Reinforcement Learning for Guiding the E Theorem Prover

Automated Theorem Proving (ATP) systems search for aproof in a rapidly growing space of possibilities. Heuristicshave profound impact on search, and ATP makeheavy use heuristics. This work uses reinforcement learn-ing to learn metaheuristic that decides which heuristic useat each step proof the E system. Proximalpolicy optimization is used dynamically select heuristicfrom fixed set, based current state E. The approachis evaluated its ability reduce number inferencesteps successful searches, as an indicator in-telligent search.