DISCOVERY AND REASONING IN MATHEMATICS'

Alan Bundy

Department of Artificial Intelligence,

University of Edinburgh

Abstract ematicians. The main theme will be that math-

We discuss the automation of mathematical ematical reasoning consists of more than just
reasoning, surveying the abilities displayed by human theorem proving, and that the simultaneous automa-
mathematicians and the computational techniques tion of other reasoning processes, e.g. learning,
available for automating these abilities. We argue simplifies the automation task. Each reasoning
the importance of the simultaneous study of these process outputs knowledge but also demands
techniques, because problems inherent in one tech- knowledge as input. If a computational technique
nique can often be solved if it is able to interact for automating a particular reasoning process is
with others. studied in isolation then the provision of its input
knowledge may prove a major barrier to automation.
But the input knowledge of one technique is the
Keywords output knowledge of another, so that the tech-
Reasoning, mathematics, deduction, search, learning, niques f i t together in an intercommunicating net-
proof analysis. work. Some techniques for automated reasoning may
involve search. In isolation the search involved may
be computationally expensive. But one technique
may be used to control the search of another,
1. Introduction especially if the two techniques are co-routined or
A major goal of artificial intelligence is to even merged into one. The power of a system in
automate reasoning. Solutions to this goal will have which the various techniques interact in well-crafted
both technological applications, enabling us to build ways will be more than just the sum of the power
more powerful expert systems, and scientific ap- of the parts.
plications, providing models to compare with human
reasoning. Mathematics is an excellent domain for I advocate the simultaneous study of all the the
exploring the automation of reasoning because: processes of mathematical reasoning with particular
(a) it provides a wide range of examples of emphasis on the possible interactions between the
reasoning, from the simple and shallow techniques that automate them.
to the complex and deep;

(b) it is possible to detach this reasoning 2. Reasoning Abilities in Mathematics

from other considerations, such as sen- What processes are involved in mathematical
sory input of data; and, reasoning? In this section we discuss the various
processes and the AI techniques that have been
(c) to a first approximation, the problems proposed to automate them. More details about
of knowledge representation have been these techniques can be found in [Chang &, Lee
solved. 73. Bundy 83].

Compare mathematics with other AI domains in which The first problem we face is how to classify
reasoning plays a part, e.g. natural language under- mathematical reasoning into different processes
standing or visual perception. In mathematics one is - what my psychology friends call defining the task
not bogged down with huge amounts of noisy data, ecology. Our starting point is a 'folk ecology' of
nor concerned that what seems to be a difficult reasoning processes, i.e. the terms used in pre-
reasoning issue may evaporate if the knowledge scientific discussion of reasoning. We will call these
representation were changed. One is free to con- mathematical abilities. With these abilities we must
centrate on the reasoning problems and then make associate computational techniques which implement
an attempt to translate any solution found to them. These techniques will form our scientific
other domains. classification. But the association of abilities to
techniques not a neat 1-1 mapping. We are likely
In this paper I discuss the automation of math- to find many techniques to associate with each
ematical reasoning, by which I mean any cognitive ability. Some abilities, e.g. learning, have so many
activity that mathematicians engage in as math- diverse techniques associated with them that they
seem highly unsuitable as scientific categories,
whereas the techniques associated with some
This research was supported by SERC grant. number abilities, e.g. theorem proving, all have a strong
GR/C/20826.
1222 A. Bundy

