Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Forml: Research in Advanced Theorem-Proving Techniques

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

RESEARCH IN ADVANCED FORML THEOREM-PROVING TECHNIQUES

Cordell Green
Robert Yates
Bertram Raphael
Charles Rosen

Artificial Intelligence Group


Information Science Laboratory

June 1969
I NTODUCT I

Stanford Research Inst itute is now engaged in a continuing project


to invest igate and develop techniques in art ificial intelligence and to
apply t hem to the cont rol of a mobile automaton. Major topics being
studied under that project include problem solving, visual pattern recog-
nit ion , computer represent at ions, quest ion answering, and system integra-
t ion The problem-solving research thus far has been based largely upon
theorem-proving methods in t he first-order predicate calculus. Cont inued
work is planned under the cited project on improving the eff iciency of
exist ing t heorem-pro.ving techniques and increasing t heir usefulness in
automaton problem-solving situat ions.

As a by- product of this research , some ideas have recently been


developed for a completely new formal mechanical theorem-proving system.
This system would be based upon higher order logic and would emphasize
the role of semantic information and of flexible control strategies.
These ideas are beyond the scope of the main lines of research under the
automaton project at this st
age. However , a successful implement at ion
of the proposed system would be of tremendous value in several research
st udies.
ideas.
fields including automaton Therefore we are here proposing a
major independent effort in the development of these new theorem-proving

OBJECT I VE

This proposal describes a program of research in the development


and applicat ion of advanced formal theorem-proving techniques. The
objective of the proposed work is to design and implement a computer
program with general , powerful , and extremely flexible capabilities for
bot h logical inference and data management. The proposed system would
have several important potent ial uses, which can be explored as part of
t he wo r k :
(1) As a b Bsis for "question-answering systems " large fact
storage and inferent ial retrieval systems;

(2) As a vehicle for research in automat ic problem sol ing


in general , wit h special emphas is on the problem of
automatically writing, verifying, and debugging computer
programs;
(3) As a problem-solving framework for use by the mobile
automaton system being developed at SRI;

*Contract F30602- 0056 , Advanced Research Projects Agency , Depart ment


of Defens ' and Rome Air Deve lopment Ce nter .
(4) As an automated or semiautomated mat hemat ical theorem-
proving system.

I II BACKGROUND

Problem Solving
The development of general methods applicable to the solut ion
of many problems is one of the major goals of current art ificial intel-
ligence research. Many research projects that have been classified as
problem solving, question answering, inferential informat ion storage and
retrieval , theorem proving, program writ ing, and complex informat ion pro-
cessing, all have certain basic underlying similarities. We shall refer
to any such problem that raises the common central questions of representa-
tion , search , and . answer generation as a problem-solving task.
Several approaches to the development of general problem-solving
systems have been suggested and implemented. Three of the most promising
candidates are:
(1) The GPS approach of Newell. This approach is based
upon a pattern-matching capability that ident ifies
differences between objects , followed by the use of
transformat ion operators that potent ially can reduce
those differences . Although pattern matching is a
useful feature , the choice of operators is usually
determined by a table and is thus rather inflexible.

(2) The formal language approach. Fikes, 2 Pople 3 and


others have designed formal languages for describing
problem-sol ving processes. Fikes ' system includes a
search algorithm for evaluat ing procedures expressed
in his language. Such systems , while providing
considerable control , require complex encoding (by
the user) of information in the usual problem state-
ment .

(3) The t heore m-proving approach of Green. 4 Thi s


approach uses informat ion generated during certain
formal. proofs oftheorems in first-order predicate
calculus to construct solut ions to problems.
Although extremely general , t his approach has been
somewhat disappointing in performance because of
the low efficiency of current mechanical theorem-
proving algorithms and the difficulty of intro-
ducing strategic control principles into these
algor it hms .
Each of these problem- sol ving approaches , and perhaps several ot hers
appears to have both advantages and disadvantages. They are all currently

*References are listed at the end of this proposal


being studied and compared at SRI (under the automaton project previously
cited). This present proposal is for a drastic modification of the
theorem-proving approach mentioned above. By designing a new , more
powerful mechanical theorem-proving system with integral pattern matching
and control language capabilities, we hope to establish a new level of
automated problem-solving ability.
B. Mechanical Theorem Proving
The main line of recent theorem-proving research may be traced
from Robinson s landmark paper 6 that introduced a rule of inference
called the resolution principle.
Since then, several research groups
have added theorems and heuristics that result in improved performance
by resolution-based theorem-proving programs for first-order logic. A
recent paper 7 des cribes how one such program aided in the discovery of
a new mathematically significant theorem in lattice theory.

