Interpretable Reinforcement Learning With Neural Symbolic Logic

Zhihao Ma, Yuzheng Zhuang, Paul Weng, Dong Li, Kun Shao, Wulong Liu, Hankz Hankui Zhuo, Jianye HAO

Keywords: Interpretable Reinforcement Learning, Neural Symbolic Logic

Abstract: Recent progress in deep reinforcement learning (DRL) can be largely attributed to the use of neural networks. However, this black-box approach fails to explain the learned policy in a human understandable way. To address this challenge and improve the transparency, we introduce symbolic logic into DRL and propose a Neural Symbolic Reinforcement Learning framework, in which states and actions are represented in an interpretable way using first-order logic. This framework features a relational reasoning module, which performs on task-level in Hierarchical Reinforcement Learning, enabling end-to-end learning with prior symbolic knowledge. Moreover, interpretability is enabled by extracting the logical rules learned by the reasoning module in a symbolic rule space, providing explainability on task level. Experimental results demonstrate better interpretability of subtasks, along with competing performance compared with existing approaches.

One-sentence Summary: Leverage neural symbolic logic to provide interpretable policy in Reinforcement Learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): /references/pdf?id=1fFwCIGYuq