family resemblance. Some abilities require a com- using the above techniques.
bination of techniques. Consider, for instance, anal-
ogy which requires separate techniques for finding - Using Analogy: Mathematicians use analogy
and then using the analogy. Each of these tech- to suggest conjectures and new defini-
niques will require sub-techniques. Some of these tions, and to guide proofs. All uses
techniques may contribute to several different require analogical matching techniques to
abilities. Consider, for instances, resolution-type find and apply the relationship between
deduction techniques, which are ubiquitous. the target and the source of the analogy
(see [Owen 85] for a survey). This is all
We list below the mathematical abilities we will be that is required for suggesting conjec-
considering under the heading of mathematical tures and definitions, but to use a
reasoning and mention some of the techniques which source proof to control the search for a
have been used to implement these abilities. target proof a further proof plan ap-
plication technique is required.
- Proving theorems: The ability that has
received the most attention in the AI This is by no means an exhaustive list; math-
study of mathematical reasoning, almost ematicians also find counterexamples, write, publish
to the exclusion of all others, has been and deliver papers, teach students, read textbooks,
theorem proving. Automated theorem etc. Unfortunately, we have nothing to say about
proving has been an important subfield of these other abilities, so they are omitted.
AI throughout its history. The main
techniques required to automate theorem
proving are deduction which involves 3. Mathematical Reasoning Techniques
search control. Deduction traverses a In this section we classify and discuss some of
search space of legal inference steps. the techniques that have been developed for
Search control decides which of these automating the mathematical abilities described in
steps to try. the last section. There is only space to discuss
those we consider particularly promising. We group
- Formalizing Problems: However, math- these techniques according to their computational
ematicians, especially applied math- purpose and mutual similarity, e.g. deduction,
ematicians, spend a large part of their search control. inductive inference. etc. This
time translating informal problem state- provides a more scientific classification of math-
ments into mathematical formulae to ematical reasoning processes than the list of math-
which deduction, etc. may be applied. ematical abilities above. But even this classification
The initial stages of this translation in- is bound to be improved as we discover new tech-
volve natural language understanding, visual niques and gain a better understanding of the ex-
perception, etc - but these are not isting techniques and their inter-relationships. It
specifically mathematical techniques. should therefore be regarded as both incomplete
Later stages involve the specifically math- and highly preliminary.
ematical technique of formula extraction
from the internal meaning representation,
which may itself involve deduction (see
section 3.5). 3.1. Deduction
The most well known deduction technique is
- Learning: This term covers a multitude of resolution, [Robinson 65]. Resolution is a rule of
processes, for instance, the learning of inference for Predicate Calculus, that is. it is a
new mathematical theorems, the learning rule for deducing new logical formulae from old.
of new proof methods, the defining of To prove that a conjecture is a theorem of a
new concepts, the conjecturing of mathematical theory both the axioms of the theory
results, etc. New theorems must be as- and the negation of the conjecture are expressed
similated into the theorem proving ability as clauses (a normal form for Predicate Calculus
so that they may be used effectively in formulae). Resolution takes a pair of clauses and
the future. The assimilation technique makes subparts of each identical using a matching
required will depend particularly on the technique called unification. The remaining parts are
search control technique being used, e.g. combined together to make a new clause. Resolu-
the new theorem will need to be labelled tion applied repeatedly to the initial clauses and to
so that it can be accessed when needed. their successors defines a search space of clauses.
Similarly, new proof methods must be in- This is searched for a contradiction. If the search
corporated into the current search con- is successful then the conjecture is a theorem.
trol technique. This may involve a proof
analysis technique to analyse new proofs The representation of axioms and conjectures as
and generalize them to extract control logical formulae is given in the texts of math-
information, e.g. proof plans. Defining ematics and logic. It is in this sense that I
new concepts can be done by inductive claimed that the knowledge representation problem
inference from descriptions of 2examples
and non-examples of the concept. Con-
jecturing of theorems can be done by Note that inductive infaranca it not tha same as math-
considering the hypothesis of the conjec- ematical induction (••• •action 3.2 below), which ia a
ture to be a concept to be defined daductive rule of inference.
 A. Bundy 1223

was solved to a first approximation.