Unfortunately, virtually all exist ing theorem-proving programs


suffer from certain inherent limitations. The major limitations lie in
the following areas:

(l) Semantics Current theorem-proving systems are


essentially pure synt act ic procedures. In complex
problems, however , significant gains in performance
can be made only by focusing attention on the
semantics of the problem domain , and by invoking
strategies that are highly problem-oriented.
mechanism for doing this is presentlyavailable.
(2) Strategies . The major fault of current theorem
provers is their low efficiency--even simple proofs
invariably entail a tremendous amount of computation
that is irrelevant to the final solution. Efficiency
ultimately depends upon the particular strategies
and heurist ics employed to search the space of
possible proofs. New , more efficient strategies are
frequently being discovered or proposed. Unfortunately,
however , the stragegy used by a part icular theorem-
. proving program is generally " frozen " into it s code so
that . newer strategies are extremely difficult , if not
imposs ible , to use.

(3) Level of Logic No exist ing program operates in a


domain of logic significantly beyond the first- order
predicate calculus. However , many interest ing pro-
blems have a natural , compact formulation in higher
order logic. (These include problems of writ ing
computer programs, and of describing the strategies
of the theorem prover itself. Although such formu-
lations can always be reduced to first-order logic
the reduction is awkward , tedious , and usually
obscures the significance of the statements.
The work described be low is aimed at overcoming the above limit at ions.
PROPOSED WORK

During the first year we propose to design and implement a prelim-


inary version of a new theorem-proving program. This version will have
several features not present in other theorem provers and will provide
n basis for future continued expansion and evolution of the system. In
the course of system design; particular attention will be given to the
following research areas:
(l) Data structures and data manipulation. This includes the
representation of objects , sets , and expressions , and the
specificat ion of operat ions including pattern mat ching,
object construct ion and transformat ion , and dat a scanning.

(2) Strategy specification and control.This refers to


methods for specifying, modifying, and monitoring the
search and decision procedures used in the system.

(3) Higher order logic and semantics. We shall try to discover


efficient methods for , extending mechanical theorem-proving
techniques to urorder predicate, calculus, and for embedding
semant ic descript ion and expression evaluat ion into the
basic logical formalism.

In addit ion , we propose to explore potent ial applicat ions for the new
theorem-proving system while it is being developed. One major applicat ion
to be studied under this proposal is the area of computer program writing.
(Applicat ion of this system to robot problem solving will be pursued , when
appropriate , under the separate support of the a omaton project previously
c it ed . )

METHOD OF APPROACH

During the past four years we have developed a series of quest ion-
answering and problem-solving systems based upon formal theorem- proving
methods. As a result of this experience , we now have
(l) A working question-answering program called QA3 , which consists
of a theorem prover based upon Robinson s resolut ion rule for
the first-order predicate calculus and associated search
strategy and data management routines , and

(2) A list of features that have been found desirable but diffi-
cult or impossible to implement in the present system in any
nat ural way.

The QA3 program provides a benchmark for comparison of future systems.


Our initial goal will be to construct a system that matches the basic capa-
bilities of QA3, but is considerably easier to modify and extend.

The list of features represents design goals for the proposed system.
Moreover , these features appear to be highly interrelated. Experienc has
shown that all of these aspects should be present in order to achieve
the desired power in the resulting system.
In each case , our goal will
be to achieve the desired capability in as general and flexible a manner
as possible. These desired capabilities are discussed below.

