## Fachbereich Informatik

### Refine

#### Keywords

- Case-Based Reasoning (2)
- EBG (1)
- Fallbasiertes Schliessen (1)
- Fallbasiertes Schließen (1)
- Similarity Assessment (1)
- analogical reasoning (1)
- analogy (1)
- automated proof planner (1)
- case-based reasoning (1)
- concept representation (1)

Planning means constructing a course of actions to achieve a specified set of goals when starting from an initial situation. For example, determining a sequence of actions (a plan) for transporting goods from an initial location to some destination is a typical planning problem in the transportation domain. Many planning problems are of practical interest.

This report presents the main ideas underlyingtheOmegaGamma mkrp-system, an environmentfor the development of mathematical proofs. The motivation for the development ofthis system comes from our extensive experience with traditional first-order theoremprovers and aims to overcome some of their shortcomings. After comparing the benefitsand drawbacks of existing systems, we propose a system architecture that combinesthe positive features of different types of theorem-proving systems, most notably theadvantages of human-oriented systems based on methods (our version of tactics) andthe deductive strength of traditional automated theorem provers.In OmegaGamma mkrp a user first states a problem to be solved in a typed and sorted higher-order language (called POST ) and then applies natural deduction inference rules inorder to prove it. He can also insert a mathematical fact from an integrated data-base into the current partial proof, he can apply a domain-specific problem-solvingmethod, or he can call an integrated automated theorem prover to solve a subprob-lem. The user can also pass the control to a planning component that supports andpartially automates his long-range planning of a proof. Toward the important goal ofuser-friendliness, machine-generated proofs are transformed in several steps into muchshorter, better-structured proofs that are finally translated into natural language.This work was supported by the Deutsche Forschungsgemeinschaft, SFB 314 (D2, D3)

Die Beweisentwicklungsumgebung Omega-Mkrp soll Mathematiker bei einer ihrer Haupttätigkeiten, nämlich dem Beweisen mathematischer Theoreme unterstützen. Diese Unterstützung muß so komfortabel sein, daß die Beweise mit vertretbarem Aufwand formal durchgeführt werden können und daß die Korrektheit der so erzeugten Beweise durch das System sichergestellt wird. Ein solches System wird sich nur dann wirklich durchsetzen, wenn die rechnergestützte Suche nach formalen Beweisen weniger aufwendig und leichter ist, als ohne das System. Um dies zu erreichen, ergeben sich verschiedene Anforderungen an eine solche Entwicklungsumgebung, die wir im einzelnen beschreiben. Diese betreffen insbesondere die Ausdruckskraft der verwendeten Objektsprache, die Möglichkeit, abstrakt über Beweispläne zu reden, die am Menschen orientierte Präsentation der gefundenen Beweise, aber auch die effiziente Unterstützung beim Füllen von Beweislücken. Das im folgenden vorgestellte Omega-Mkrp-System ist eine Synthese der Ansätze des vollautomatischen, des interaktiven und des planbasierten Beweisens und versucht erstmalig die Ergebnisse dieser drei Forschungsrichtungen in einem System zu vereinigen. Dieser Artikel soll eine Übersicht über unsere Arbeit an diesem System geben.

Typical instances, that is, instances that are representative for a particular situ-ation or concept, play an important role in human knowledge representationand reasoning, in particular in analogical reasoning. This wellADknown obser-vation has been a motivation for investigations in cognitive psychology whichprovide a basis for our characterization of typical instances within conceptstructures and for a new inference rule for justified analogical reasoning withtypical instances. In a nutshell this paper suggests to augment the proposi-tional knowledge representation system by a non-propositional part consistingof concept structures which may have directly represented instances as ele-ments. The traditional reasoning system is extended by a rule for justifiedanalogical inference with typical instances using information extracted fromboth knowledge representation subsystems.

This paper addresses two modi of analogical reasoning. Thefirst modus is based on the explicit representation of the justificationfor the analogical inference. The second modus is based on the repre-sentation of typical instances by concept structures. The two kinds ofanalogical inferences rely on different forms of relevance knowledge thatcause non-monotonicity. While the uncertainty and non-monotonicity ofanalogical inferences is not questioned, a semantic characterization ofanalogical reasoning has not been given yet. We introduce a minimalmodel semantics for analogical inference with typical instances.

The hallmark of traditional Artificial Intelligence (AI) research is the symbolic representation and processing of knowledge. This is in sharp contrast to many forms of human reasoning, which to an extraordinary extent, rely on cases and (typical) examples. Although these examples could themselves be encoded into logic, this raises the problem of restricting the corresponding model classes to include only the intended models.There are, however, more compelling reasons to argue for a hybrid representa-tion based on assertions as well as examples. The problems of adequacy, availability of information, compactness of representation, processing complexity, and last but not least, results from the psychology of human reasoning, all point to the same conclusion: Common sense reasoning requires different knowledge sources and hybrid reasoning principles that combine symbolic as well as semantic-based inference. In this paper we address the problem of integrating semantic representations of examples into automateddeduction systems. The main contribution is a formal framework for combining sentential with direct representations. The framework consists of a hybrid knowledge base, made up of logical formulae on the one hand and direct representations of examples on the other, and of a hybrid reasoning method based on the resolution calculus. The resulting hybrid resolution calculus is shown to be sound and complete.