The resolution search space is guaranteed to con-

tain a contradiction if and only if the conjecture is
indeed a theorem. The if part is called complete-
ness and the 'only if part is called soundness. The
snag is that if the conjecture is not a theorem
then the search may never terminate. The number
of new clauses generated rises exponentially, or
worse, with the length of the proof. If the proof
is non-trivial then either the storage capacity of
the computer or the patience of the human Note that this precondition is too general; trig
operator is exhausted before the proof is found. includes tan, for which the rule does not work.
This is an example of the phenomenon called the We discuss this problem in section 3.3 below.
combinatorial explosion.
Note that these preconditions can get very com-
Various refinements of or alternatives to resolu- plicated since the same rule may be used success-
tion have been designed and implemented, with the fully in a number of different circumstances and
aim of improving on its space and/or time efficiency the precondition must be the disjunct of these
without losing completeness. One of the most circumstances. Note also that the rule now con-
powerful alternative deduction techniques is the tains a mixture of factual and control knowledge.
Connection Calculus, [Bibel 82]. These improvements These complications can cause difficulties, e.g. in
reduce the combinatorial explosion, however, but do the automatic learning of such rules, so some
not conquer it. researchers prefer to separate factual and control
knowledge.
There have also been attempts to avoid the com-
binatorial explosion by proposing radically new For instance, the Boyer/Moore Theorem Prover,
deductive techniques. This has sometimes led to [Boyer & Moore 79], represents its control
techniques which are efficient at proving restricted knowledge as procedures for manipulating the math-
classes of theorems, e.g. rewrite rules [Huet & ematical formulae. This theorem prover exploits the
Oppen 80]. It has also led to the reinvention of relationship between recursion and mathematical in-
the wheel - some researchers have rejected the duction to guide inductive proofs of the properties
logical approach, only to reinvent it together with of Lisp functions. The recursive definitions of the
the attendant problems of incompleteness and/or Lisp functions are first used to symbolically evaluate
combinatorial explosion. the conjecture to be proved. This may fail because
the conjecture contains arbitrary constants and the
evaluation process requires lists with internal struc-
Rather than reject logical deduction techniques it ture. This failure is used to guide the choice of
is necessary to augment them with heuristic tech- induction scheme and variable, so that when symbolic
niques to control the search for a proof, [Hayes evaluation is applied to the induction conclusion just
77]. Search control techniques developed have enough structure will be available to enable the
ranged from weak but general purpose ones, e.g. induction step to be performed. We will call this
evaluation functions based on the complexity of the technique, recursion guidance.
formulae, to powerful but special purpose ones,
such as those described in the next section. For a
survey see [Bledsoe 77]. Note that recursion guidance involves an analysis
of the conjecture and its failed proof, and the
choice of an appropriate proof technique on the
basis of this analysis. This analysis uses meta-
3.2. Search Control concepts4 to describe the conjecture, its proof and
A number of powerful search control techniques the proof methods, e.g. the terms induction
have been developed by careful analysis of the hypothesis, induction step, etc in the Boyer/Moore
proofs of human mathematicians to extract the un- Theorem Prover or trigonometric function in LEX.
derlying control ideas and express them computa- These meta-concepts must be discovered by study
tionally. of existing proofs, introspection, etc. It is in this
sense that the knowledge representation problem is
The simplest technique, called heuristic rules, is to only solved to a first approximation, since the num-
add these control ideas into the mathematical for- ber of possible meta-concepts is open ended and
mulae as preconditions for their application. For the choice of appropriate ones determines the suc-
instance, the LEX program, [Mitchell et a/ 81], uses cess or failure of the technique.
rewrite rules to symbolically integrate algebraic
terms, e.g the rule:

Or an additional rule uaad for aach circumtances.

is used to integrate by parts. One class of terms 4

Tha term meta' ia uaad bacause tha concepts in quaation
which this rule successfully integrates is those daacriba the rapraaantation of the problem, i.a. they ara
where u is a variable, x, and dv/dx is a constant about it rathar than of i t .
1224 A. Bundy

This process of analysis and guidance using meta- applied and how this reason fitted into the overall
concepts is itself a reasoning process and can be proof plan. For instance, sin u.cos u = 1/2.sin 2u
conveniently represented as deduction - but deduc- was applied to reduce the occurrences of u from 2
tion in a meta-theory, rather than the theory it- to 1 in order that that single occurrence could be
self. Such a use of deduction to guide deduction isolated on the left hand side of the equation.
is made explicit in the PRESS system, [Bundy & Wel- This kind of analysis is performed by the LP
ham 81], for solving equations. The PRESS meta- program. [Silver 84], which was able to learn new
concepts describe the equation to be solved, e.g. methods of equation solving and extend the range
the number of unknowns, their distance apart, the of problems that PRESS could solve. The technique
kind of functions involved, etc, and they describe used by LP is called Precondition Analysis because it
various methods of solving equations or of achieving discovers new methods by considering how un-
useful subgoals, e.g. reducing the number of un- explained steps establish the preconditions of suc-
knowns, moving them closer together, making them cessive steps.
occur within identical subterms, etc. Deduction
with these meta-concepts induces an implicit, but In the case of the LEX program for symbolic
highly controlled, search for the solution to an integration (also described in section 3.?) the tech-
equation. We call this technique, meta-level in- nique of back propagation was developed to analyse
ference (see figure 3-1). successful solutions and extract the control infor-
mation. This technique was incorporated in the LEX2
program, [Mitchell et al 833. In back propagation
the successful sequence of rewrite rules was applied
in reverse order to a generalized answer to see
what constraints this would impose on the original
problem. These constraints then became the precon-
ditions of the first rule of the sequence. For
instance, given a successful integration of cos7(x),
LEX? used back propagation to discover that the
same sequence of rules n would have worked on any
term of the form cos (x). where n was an odd
integer. The first rule of this sequence,