Pattern Matching and Transformat ions


Pattern matching and transformat ion operat ions consist of
recognizing and naming certain substructures within an expression and
reforming these subparts into new expressions. The system will make
several fundamental uses of pattern-matching capabilities. First , all
strategy operat ions , eval uat ion , and state transformat ions need access
to the properties of expressions; pattern matching offers a convenient
technique for extract ing those propert ies which are dependent upon the
syntact ic structure , of the expression. Secondly, some form of pattern
matching is necessary in the task of comparing expressions for difference
techniques . Finally, most inference rules require a given set of expres-
sions to satisfy some pattern to infer a new expression transformed from
the given set.
Representat ion Changes

These operations will facilitate the use of various repre-


sentat ions of information. For example , suppose that during a complex
problem-solving task three different subprocessors require access to a
certain piece of data. Each subprocessor might require the data to be
represented in a different form.
The three different forms might be
functions , predicates , and sets. The representation operators will be
able to transform the data into the alternate forms.This facility will
also allow experiment at ion with alternate representat ions of informat ion
and with problems t hat require finding new represent at ions.
Ordered Set Operat ions

The purpose of these operations is to manipulate ordered sets


indexed sets ,implicit sets , computed sets , etc.

There should be a set manipulation language describing set


operations. The set operations should include:

Enumerate a set
Enumerate a set according to a gi ven ordering relat ion
Find element s of a
set having specified propert ies
Store a set according to a gi ven ordering re lat ion
Reorder a set and store or enumerate according to the
new relat ion
Ordered set union , intersection , relative complements
Add and delete element s of set
Cre te indices for sets
Find' the extension (closure) of a set under specified
operators
Find the extension of a set under specified
operators , but grow the extension according to a
specified ordering relat ion on the set
Create partitions and covering subsets of sets.

Informat ion storage and retrieval of dat a relevant for inference


is an import ant aspect of a theorem-proving sy stem. Ordered set operat ions
provide a basis for efficient storage and retrieval.

Higher Order Logic

The system will be embedded in the formal framework of w-order


predicate calculus. This higher order system will provide great generality
and expressive power. The synt ax and semant ics of any desired port ion of
t he system it self can be expressed in it s own language. A recent paper
by Robinson 8 developS a foundat ion for extending resolut ion theorem provers
to this domain.

Expression Evaluat ion


The evaluat ion of an expression is t he process of discovering
that the expression denotes a part icular object , or has a particular
object as its value. The concept of the value of an expression is central
to higher order logic since the semantics of the language as well as all
proof procedures are defined in terms of the interpretations (values).
Consequently, most of the problem-solving system is devoted to finding
the values of expressions. In the case where the expression is computable
that is , expressible in terms of computable constructs , then finding the
value reduces to the ordinary process of computation or code evaluation.

One major difference between the proposed system and the usual
theorem-proving " systems is illustrated by the importance to the new
system of expression eval uation , an essent ially semant ic operat iono
Methods called " theorem proving " usually denote a procedure that uses
syntactic rules of inference to establish theorems. The significance of
such methods is that , in an adequate logical inference system such
resolution, the set of theorems coincides with the set of semantically
valid statements. However , in the proposed work we are really only
interested in the semantic validity of a statement , and theorem proving
is merely a tool that allows us to determine validity. In the higher
order system that we are considering, we will directly use semantic
methods for determining validity. Such methods promise to be more effec-
t i ve . However , unt il it is clear that semant ic met hods will complet ely
dominate syntact ic methods , we will of course also carry along a complete
inference system.

Semant ic Descript ion


By semantic description we mean the ability to specify
completely the semant ics of all expressions in the system. Thus , for
example , not only will the system have statement s in a strategy language
which specify procedures , but it will have t he ability to describe and
deal with the effect of carrying out a strategy. Another example of this
not ion is se If-d script ion of t he program s semant ics Anot her example
is a description of an internal representation of data ' which represents
say, a set stored sequent ially according to a given ordering relat ion.
The expressive power of higher order logic lies in its ability to make
statements about other statements, sets of statements , or any functional
port ion of a st atement .
Semantic descriptions will be stated in the language of higher
order logic , and evaluation procedures will be used to find denotations.

Inference
Inference , along with strategies and represent at ions , const i-
t utes one of the major aspect s of theorem proving. The new system will
possess at least the inference capabilit ies of earlier resolut ion- based
systems such as QA3 , although it is expected that gradually semantic
methods will dominate the syntactic inference methods. Such is already
t he case for proposit ional calculus.
We ant icipate t hat the unification process t hat is cent ral
to the resolut ion inference principle will be ext ended , in theory and
practice , to higher order logic , with special facilities for equality
(expressions denot ing the same object).

