Towards Explainable Reinforcement Learning Using Scoring Mechanism Augmented Agents

Deep reinforcement learning (DRL) is increasingly used in application areas such as medicine and finance. However, the direct mapping from state to action in DRL makes it challenging to explain why decisions are made. Existing algorithms for explaining DRL policy are posteriori, explaining to an agent after it has been trained. As a common limitation, these posteriori methods fail to improve training with the deduced knowledge. Face with that, an end-to-end trainable explanation method is proposed, in which an Adaptive Region Scoring Mechanism (ARS) is embedded into DRL system. The ARS explains the agent's action by evaluating the features of the input state that are most relevant action before DRL re-learn from task-related regions. The proposed method is validated on Atari games. Experiments demonstrate that agent using the explainable proposed mechanism outperforms the original models.