These meta-concepts can also be used to express was then given the precondition, f-cos &, odd(r). In
proof plans, e.g. the plan to move two unknowns forming this constraint. back propagation also
closer together, then merge them and then isolate defined the concept. odd* as any number of the
the remaining occurrence. This plan might be form 2k+ 1, where k is an integer.
provided by the programmer or learnt by proof
analysis of a worked example (see section 3.3). Al-
ternatively the proof plan may be at the object- Proof analysis can also be used to analyse faulty
level, e.g. the proof of an analogous theorem, or a proofs and repair them. For instance, a classic
generalized proof. [Plummer & Bundy 84]. In each faulty proof in the history of analysis is due to
case it must then be applied by a proof plan Cauchy. namely that a convergent series of con-
application technique. This application might be tinuous functions is continuous. This proof can
straightforward - the target proof exactly following be analysed using deduction and the fault identified
the plan - or the plan may need to be relaxed or as a missing occurs check during unification (see
augmented at various points. A variety of tech- section 3.1. This suggests an obvious patch, namely
niques have been suggested for realising such a changing the order of quantifiers in the theorem
flexible plan application technique (see, e.g. [Silver statement, and this generates three new concepts
84]) and three new and correct theorems. One of these
is the traditional replacement of convergence by
uniform convergence; one is a trivial theorem whose
3.3. Proof Analysis conclusion is always true; and one is a new theorem
The process of constructing and augmenting involving the concept of equi-continuity. This
powerful search control techniques, by adding meta- reasoning process, which we will call argument
knowledge gained from analysing proofs, can itself removal, has been implemented in a program, SEIDEl ,
be automated. This analysis will depend crucially on described in [Bundy 85].
the search control technique in question. For in-
stance, in the case of the PRESS system, described
in section 3.2. the solution to an equation must be
analysed using the meta-concepts of PRESS. For each
step an account must be given of not only what It would have been more powerful to remember the whole
algebraic identity was applied but also why it was sequence as a proof plan and make the constraints the
preconditions of the plan rather than just its first step.

In contrast to the deduction of the equation solution, which See [Lakatos 76] for a discussion of this proof and its
is called object-level inference. history.
 A. Bundy 1225

For a survey of the use of proof analysis for the unknowns and givens of the problem. The vari-
learning see [Boswell 85]. ables are then instantiated by inferring the logical
description from the meaning of the English problem
statement.
3.4. Inductive Inference
Inductive inference has received a lot of atten- This inference process may use deduction and
tion in the learning literature, with a number of default reasoning to fill gaps between the state-
techniques being developed. For a survey see ment of the problem and the knowledge required to
[Dietterich et al 82] and for an analytic com- instantiate the law. MECHO used resolution guided
parison of some of these techniques see [Bundy et by meta-level inference for the deduction and the
al 833. They all learn a concept from examples and, closed world assumption for the default reasoning.
sometimes, non-examples of it. From examples of Deduction was needed, for instance, to work out
the concept the techniques might generalise: the contributions to the sum of forces from
replacing specific relations and terms with more gravity, the tensions of strings, the reactions of
general ones. From non-examples of the concept contact surfaces, etc. Default reasoning was used,
the techniques might discriminate: pruning away rela- for instance, to assume that the only surfaces in
tions and terms that are not an essential part of contact were those mentioned in the problem
the concept. statement.