Inference ope rat ions will ut i 1 ize set operat ion s , pat tern-
match-and- transform operat ions , and higher order logic ope rat ions.
Strategy Operat ions
The purpose of t he strategy operat ions is to create a flexible
system capable of change and self-descript ion. The funct ion of a strategy
is to specify (schedule) the next operation(s) to be executed or attempted
during a process.

There will be a strategy language in which strategies can be


described. The system will include an interpreter that can execute
strategies specifi d in this language. The language will be capable of
concisely expressing complex and interest ing strategies. The strategy
operations that will be included will be selected on the basis of their
utility in describing strategies and their semantics being expressible
rather than by whether they are easily or most efficiently implement able .

Monitoring and Control Operat ions


Monitoring and control operations deal with the flow and pro-
cessing of information within the system and with respect to the external
world. These operations consist of task scheduling, priorities , interrupts
interact iveness , and input-output. The function of the monitor would be
as an overriding control unit which relegates the tasks to be performed
accept s and acknowledges interrupt s , and act s as the 1- 0 controller.
(Informat ion requests and transmission subprocesses would interface
through the monitor. Since many problem-solving tasks have independ-
ent status , it is important to be able to IOnitor the performances of
the individual tasks to determine which subproblems are explored first
and to interrupt control when other subgoals seem more promising.

PRESENT STATUS

The QA3 system is an operational program written in the LISP l.


language and running on the SDS 940 computer at SRI. It has been used
for experimental studies in theorem proving, quest ion answering, robot
problem solving, information retrieval , and reasoning by analogy.

Preliminary work on the new system proposed here is already in


progress. Candidate forms for the strategy, monitoring, set and pattern-
match and transformation operations have been proposed and are being
evaluated. As our next step we plan to implement a version of QA3 in
the new formalism in order to test these choices.

The SDS 940 computer will be replaced by a PDP- lO from Digital


Equipment Corp. by. the end of 1969. The large core memory more powerful
instruction set , and better LISP implementation on the PDP- lO is ex-
pected to result in a vast improvement in the efficiency of our experi-
mental programming work. Although the new computer will be a dedicated
facility purchased by the U. S. Government for the automaton project
we are confident that permission can be obtained to use the system for
t he work proposed here.
VI I PERSONNEL

The following key personnel are expected to participate in the


proposed research.
c. CORDELL GREEN, RESEARCH MATHEMATICIAN
INFORMATION SCIENCE LAORATORY
INFORMATION SCIENCE AND ENGINEERING DIVISION

Specialized professional competence


. Quest ion-answering systems
Problem-solving systems
. Mat he mat ical logic
. Automat ic theorem proving

Representative research assignments joined 1966)


. Design and implement at ion of three quest ion-answering and
theorem- proving computer programs
. Finding first-order logic representation of English language
fact s , quest ions, and answers
. Extension of resolution-type automatic mathematical proof
procedures to const ruct i ve proof procedures for que st ion-
answering applicat ions
. Design of experimental problem-solving system for SRI robot
. Investigation of automatic program writing, verifying, and
debugging techniques
Other professional experience
o Researcher , Synnoet ics Systems (Dr. Louis Fe in), Los Altos
Calif. , 1965
. Engineer , Texas Instruments , Houston , Texas, 1963-
Academic backgrounq
B. S. in electrical engineering (1964), Rice University
M. S. in electrical engineering (1965), Stanford University
Ph. D. in electrical engineering (1969), Stanford University
Publicat ions

The Use of Theorem-Proving Techniques in Question- Answering


Systems coauthored with B. Raphael Proco 23rd Natl. conf. ACM
(Brandon/Systems Press , Inc. , Princeton , N. J., 1968)
Theorem Proving by Resolution as the Basis for Ques.tion-
Answering Systems Machine Intelligence 4 , Michie , ed.
(Edinburgh University Press , Edinburgh , Scot land , 1969)
o " Appl icat ion of Theorem Proving to Problem Solving, " Proc.
Int ' l. Joint conf. on Art if icial Intelligence Washi ngton
C. (l969)
Professional associat ions and honors