Typical examples, that is, examples that are representative for a particular situationor concept, play an important role in human knowledge representation and reasoning.In real life situations more often than not, instead of a lengthy abstract characteriza-tion, a typical example is used to describe the situation. This well-known observationhas been the motivation for various investigations in experimental psychology, whichalso motivate our formal characterization of typical examples, based on a partial orderfor their typicality. Reasoning by typical examples is then developed as a special caseof analogical reasoning using the semantic information contained in the correspondingconcept structures. We derive new inference rules by replacing the explicit informa-tion about connections and similarity, which are normally used to formalize analogicalinference rules, by information about the relationship to typical examples. Using theseinference rules analogical reasoning proceeds by checking a related typical example,this is a form of reasoning based on semantic information from cases.

Planning for realistic problems in a static and deterministic environment with complete information faces exponential search spaces and, more often than not, should produce plans comprehensible for the user. This article introduces new planning strategies inspired by proof planning examples in order to tackle the search-space-problem and the structured-plan-problem. Island planning and refinement as well as subproblem refinement are integrated into a general planning framework and some exemplary control knowledge suitable for proof planning is given.

The paper addresses two problems of comprehensible proof presentation, the hierarchically structured presentation at the level of proof methods and different presentation styles of construction proofs. It provides solutions for these problems that can make use of proof plans generated by an automated proof planner.

Proof planning is an alternative methodology to classical automated theorem prov-ing based on exhausitve search that was first introduced by Bundy [8]. The goal ofthis paper is to extend the current realm of proof planning to cope with genuinelymathematical problems such as the well-known limit theorems first investigated for au-tomated theorem proving by Bledsoe. The report presents a general methodology andcontains ideas that are new for proof planning and theorem proving, most importantlyideas for search control and for the integration of domain knowledge into a general proofplanning framework. We extend proof planning by employing explicit control-rules andsupermethods. We combine proof planning with constraint solving. Experiments showthe influence of these mechanisms on the performance of a proof planner. For instance,the proofs of LIM+ and LIM* have been automatically proof planned in the extendedproof planner OMEGA.In a general proof planning framework we rationally reconstruct the proofs of limittheorems for real numbers (IR) that were first computed by the special-purpose programreported in [6]. Compared with this program, the rational reconstruction has severaladvantages: It relies on a general-purpose problem solver; it provides high-level, hi-erarchical representations of proofs that can be expanded to checkable ND-proofs; itemploys declarative contol knowledge that is modularly organized.

This paper addresses the decomposition of proofs as a means of constructingmethods in plan-based automated theorem proving. It shows also, howdecomposition can beneficially be applied in theorem proving by analogy.Decomposition is also useful for human-style proof presentation. We proposeseveral decomposition techniques that were found to be useful in automatedtheorem proving and give examples of their application.

This paper analyzes how mathematicians prove the-orems. The analysis is based upon several empiricalsources such as reports of mathematicians and math-ematical proofs by analogy. In order to combine thestrength of traditional automated theorem provers withhuman-like capabilities, the questions arise: Whichproblem solving strategies are appropriate? Which rep-resentations have to be employed? As a result of ouranalysis, the following reasoning strategies are recog-nized: proof planning with partially instantiated meth-ods, structuring of proofs, the transfer of subproofs andof reformulated subproofs. We discuss the represent-ation of a component of these reasoning strategies, aswell as its properties. We find some mechanisms neededfor theorem proving by analogy, that are not providedby previous approaches to analogy. This leads us to acomputational representation of new components andprocedures for automated theorem proving systems.

This paper shows how a new approach to theorem provingby analogy is applicable to real maths problems. This approach worksat the level of proof-plans and employs reformulation that goes beyondsymbol mapping. The Heine-Borel theorem is a widely known result inmathematics. It is usually stated in R 1 and similar versions are also truein R 2 , in topology, and metric spaces. Its analogical transfer was proposedas a challenge example and could not be solved by previous approachesto theorem proving by analogy. We use a proof-plan of the Heine-Boreltheorem in R 1 as a guide in automatically producing a proof-plan of theHeine-Borel theorem in R 2 by analogy-driven proof-plan construction.

This paper addresses a model of analogy-driven theorem proving that is more general and cognitively more adequate than previous approaches. The model works at the level ofproof-plans. More precisely, we consider analogy as a control strategy in proof planning that employs a source proof-plan to guide the construction of a proof-plan for the target problem. Our approach includes a reformulation of the source proof-plan. This is in accordance with the well known fact that constructing ananalogy in maths often amounts to first finding the appropriate representation which brings out the similarity of two problems, i.e., finding the right concepts and the right level of abstraction. Several well known theorems were processed by our analogy-driven proof-plan construction that could not be proven analogically by previous approaches.