This can be used for the learning of both new

object-level (factual) and new meta-level (search 4. The Interaction of Reasoning Techniques
control). For instance, Mitchell et al's LEX program, The mathematical reasoning techniques outlined
[Mitchell et al 81], used induction to learn the above can interact in a variety of ways. For
meta-level conditions for applying a particular instance, successful deductions might provide the
strategy for integrating by parts. Given that the material for proof analysis and for analogy. Proof
strategy worked correctly to integrate 3x.cos(x) and analysis may suggest proof plans to aid deduction.
5x.sin(x), LFX generalized these terms to hypothesise The desire to make a particular theorem hold may
that the strategy will work for terms of the form trigger a process of proof analysis that leads to
nx.trig(x), where n is any integer and trig any changes in definitions and axioms. Sometimes these
trigonometric function. interactions can be more intimate. Several deduction
techniques may be co-routined or even merged so
Both the sub-techniques of generalization and dis- that each prunes the search space of the others.
crimination are crucially dependent on the descrip- We discuss some of these possibilities in more detail
tion space. For instance, in LEX the description below.
space contains concepts like 3, 5, n, cos, sin, trig,
etc. The generalization of two concepts is the
most specific concept in the description space 4.1. The Interaction of Techniques in the PRESS
which includes both of them, e.g. trig is the Family
generalization of cos and sin. However, if a term In my research group work has continued over a
for sine or cosine" were added to the description number of years on a family of programs working
space then it would replace trig as the generaliza- on the common domain of symbolic equation solving.
tion of cos and sin. In fact, this is the concept
required, as discussed in section 3.1. Without it,
LEX overgeneralises to trig. - As described in [Bundy &, Welham 81] and
section 3.2 above, the PRESS program
used the deduction technique of rewrite
rules to generate solutions to equations.
3.5. Formula Extraction This deduction technique was guided by
A key technique in the mathematical formalisation the search control technique of meta-
of problems is the extraction of formulae from a level inference.
meaning representation, e.g. the semantic represen-
tation of an English description of the problem. - The LP (Learning PRESS) program (see
Formulae extraction techniques have received little [Silver 84] and section 3.3) used the
attention in Al despite their importance in human proof analysis technique of precondition
mathematical reasoning. However, the MECHO analysis to extract and conjecture new
program, [Bundy et al 79], contained a formula ex- methods of solving equations which were
traction technique called the Marples algorithm, for then used by PRESS.
forming equations to describe mechanics problems.
- The IMPRESS (Inferring Meta-knowledge
The Marples algorithm is a kind of plan formation about PRESS) program, [Sterling & Bundy
technique. Equations are formed by instantiating 82], was a theorem proving program for
physical laws, e.g. F=MA. With each law is stored a proving properties of logic programs. It
list of the things it is about and a logical descrip- was used to prove properties of the
tion of how these things relate to the variables in Prolog code of PRESS using a modified
the law. For instance, F=MA is about an object and version of recursion guidance (see section
a direction. The variable M is the mass of the 3.2). For instance, it proved the correct-
object, A its acceleration in that direction, and F ness of some PRESS equation solving
the sum of the forces acting in that direction. A methods, i.e. that under appropriate
law and a situation are chosen after an analysis of preconditions the methods would achieve
1226 A. Bundy

their goal. gence.

- As described in [Bundy et al 79] and sec- But proof analysis techniques cannot merely
tion 3.5 above, the MECHO program solved replace inductive inference techniques as learning
mechanics problems stated in English by processes. Proof analysis techniques work only on
extracting and solving equations using the single examples and this limits the amount of
Marples algorithm. The equation solving generalization that they can do unaided. For in-
part was done using PRESS as a sub- stance, we saw that back propagation was able nto
program generalize a particular problem from cos x to cos x,
where n is an odd integer. 7 is generalized to n
Each of the above techniques is incomplete on its by considering the constraints forced by the par-
own. Meta-level inference requires a rich supply of ticular sequence of rules used in the successful
meta- level concepts to analyse problems and bring solution. However, a similar sequence will also in-
to bear appropriate methods of solution. These tegrate terms of the form sinnx where n is an odd
meta-level concepts can be extracted from example integer; a rule for cos needs to be replaced by a
solutions by precondition analysis. The new methods similar rule for sin. To recognise this similarity and
conjectured by precondition analysis can be shown build a general proof plan for both cases requires
to be correct using recursion guidance. The ex- inductive generalization, [Boswell 84]. Alternatively,
ample solutions required by preconditions can be one might use analogical matching to recognise the
supplied by deduction with a less constraining search similarity and a flexible proof plan application tech-
control. The problems to be solved by deduction nique to apply the cos sequence to the sin problem.
can be supplied by the Marples algorithm, but this Note how back propagation narrows down the search
technique requires deduction together with meta- which would otherwise be involved in finding an
level inference to bridge gaps between the analogous solution, but then uses analogical matching
knowledge it requires and that provided in the to further narrow its own search. Therefore,
problem statement. The interactions are summarised techniques of proof analysis, inductive inference
in table 4-1. 8 analogical matching and proof application need to
work in concert to achieve maximum learning power.

4.2. The Interaction of Learning Techniques

In section 3.4 we described the importance of
the description space in constraining the kinds of
inductive inference that were possible, old concepts
can only be generalized or specialized to new con-
cepts that are contained in the description space.
In section 3.3 we described how the techniques of Faced with formulae A -> B' and B" -> C the
back propagation and argument removal could be resolution rule applies substitutions to either for-
used to define new concepts and thus extend the mula in an attempt to make the Bs identical so
description space in a principled way by adding a that cut can be applied. Unification (see section
needed concept, e.g. odd integer, uniform conver- 3.1) will find the most general such substitution.
Up to renamings of variables this substitution is
unique - a far cry from the infinite branching of
undirected substitution. Thus the merging of modus
ponens and substitution controls the search implicit
8,Not all these interactions have been implemented. in the later rule by making its application subser-
 A. Bundy 1227