. Associat ion for Comput ing Machinery


ute of Electrical and Elect ronics Engineers
. Inst it
. Tau Beta Pi
. Sigma Tau
NILS J. NILSSON , SENIOR RESEARCH ENGINEER
INFORMTION SCIENCE LAORATORY
INFORMTION SCIENCE AND ENGINEERING DIVISION

Specialized professional competence

. Artificial intelligence
. Systems theory
. Pattern recognition

Representative research assignments at SRI joined 1961

. Studies in the theory of pattern recognition


. Feature detection studies
. Studies in heuristic search procedures
. Automatic theorem- proving studies
. Planning, promotion , and direction of robot sys ems research
. Head , Artificial
Intelligence Group (1963-67)

Other professional experience

. Taught courses in pattern recognition at Stanford Uni versi ty


and at University of California , Berkeley, 1962-63
Acting Associate Professor at Stanford Uni versi ty, Computer
Science Department (one-half time , 1968-69)

Academic background
S . in electrical engineering (communication theory) (l956),
Stanford University
. Ph. D. in electrical engineering (communication theory) (l958), S.
Publ ica t ions
. Twelve articles on pattern recognition and artificial intelligence
. Learning Machines (McGraw- Hill , 1965)

Professional associations

. Institut of Electrical and Electronics Engineers


. Association for Computing Machinery
. Tau Beta
. Sigma Xi
BERTRA RAPHAEL t SENIOR RESEARCH MATHEMTICIAN
INFORMTION SCIENCE LABORATORY
INFORMTION SCIENCE AND ENGINEERING DIVISION

Specialized prof ssional competence


Question-answering systems
o Heuristic problem solving
o Symbol manipulation techniques
o Theorem- proving methods
Representati ve research assignments at SRI joined 1965

. Development of data structures and deduct i ve techniques for


on-line question-answering systems
o Studies of problem- solving act ivity in a simulated robot
o Direction of system design for an experimental " intelligent
automaton
o Survey of computer languages for symbolic and algebraic
manipulation
Other professional experience

o Lecturer ,Electrical Engineering and Computer Science , University


of California at Berkeley; Lecturer , Computer Science , Stanford
University; and Instructor , summer course at University of
California at Los Angeles
o Consultant , Computer Science Department , RAND Corp 0, Santa
Monica , California
o Assistant Research Scientist , University of California at
Berkeley
o Part-time research staff , Bol t , Beranek and Newman , Inc
Cambridge , Massachuset ts

Academi c background
. B oS 0 in physics (l957), Rensselaer Polytechnic Insti tute
. Mo So in applied mathematics (l959), Brown University
. Ph. Do in mathematics (1964), Massachusetts Institute of
Technology
Publications
o More than a dozen papers in technical journals and in the
proceedings of national and international computer conferences
Professional associations
o Association for Computing Machinery (National Lecturer 1967-68 ;
founding ditor , Newsletter of the group on artificial
intelligence)
o Association for Computational Linguistics
o Sigma Xi
CHARLES A. ROSEN , MANAGER , ARTIFICIAL INTELLIGENCE GROUP
INFORMTION SCIENCE LABORATORY
INFORMATION SCIENCE AND ENGINEERING DIVISION
Specialized professional competence
. Artificial intelligence
. Pattern recognition
. Solid- state devices (especially piezoelectri c)
. Electron physics
Representative research assignments at SRI joined 1957
. Developed the Electron Physics Group and the Artificial
Intelligence Group, as group manager
. Development of mobile automaton system
. Pattern- recogni tion and learning-machine studies and applica tions
. Development of microelectronic devices and systems
Other professional experience
. Assistant Head , Transistor Circui t Group; Head , Dielectric
Devices Group; Consulting Engineer , Dielectrics and Magnetics
General Electric Company
. Manager of Radio Department and Spot Weld Department , Fairchild
Aircraft , Canada
. Technical investigations for radio and instruments , Bri tish Air
Commission
. Co- owner , Electrolabs Reg ' d. , Montreal , Canada , Alarm Intercom
Systems
. Lecturer , Stanford University, piezoelectric and ferroelectric
devices
Academic background
. B. E. E. (1940), Cooper Union Insti tute
of Technology
. M.Eng. in communications (l950), McGill University
. Ph. D. in electrical engineering (minor in solid- state physics)
(l956), Syracuse Universi
Publications and patents
. Coauthor of Principles of Transistor Circui ts , R. F. Shea
editor (John Wiley and Sons , Inc. , 1953)
. Coauthor of Solid State Dielectric and Magnetic Devices , H. Ka t z ,
editor (Joh n Wile sons , Inc. 195
. Author or coauthor of several papers in the fields of piezo-
electric devices , learning machines , pa ttern recogni tion
. Six patents relating to solid- state devices
Professional associations
. Senior Member of the Inst i tute of Electrica 1 and Electroni cs
Engineers
. Member of the American Physica 1 Society
. Member of the Scientific Research Society of America
RICHAD J. WALDINGER RESEACH MATHEMATICIAN
INFORMATION SCIENCE LAORATORY
INFORMATION SCIENCE AND ENGINEERING DIVISION

