Flexible Proof-Replay with Heuristics

We present a general framework for developing search heuristics for automated theorem provers. This framework allows for the construction of heuristics that are on the one hand able to replay (parts of) a given proof found in the past but are on the other hand exible enough to deviate from the given proof path in order to solve similar proof problems. We substantiate the abstract framework by the presentation of three distinct techniques for learning appropriate search heuristics based on so-called features. We demonstrate the usefulness of these techniques in the area of equational deduction. Comparisons with the renowned theorem prover Otter validate the applicability and strength of our approach.