VIent to the former. trying to solve, faulty proofs it is trying to cor-

rect or successful solutions it is trying to
Similar mergings of other deduction techniques generalise. Its sense of direction comes entirely
have been suggested. For instance, some axioms from its evaluation function which is guided by the
have been built-in to the unification algorithm. In patterns, coincidences, etc that it notices in its
associative resolution, [Plotkin 72], the associative example finding. It would be interesting to link
axiom for a function, f, is deleted from the set of AM's techniques to those outlined above to get a
axioms and built-in to the unification algorithm, i.e. better directed process of mathematical discovery.
it can be used in the attempt to match two This would involve separating the different tech-
expressions. This merging of unification and as- niques used in AM and implementing a wider inter-
sociativity has the effect of controlling the ap- action of mathematical techniques. However, such a
plications of associativity by making it subservient programme would not be easy. The techniques used
to unification. Other examples are higher order by AM are not clearly explained, are embedded in
unification, [Huet 75], which builds the axioms and complicated Lisp code, and are difficult to dis-
rules of lambda calculus into the unification algo- entangle from the heuristic rules.
rithm, and E-Resolution, [Morris 69], which builds the
equality axioms into the unification algorithm.
5. The DReaM Programme
The advantage of such mergings is not just in the In this paper I have advocated the simultaneous
shrinking of the search space; they can also assist study of a number of different techniques for
the application of proof plans by bringing the key mathematical reasoning, especially how these tech-
steps of a proof to the top of the search space. niques may be fitted together. I believe that
A proof plan may identify a particular step as a problems associated with the individual reasoning
key one, but it might take several minor steps to techniques can often be solved by combining them
transform the problem into a state where this key together, and I gave a number of examples of this
step can be applied. For instance, the key step phenomenon in section 4 above.
may be to resolve with a particular clause, but
several applications of associativity may be required To realise these ideas I have instituted the DReaM
to allow the resolution to go through. These minor (Discovery and Reasoning in Mathematics) Programme
steps may create a combinatorial explosion of their at the Department of Artificial Intelligence at Edin-
own before the key step is reached. By making burgh University. This programme gives explicit
the minor steps subservient to the key one the key recognition to an implicit programme of development
step can be applied first and the application of of mathematical reasoning programs over a period of
the minor ones can be controlled. several years. Figure 5-1 explains the relationship
between the various programs built in our group
Merging may also be applicable to non-deduction during this period. Each node is a program and
techniques. For instance, given a problem to solve, each arc represents some historical dependence of
an analogical matcher might be able to find a the earlier program upon the later one. PRESS, LP,
similar solved problem in order to use its solution
as a proof plan. Sometimes a solved problem will
match the given one in several different ways.
Rather than work through each way in turn, the
match may be left incomplete until further instan-
tion is required to continue with the proof plan
application. Thus the analogical matching will be
made subservient to the plan application technique
and its search thus controlled.

4.4. Lenat's AM Program

The AM program, [Lenat 8?], is an interesting ex-
periment in the interaction of a number of tech-
niques for finding examples, defining concept,
making conjectures, etc. These are set in a
framework of heuristic rules controlled by heuristic
search. An evaluation function is used to decide
what concepts to define or find examples of, what
conjectures to make. AM's performance is impres-
sive; starting with some simple set-theoretic con-
cepts, it defines some relatively complex and inter-
esting concepts, e.g. prime numbers, and makes
some interesting conjectures, e.g. the prime unique
factorization theorem. It also defines a lot of
uninteresting concepts and makes a lot of silly con-
jectures.

AM has no theorem proving ability. Its definitions

and conjectures are not motivated by problems it is
1228 A. Bundy

niques for search control, proof analysis, inductive

ecological modelling statistics inference, matching, formula extraction, etc. In
packages, respectively. this paper I have outlined some of the most
- RUT, [Plummer 84], is a rational promising such techniques drawn from the work of
reconstruction of Bledsoe's natural deduc- my own group and from that of others.
tion theorem prover, [Bledsoe 77], and
VOYEUR, [Plummer 8, Bundy 84], extends I have given examples of the interactions of
RUT with the gazing technique described these techniques and shown how these interactions
below. can solve problems which can appear insuperable if a
technique is studied in isolation. This constitutes a
Our specific, short term objectives are to extend strong argument for the simultaneous study of
our existing reasoning techniques and invent new reasoning techniques; to see how the total can be
techniques in a variety of domains, and more more than the sum of the parts.
centrally to investigate the interaction of: deduc-
tion, search control, proof analysis and inductive The DReaM project aims to conduct such a simul-
inference within a single reasoning system. The taneous study. Some of the preliminary results are
understanding gained from this investigation will be reported above together with our plans to incor-
exhibited in a program for reasoning primarily in porate several techniques within a single system.
mathematics, but adaptable (we hope) to other
forms of problem domain.

