Synthesising Reinforcement Learning Policies Through Set-Valued Inductive Rule Learning

Today’s advanced Reinforcement Learning algorithms produce black-box policies, that are often difficult to interpret and trust for a person. We introduce policy distilling algorithm, building on the CN2 rule mining distills into rule-based decision system. At core of our approach is fact an RL process does not just learn policy, mapping from states actions, but also produces extra meta-information, such as action values indicating quality alternative actions. This meta-information can indicate whether more than one near-optimal certain state. extend make it able leverage knowledge about equally-good actions distill fewer rules, increasing its interpretability by Then, ensure rules explain valid, non-degenerate we refinement algorithm fine-tunes obtain good performance when executed in environment. demonstrate applicability Mario AI benchmark, complex task requires modern reinforcement learning including neural networks. The explanations capture learned only few allow person understand what agent learned. Source code: https://gitlab.ai.vub.ac.be/yocoppen/svcn2.