FEMaLeCoP: Fairly efficient machine learning connection prover

Abstract

FEMaLeCoP is a connection tableau theorem prover based on leanCoP which uses efficient implementation of internal learningbased guidance for extension steps. Despite the fact that exhaustive use of such internal guidance can incur a significant slowdown of the raw inferencing process, FEMaLeCoP trained on related proofs can prove many problems that cannot be solved by leanCoP. In particular on the MPTP2078 benchmark, FEMaLeCoP adds 90 (15.7%) more problems to the 574 problems that are provable by leanCoP. FEMaLeCoP is thus the first AI/ATP system convincingly demonstrating that guiding the internal inference algorithms of theorem provers by knowledge learned from previous proofs can significantly improve the performance of the provers. This paper describes the system, discusses the technology developed, and evaluates the system.