This paper addresses analogy-driven auto-mated theorem proving that employs a sourceproof-plan to guide the search for a proof-planof the target problem. The approach presen-ted uses reformulations that go beyond symbolmappings and that incorporate frequently usedre-representations and abstractions. Severalrealistic math examples were successfully pro-cessed by our analogy-driven proof-plan con-struction. One challenge example, a Heine-Borel theorem, is discussed here. For this ex-ample the reformulaitons are shown step bystep and the modifying actions are demon-strated.

Analogy in CLAM
(1999)

CL A M is a proof planner, developed by the Dream group in Edinburgh,that mainly operates for inductive proofs. This paper addresses the questionhow an analogy model that I developed independently of CL A M can beapplied to CL A M and it presents analogy-driven proof plan construction as acontrol strategy of CL A M . This strategy is realized as a derivational analogythat includes the reformulation of proof plans. The analogical replay checkswhether the reformulated justifications of the source plan methods hold inthe target as a permission to transfer the method to the target plan. SinceCL A M has very efficient heuristic search strategies, the main purpose ofthe analogy is to suggest lemmas, to replay not commonly loaded methods,to suggest induction variables and induction terms, and to override controlrather than to construct a target proof plan that can be built by CL A Mitself more efficiently.

In recent years several computational systems and techniques fortheorem proving by analogy have been developed. The obvious prac-tical question, however, as to whether and when to use analogy hasbeen neglected badly in these developments. This paper addresses thisquestion, identifies situations where analogy is useful, and discussesthe merits of theorem proving by analogy in these situations. Theresults can be generalized to other domains.

Constructing an analogy between a known and already proven theorem(the base case) and another yet to be proven theorem (the target case) oftenamounts to finding the appropriate representation at which the base and thetarget are similar. This is a well-known fact in mathematics, and it was cor-roborated by our empirical study of a mathematical textbook, which showedthat a reformulation of the representation of a theorem and its proof is in-deed more often than not a necessary prerequisite for an analogical inference.Thus machine supported reformulation becomes an important component ofautomated analogy-driven theorem proving too.The reformulation component proposed in this paper is embedded into aproof plan methodology based on methods and meta-methods, where the latterare used to change and appropriately adapt the methods. A theorem and itsproof are both represented as a method and then reformulated by the set ofmetamethods presented in this paper.Our approach supports analogy-driven theorem proving at various levels ofabstraction and in principle makes it independent of the given and often acci-dental representation of the given theorems. Different methods can representfully instantiated proofs, subproofs, or general proof methods, and hence ourapproach also supports these three kinds of analogy respectively. By attachingappropriate justifications to meta-methods the analogical inference can oftenbe justified in the sense of Russell.This paper presents a model of analogy-driven proof plan construction andfocuses on empirically extracted meta-methods. It classifies and formally de-scribes these meta-methods and shows how to use them for an appropriatereformulation in automated analogy-driven theorem proving.

This case study examines in detail the theorems and proofs that are shownby analogy in a mathematical textbook on semigroups and automata, thatis widely used as an undergraduate textbook in theoretical computer scienceat German universities (P. Deussen, Halbgruppen und Automaten, Springer1971). The study shows the important role of restructuring a proof for findinganalogous subproofs, and of reformulating a proof for the analogical trans-formation. It also emphasizes the importance of the relevant assumptions ofa known proof, i.e., of those assumptions actually used in the proof. In thisdocument we show the theorems, the proof structure, the subproblems andthe proofs of subproblems and their analogues with the purpose to providean empirical test set of cases for automated analogy-driven theorem proving.Theorems and their proofs are given in natural language augmented by theusual set of mathematical symbols in the studied textbook. As a first step weencode the theorems in logic and show the actual restructuring. Secondly, wecode the proofs in a Natural Deduction calculus such that a formal analysisbecomes possible and mention reformulations that are necessary in order toreveal the analogy.

The amount of user interaction is the prime cause of costs in interactiveprogram verification. This paper describes an internal analogy techniquethat reuses subproofs in the verification of state-based specifications. Itidentifies common patterns of subproofs and their justifications in orderto reuse these subproofs; thus significant savings on the number of userinteractions in a verification proof are achievable.

Many mathematical proofs are hard to generate forhumans and even harder for automated theoremprovers. Classical techniques of automated theoremproving involve the application of basic rules, of built-in special procedures, or of tactics. Melis (Melis 1993)introduced a new method for analogical reasoning inautomated theorem proving. In this paper we showhow the derivational analogy replay method is relatedand extended to encompass analogy-driven proof planconstruction. The method is evaluated by showing theproof plan generation of the Pumping Lemma for con-text free languages derived by analogy with the proofplan of the Pumping Lemma for regular languages.This is an impressive evaluation test for the analogicalreasoning method applied to automated theorem prov-ing, as the automated proof of this Pumping Lemmais beyond the capabilities of any of the current auto-mated theorem provers.

While most approaches to similarity assessment are oblivious of knowledge and goals, there is ample evidence that these elements of problem solving play an important role in similarity judgements. This paper is concerned with an approach for integrating assessment of similarity into a framework of problem solving that embodies central notions of problem solving like goals, knowledge and learning.