Explain Reinforcement Learning Agents Through Fuzzy Rule Reconstruction

There has been a lot of interest in making reinforcement learning (RL) models more explainable, as their applications become more widespread. Currently, most of the explainable RL (xRL) models focus on improving the transparency of the agent's observations, rather than the relationship between the agent's states and actions. This study introduces the Explainable Fuzzy Reconstruction Net (EFRN), which aims to interpret these relationships in RL. The EFRN utilizes the interpretability of Fuzzy Neural Networks (FNNs) to create IF-THEN rules and a generative model to showcase the learned knowledge. The IF-THEN rules can be expressed in a way that is easily understandable for humans, such as “IF A THEN B.” The generative model then visualizes the state as patterns that is easy for humans to comprehend. The results of the study shows that the proposed EFRN maintains the same level of performance as traditional RL methods and significantly improves the explainability of the RL agents both globally and locally.