Specialized professional competence

. Art ificial intelligence


. Automata theory
. Log i
. Recursive funct ion theory
. Automat ic theorem proving

Other professional experience

. Research assist ant , teaching assist ant , project scient ist


Carnegie-Mellon Uni vers i ty , Pitt sburgh , Pennsylvania

. Mat he mat ician , Heurist ics Laboratory, Nat ional Inst i tute of
Health , Bethesda , Maryland

Academic background
. A. Bo in mathematics (l964), Columbia College , New York, No Y.

. Ph. D. in computer science (1969), Carnegie-Mellon University,


Pi t t sburgh, Pennsylvania
Publications
PROW: A Step toward Automatic Program Writing,
Internat ional Joint Conference on Art ificial Intelligence,
Washington , D. C. (1969)
ROBERT A. YATES , SYSTEMS PROORAER
INFORMTION SCIENCE LAORATORY
INFORMTION SCIENCE AND ENGINEERING DIVISION

Specialized professional competence

Mathematics
Programming languages
. Compi 1 ers
. Diffraction

Representative research assignments at SRI joined 1967

. Design and development of a question-answering computer system


based on first-order predicate calculus

Other professional experience

Member of technical staff , Bell Telephone Laboratories , Holmdel,


New Jersey; work on design and implementation of SNOBOL4
programming language
. Programmer , Johns Hopkins Universi ty, Physics Department
. Programmer , Poli tecnico Mexico City; design and implementation
of LISP system and compiler

Academic background
A. in mathematics (l965), Johns Hopkins Uni versi ty
A. in mathematics (l967), ' Stanford University

Professional associations

. Phi Beta Kappa


REFERENCES

G. Ernst and A. Newell Generality and GPS " Technical Report


Carnegie Institute of Technology, Pittsburgh , Pa. (January 1967).

2. R. E. Fikes St at ing Problems as Procedures to a General Problem


Solving Program Proc. Fourth Systems Symposium , Case Western
Reserve University, Cleveland , Ohio , November 1968 (to be published).

3. H. E. Pople , Jr., "A Goal-Oriented Language for the Computer


Ph. D. Thesis, Graduate School of Industrial Administrat ion , Carnegie
Mellon University, Pittsburgh , Pa.
(1969).
4. C. Green Theorem Proving by Resolut ion as a Basis for Quest ion-
Answering Systems, Machine Intelligence 4 , Michie and Meltzer , eds.
(Edinburgh University Press, Edinburgh , Scotland , 1969)

C. Green App1icat ion of Theorem Proving to Problem Solving, Proc


Internat ional Joint Conference on Art ificial Intelligence , Washington
C. (May 1969).
J. A. Robinson A Machine-Oriented Logic Based on the Resolution
Principle " J. AcM, Vol. l2 , No. , pp. 23- 41 (January 1965).

J. R. Guard , F. Co Oglesby, J. H. Bennett , and L. G. Settle


Semi- Automated Mathematics, , Vol. 16 , No. , pp. 49-62
(January 1969).

J. A. Robinson Mechanizing Higher Order Logic Machine Intelligence


, Michie and Melt zer , eds. (Edinburgh Uni vers i ty Press, Scot land,
1969) .

You might also like