Reinforcement Learning Your Way: Agent Characterization through Policy Regularization

Recent advances in reinforcement learning (RL) have increased complexity which, especially for deep RL, has brought forth challenges related to explainability [1]. The opacity of state-of-the-art RL algorithms has led to model developers having a limited understanding of their agents' policies and no influence over learned strategies [2]. While concerns surrounding explainability have been noted for AI in general, it is only more recently that attempts have been made to explain RL systems [3, 1, 4, 5]. These attempts have resulted in a wide suite of methods requiring various degrees of expert knowledge, either about the state-action domain or about the specific RL algorithm. Further, they typically rely on post-hoc analysis of learned policies which give only observational assurances of agents' behaviour. We instead propose an intrinsic method of regularizing agents' actions based on a given prior. While current methods for RL regularization aim to improve training performance - e.g., by maximizing the entropy of the action distribution [6], or by minimising the distance to a prior sub-optimal state-action distribution [7] - our aim is the characterization of our agents' behaviours. We also extend the current regularization techniques to accommodate multi-agent systems which allows intrinsic characterization of individual agents. We provide a formal argument for the rationale of our method and demonstrate its efficacy in a toy problem where agents learn to navigate to a destination on a grid by performing, e.g., only right turns (under the premise that right turns are