Application of instruction-based behavior explanation to a reinforcement learning agent with changing policy

Abstract

Agents that acquire their own policies autonomously have the risk of accidents caused by the agents’ unexpected behavior. Therefore, it is necessary to improve the predictability of the agents’ behavior in order to ensure the safety. Instruction-based Behavior Explanation (IBE) is a method for a reinforcement learning agent to announce the agent’s future behavior. However, it was not verified that the IBE was applicable to an agent that changes the policy dynamically. In this paper, we consider agents under training and improve the IBE for the application to agents with changing policy. We conducted an experiment to verify if the behavior explanation model of an immature agent worked even after the agent’s further training. The results indicated the applicability of the improved IBE to agents under training.