We present first steps towards fully automated deduction that merely requiresthe user to submit proof problems and pick up results. Essentially, this necessi-tates the automation of the crucial step in the use of a deduction system, namelychoosing and configuring an appropriate search-guiding heuristic. Furthermore,we motivate why learning capabilities are pivotal for satisfactory performance.The infrastructure for automating both the selection of a heuristic and integra-tion of learning are provided in form of an environment embedding the "core"deduction system.We have conducted a case study in connection with a deduction system basedon condensed detachment. Our experiments with a fully automated deductionsystem 'AutoCoDe' have produced remarkable results. We substantiate Au-toCoDe's encouraging achievements with a comparison with the renowned the-orem prover Otter. AutoCoDe outperforms Otter even when assuming veryfavorable conditions for Otter.
In this report we present a case study of employing goal-oriented heuristics whenproving equational theorems with the (unfailing) Knut-Bendix completion proce-dure. The theorems are taken from the domain of lattice ordered groups. It will bedemonstrated that goal-oriented (heuristic) criteria for selecting the next critical paircan in many cases significantly reduce the search effort and hence increase per-formance of the proving system considerably. The heuristic, goalADoriented criteriaare on the one hand based on so-called "measures" measuring occurrences andnesting of function symbols, and on the other hand based on matching subterms.We also deal with the property of goal-oriented heuristics to be particularly helpfulin certain stages of a proof. This fact can be addressed by using them in a frame-work for distributed (equational) theorem proving, namely the "teamwork-method".
We present a novel approach to classification, based on a tight coupling of instancebased learning and a genetic algorithm. In contrast to the usual instance-based learning setting, we do not rely on (parts of) the given training set as the basis of a nearestneighbor classifier, but we try to employ artificially generated instances as concept prototypes. The extremely hard problem of finding an appropriate set of concept prototypes is tackled by a genetic search procedure with the classification accuracy on the given training set as evaluation criterion for the genetic fitness measure. Experiments with artificial datasets show that - due to the ability to find concise and accurate concept descriptions that contain few, but typical instances - this classification approach is considerably robust against noise, untypical training instances and irrelevant attributes. These favorable (theoretical) properties are corroborated using a number of hard real-world classification problems.
We present a method for learning heuristics employed by an automated proverto control its inference machine. The hub of the method is the adaptation of theparameters of a heuristic. Adaptation is accomplished by a genetic algorithm.The necessary guidance during the learning process is provided by a proof prob-lem and a proof of it found in the past. The objective of learning consists infinding a parameter configuration that avoids redundant effort w.r.t. this prob-lem and the particular proof of it. A heuristic learned (adapted) this way canthen be applied profitably when searching for a proof of a similar problem. So,our method can be used to train a proof heuristic for a class of similar problems.A number of experiments (with an automated prover for purely equationallogic) show that adapted heuristics are not only able to speed up enormously thesearch for the proof learned during adaptation. They also reduce redundancies inthe search for proofs of similar theorems. This not only results in finding proofsfaster, but also enables the prover to prove theorems it could not handle before.
We present an overview of various learning techniques used in automated theorem provers. We characterize the main problems arising in this context and classify the solutions to these problems from published approaches. We analyze the suitability of several combinations of solutions for different approaches to theorem proving and place these combinations in a spectrum ranging from provers using very specialized learning approaches to optimally adapt to a small class of proof problems, to provers that learn more general kinds of knowledge, resulting in systems that are less efficient in special cases but show improved performance for a wide range of problems. Finally, we suggest combinations of solutions for various proof philosophies.
We investigate in how far interpolation mechanisms based on the nearest-neighbor rule (NNR) can support cancer research. The main objective is to usethe NNR to predict the likelihood of tumorigenesis based on given risk factors.By using a genetic algorithm to optimize the parameters of the nearest-neighbourprediction, the performance of this interpolation method can be improved sub-stantially. Furthermore, it is possible to detect risk factors which are hardly ornot relevant to tumorigenesis. Our preliminary studies demonstrate that NNR-based interpolation is a simple tool that nevertheless has enough potential to beseriously considered for cancer research or related research.
The team work method is a concept for distributing automated theoremprovers and so to activate several experts to work on a given problem. We haveimplemented this for pure equational logic using the unfailing KnuthADBendixcompletion procedure as basic prover. In this paper we present three classes ofexperts working in a goal oriented fashion. In general, goal oriented experts perADform their job "unfair" and so are often unable to solve a given problem alone.However, as a team member in the team work method they perform highly effiADcient, even in comparison with such respected provers as Otter 3.0 or REVEAL,as we demonstrate by examples, some of which can only be proved using teamwork.The reason for these achievements results from the fact that the team workmethod forces the experts to compete for a while and then to cooperate by exADchanging their best results. This allows one to collect "good" intermediate resultsand to forget "useless" ones. Completion based proof methods are frequently reADgarded to have the disadvantage of being not goal oriented. We believe that ourapproach overcomes this disadvantage to a large extend.
We present a method for making use of past proof experience called flexiblere-enactment (FR). FR is actually a search-guiding heuristic that uses past proofexperience to create a search bias. Given a proof P of a problem solved previouslythat is assumed to be similar to the current problem A, FR searches for P andin the "neighborhood" of P in order to find a proof of A.This heuristic use of past experience has certain advantages that make FRquite profitable and give it a wide range of applicability. Experimental studiessubstantiate and illustrate this claim.This work was supported by the Deutsche Forschungsgemeinschaft (DFG).
We are going to present two methods that allow to exploit previous expe-rience in the area of automated deduction. The first method adapts (learns)the parameters of a heuristic employed for controlling the application of infer-ence rules in order to find a known proof with as little redundant search effortas possible. Adaptation is accomplished by a genetic algorithm. A heuristiclearned that way can then be profitably used to solve similar problems. Thesecond method attempts to re-enact a known proof in a flexible manner in orderto solve an unknown problem whose proof is believed to lie in (close) vicinity.The experimental results obtained with an equational theorem prover show thatthese methods not only allow for impressive speed-ups, but also make it possibleto handle problems that were out of reach before.
Problems stemming from the study of logic calculi in connection with an infer-ence rule called "condensed detachment" are widely acknowledged as prominenttest sets for automated deduction systems and their search guiding heuristics. Itis in the light of these problems that we demonstrate the power of heuristics thatmake use of past proof experience with numerous experiments.We present two such heuristics. The first heuristic attempts to re-enact aproof of a proof problem found in the past in a flexible way in order to find a proofof a similar problem. The second heuristic employs "features" in connection withpast proof experience to prune the search space. Both these heuristics not onlyallow for substantial speed-ups, but also make it possible to prove problems thatwere out of reach when using so-called basic heuristics. Moreover, a combinationof these two heuristics can further increase performance.We compare our results with the results the creators of Otter obtained withthis renowned theorem prover and this way substantiate our achievements.