The core of the system will be the MT program, Acknowledgements

[Wallen 83], which consists of two parts I am grateful to Robin Boswell, Steve Owen, Lin-
coln Wallen and Mike Uschold for assistance with
- the object-language: a logic for express- and/or feedback on this paper.
ing problems; and

the meta-language: a logic for expressing

proof plans. References
Bibel W.
Automated Theorem Proving.
MT uses a process of meta-level inference to Friedr. Vieweg & Sohn,
analyse a conjecture, choose an appropriate proof [Bibel 82]
Braunschweig/Wiesbaden. 198?.
plan and use it to guide the search for a proof.
Bledsoe, W.W.
The object-level deduction technique is based on Non-Resolution theorem-proving.
Bibel's Connection Calculus. [Bibel 82]. The Connec- Artificial Intelligence 9(1): 1-35,
tion Calculus is particularly suitable as a vehicle for [Bledsoe 77] August. 1977.
proof plans as it does not demand that the conjec-
ture be put in a normal form and the proof is Boswell R.A.
constructed using a detailed analysrs of the conjec- Further Developments to an Algebra
ture. No new formulae need to be constructed Learning Program - A Thesis
during the proof. This makes it particularly easy to [Boswell 84] Proposal.
relate the original analysis of the conjecture to the Working Paper 174. Dept. of Artifi-
proof plan and hence to the subsequent proof. We cial Intelligence, Edinburgh, Oc-
plan to design and implement several proof plans in tober, 1984.
the MT system. Heuristics developed from natural
deduction proofs can be readily translated into the [Boswell 85] Boswell R.A.
Connection Calculus. In particular, we will try to An Analytic Survey of Analytic
implement within MT one such technique, developed Concept-Learning Programs.
in the group, called gazing, [Plummer & Bundy 84]. Working Paper. Dept. of Artificial
Gazing is a heuristic technique for controlling the Intelligence, Edinburgh, 1985.
expansion of non-logical definitions and the use of Forthcoming.
previously proved theorems during a proof attempt.
[Boyer & Moore 79]
We plan to add to MT a learning component based Boyer, R.S. and Moore J.S.
on precondition analysis and other analytic learning A Computational Logic.
techniques. This will analyse proofs using the meta- Academic Press, 1979.
level concepts already embodied in the MT proof ACM monograph series.
plans and use this analysis to modify the existing
plans and/or build new plans. These plans will then [Bundy 83] Bundy, A.
be added to MT to improve its theorem proving The Computer Modelling of Mathemati-
ability. cal Reasoning.
Academic Press. 1983.
Earlier version available from Edin-
6. Conclusion burgh as Occasional Paper 24.
The automation of mathematical reasoning involves
not just techniques for deduction, but also tech-
 A. Bundy 1229

[Bundy 85] Bundy. A. [Huet 75] Huet, G.

