Breeding Theorem Proving Heuristics with Genetic Algorithms

First-order theorem provers have to search for proofs in an infinite space of possible derivations. Proof search heuristics play a vital role for the practical performance of these systems. In the current generation of saturation-based theorem provers like SPASS, E, Vampire or Prover 9, one of the most important decisions is the selection of the next clause to process with the given clause algorithms. Provers offer a wide variety of basic clause evaluation functions, which can often be parameterized and combined in many different ways. Finding good strategies is usually left to the users or developers, often backed by large-scale experimental evaluations. We describe a way to automatize this process using genetic algorithms, evaluating a population of different strategies on a test set, and applying mutation and crossover operators to good strategies to create the next generation. We describe the design and experimental set-up, and report on first promising results.