Reinforcement Learning by Neuro-Symbolic AI

Abstract

Reinforcement learning often require many trials for obtaining an optimal policy, and no interpretability for trained policies is provided. To overcome these problems, we propose a novel reinforcement learning method by neuro-symbolic AI which is a combination of recent deep neural network and symbolic reasoning. Our experimental results show training with the proposed method converges significantly faster than other state-of-the-art neuro-only and neuro-symbolic methods in a text-based game benchmark.