1960. Reprinted as [Davis 1983a.
Davis M. [1963, Eliminating the irrelevant from me
hani
al proofs, in `Pro
. Symp. Applied
Math.', Vol. XV, pp. 15{30. Reprinted as [Davis 1983b.
Davis M. [1983a, A
omputer program for Presburger's algorithm, in J. Siekmann and
G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
',
Vol. 1, Springer Verlag, pp. 41{48. Originally published as [Davis 1957.
Davis M. [1983b, Eliminating the irrelevant from me
hani
al proofs, in J. Siekmann and
G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
',
Vol. 1, Springer, pp. 315{330. Originally published as [Davis 1963.
Davis M. [1983
, The prehistory and early history of automated dedu
tion, in J. Siekmann
and G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
',
Vol. 1, Springer Verlag, pp. 1{28.
Davis, M., ed. [1994, Solvability, Provability, Denability: The Colle
ted Works of Emil L.
Post, Birkhauser.
Davis M., Logemann G. and Loveland D. [1962, `A ma
hine program for theorem proving',
Communi
ations of the ACM 5(1962), 394{397. Reprinted as [Davis, Logemann and Loveland
1983.
Davis M., Logemann G. and Loveland D. [1983, A ma
hine program for theorem proving, in
J. Siekmann and G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
', Vol. 1, Springer, pp. 267{270. Originally published as [Davis et al. 1962.
Davis M. and Putnam H. [1958, Computational methods in the propositional
al
ulus, unpublished report, Rensselaer Polyte
hni
Institute.
Davis M. and Putnam H. [1960, `A
omputing pro
edure for quanti
ation theory', Journal of
the ACM 7(3), 201{215. Reprinted as [Davis and Putnam 1983.
Davis M. and Putnam H. [1983, A
omputing pro
edure for quanti
ation theory, in J. Siekmann and G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational
Logi
', Vol. 1, Springer Verlag, pp. 125{150. Originally published as [Davis and Putnam 1960.
Dunham B., Fridsal R. and Sward G. [1959, A non-heuristi
program for proving elementary
logi
al theorems, in `Pro
. IFIP Congr.', pp. 282{285.
Dunham B. and North J. [1962, Theorem testing by
omputer, in `Symp. Math. Theory
Ma
hines', Brooklyn Poly. Inst., pp. 172{177. Reprinted as [Dunham and North 1983.
Dunham B. and North J. [1983, Theorem testing by
omputer, in J. Siekmann and G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
', Vol. 1, Springer
Verlag, pp. 271{275. Originally published as [Dunham and North 1962.
Gelernter H. [1959, Realization of a geometry-theorem proving ma
hine, in `Pro
. Intern.
Conf. on Inform. Pro
essing', UNESCO House, pp. 273{282. Reprinted as [Gelernter 1983.
Gelernter H. [1983, Realization of a geometry-theorem proving ma
hine, in J. Siekmann and
G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
',
Vol. 1, Springer Verlag, pp. 99{122. Originally published as [Gelernter 1959.
Gilmore P. [1960, `A proof method for quanti
ation theory: its justi
ation and realization',
IBM J. of Resear
h and Development 4, 28{35. Reprinted as [Gilmore 1983.
Gilmore P. [1983, A proof method for quanti
ation theory: its justi
ation and realization, in
J. Siekmann and G. Wrightson, eds, `Automation of Reasoning. Classi
al Papers on Computational Logi
', Vol. 1, Springer Verlag, pp. 151{158. Originally published as [Gilmore 1960.
Bibel W.
13
14
Martin Davis
15
Index
armative-negative rule . . . . . . . . . . . . . . . . 8
Bibel, Wolfgang . . . . . . . . . . . . . . . . . . . . . . . . 9
Blake, A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
haining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6, 7
Chur
h, Alonzo . . . . . . . . . . . . . . . . . . . . . . . . . 5
onjun
tive normal form . . . . . . . . . . . . . . . . 8
Ents heidungsproblem . . . . . . . . . . . . . . . . . . 9
rst-order logi . . . . . . . . . . . . . . . . . . . . . . . . . 7
Post, E.L. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Prawitz, Dag . . . . . . . . . . . . . . . . . . . . . . . . 8{10
Prawitz, H
akan . . . . . . . . . . . . . . . . . . . . . . . . . 8
Presburger, M. . . . . . . . . . . . . . . . . . . . . . . . . . 5
Prin
ipia Mathemati
a . . . . . . . . . . . . . . . 6, 9
propositional
al
ulus . . . . . . . . . . . . . . . . . . . 6
pure literal rule . . . . . . . . . . . . . . . . . . . . . 8, 10
Putnam, Hilary . . . . . . . . . . . . . . . . . . . . . . . . . 8
Quine, W.V. . . . . . . . . . . . . . . . . . . . . . . . . . . . .9
resolution . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 11
Robinson, Abraham . . . . . . . . . . . . . . . . . . . . 7
Robinson, J.A. . . . . . . . . . . . . . . . . . . . . . 10, 11
rule for eliminating atomi
formulas . . 8, 9
rule of
ase analysis . . . . . . . . . . . . . . . . . . . . .8
Russell, Bertrand . . . . . . . . . . . . . . . . . . . . . . . 6
set of support . . . . . . . . . . . . . . . . . . . . . . . . . 11
Skolem fun
tions . . . . . . . . . . . . . . . . . 7, 8, 10
Skolem, T. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7
splitting rule . . . . . . . . . . . . . . . . . . . . . . . . . 8, 9
subproblems . . . . . . . . . . . . . . . . . . . . . . . . . 6, 7
Tarski, Alfred . . . . . . . . . . . . . . . . . . . . . . . . . . .7
Turing, Alan . . . . . . . . . . . . . . . . . . . . . . . . . . . .5
Logemann, George . . . . . . . . . . . . . . . . . . . . . .8
Logi
Theory Ma
hine . . . . . . . . . . . . . . . . . . 6
Loveland, Donald . . . . . . . . . . . . . . . . . . . . . . . 8
M Ilroy, D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Voghera, Neri . . . . . . . . . . . . . . . . . . . . . . . . . . .8
Wang, Hao . . . . . . . . . . . . . . . . . . . . . . . . 6, 7, 9
Whitehead, A. N. . . . . . . . . . . . . . . . . . . . . . . .6