Poof Analysis: A technique for Con- A Unification Algorithm for Lambda
cept Formation. calculus.
In Ross, P. (editor), Proceedings of Theoretical Computer Science 1:27-57,
AISB-85, pages 78-86. 1985. 1975.
Also available as DAI Research Paper
no. 198. [Huet & Oppen 80]
[Bundy & Welham 81] Huet, G. and Oppen, D.C.
Bundy, A. and Welham, B. Equations and rewrite rules: a sur-
Using meta-level inference for selec- vey.
tive application of multiple In Book, R. (editor), Formal languages:
rewrite rules in algebraic perspectives and open problems.
manipulation. Academic Press, 1980.
Artificial Intelligence 16(2):189-212, Presented at the conference on
1981. formal language theory, Santa
Also available as DAI Research Paper Barbara, 1979. Available from SRI
121. International as technical report
CSl -111.
[Bundy et al 79]
Bundy, A., Byrd, L., Luger, G., Mellish, [lakatos 76] Lakatos, I.
C, Milne, R. and Palmer, M. Proofs and refutations: The logic of
Solving Mechanics Problems Using Mathematical discovery.
Meta-Level Inference. Cambridge University Press, 1976.
In Buchanan, B.G. (editor), Proceedings
of IJCAT-79, pages 1017-1027. [Lenat 82] Lenat D.B.
International Joint Conference on AM: An Artificial Intelligence ap-
Artificial Intelligence, 1979. proach to discovery in Math-
Reprinted in 'Expert Systems in the ematics as Heuristic Search.
microelectronic age' ed. Michie, In Knowledge-based systems in artifi-
D., Edinburgh University Press, cial intelligence. McGraw Hill,
1979. Also available from Edin- 1982.
burgh as DAI Research Paper No. Also available from Stanford as
112. TechReport AIM 286.
[Bundy et al 83] [Mitchell et al 81]
Bundy, A., Silver, B. and Plummer, D. Mitchell, T.M., Utgoff. P. E., Nudel,
An Analytical Comparison of some B. and Banerji, R.
Rule Learning Programs. Learning problem-solving heuristics
In Third Annual Technical Conference through practice.
of the British Computer Society's In Proceedings of IJCAI-8 1, pages
Expert Systems Specialist Group. 127-134. International Joint Con-
British Computer Society, 1983. ference on Artificial Intelligence,
Earlier Version in Procs of ECAI-8?. 1981.
[Mitchell et al 83]
[Chang & Lee 73] Mitchell, T.M., Utgoff. P. E. and
Chang C-l. and Lee R. C-T. Banerji, R.
Symbolic logic and mechanical theorem Learning by Experimentation: Ac-
proving. quiring and modifying problem-
Academic Press, 19 73. solving heuristics.
In Michalski, R.S, Carbonell, J.F. and
[Diettench et al 82] Mitchell. T.M. (editors). Machine
Dietterich, T.G., London, R., Clarkson, Learning, pages 163-190. Tioga
K. and Dromey. G. Press, 1983.
Learning and Inductive Inference.
In Cohen, P.R. and Feigenbaum, E.A. [Morris 69] Morris, J.B.
(editors), The Handbook of Artifi- E-Resolution: Extension of Resolution
cial Intelligence, Volume 3, chapter to include the equality relation.
XIV. Pitman Books Ltd, 1982. In Walker, D. and Norton, L.M.
(editor), Proceedings of IJCAI-69,
[Fearnley-Sander 85] pages 287-294. Kaufmann Inc.,
Fearnley-Sander, D. 1969.
Using and Computing Symmetry in
Geometry Proofs. [O'Keefe 82] O'Keefe, R.A.
In Ross, P. (editor), Proceedings of Automated Statistical Analysis.
AISB-85. 1985. Working Paper 104, Dept. of Artifi-
[Hayes 77] Hayes. P. cial Intelligence, Edinburgh, 1982.
In defence of logic.
In Proceedings of IJCAI-77. Inter- [Owen 85] Owen, S.G.
national Joint Conference on Ar- Analogy in Artificial Intelligence
tificial Intelligence, 1977. - Thesis Proposal.
1230 A. Bundy

Working Paper 176, Dept. of Artifi- [Wallen 83] Wallen, L.A.

cial Intelligence. Fdinburgh. 1985. Towards the Provision of a Natural
Mechanism for Expressing Domain -
[Plotkin 72] Plotkin, G. Specific Global Strategies in
Bu»lding-in equational theories. General Purpose Theorem-Provers.
In Michie, D and Meltzer, B (editors), Research Paper 202, Dept. of Artifi-
Machine Intelligence 7, pages cial Intelligence. Edinburgh. Sep-
73-90. Edinburgh University tember, 1983.
Press. 1972.

[Plummer 84] Plummer, D.

RUT: Reconstructed UT Theorem
Prover.
Working Paper 165, Dept. of Artifi-
cial Intelligence, Edinburgh, Sep-
tember. 1984.
[Plummer & Bundy 84]
Plummer, D. and Bundy, A.
Gazing: Identifying potentially useful
inferences.
Working Paper 160, Dept. of Artifi-
cial Intelligence, Edinburgh,
February, 1984.

[Robinson 65] Robinson, J.A.

A machine oriented logic based on
the Resolution principle.
J Assoc. Comput. Mach. 12:23-41,
1965.
[Silver 84] Silver, B.
Precondition Analysis: Learning Control
Information.
In Michalski, R.S., Carbonell, J.G. and
Mitchell, T.M. (editors), Machine
Learning 2. Tioga Publishing Com-
pany, 1984.
Forthcoming. Earlier version available
from Edinburgh as Research Paper
220.
[Sterling & Bundy 82]
Sterling, L. and Bundy, A.
Meta-level Inference and Program
Verification.
In Loveland, D.W. (editor), 6th Con-
ference on Automated Deduction,
pages 144-150. Springer- Verlag.
1982.
Lecture Notes in Computer Science
No. 138. Also available from
Edinburgh as Research Paper 168.

[Uschold et al 84]
Uschold, M., Harding, N., Muetzelfeldt,
R. and Bundy, A.
An Intelligent Front End for Ecologi-
cal Modelling.
In O'Shea, T. (editor). Proceedings of
ECAI-84, pages 761-770. ECAI,
1984.
Available from Edinburgh as Research
Paper 223.