Learning Symbolic Temporal Advice for Guiding Reinforcement Learning (Extended Abstract)

Abstract

Reinforcement learning is sample-inefficient and opaque—problems that reward shaping and feature engineering often fail to address. We introduce a method to learn temporal advice from expert demonstrations via inductive logic programming, transforming the advice into answer set automata for transparent RL guidance and accelerating learning. Our advice is expressed in LTLASP f (LTLf with Answer Set queries), enabling generalizable specifications that transfer from small to large map instances. In RL benchmarks, this temporal guidance outperforms unguided Q-learning and non-temporal baselines in both convergence rate and final performance.