Policy Regularization for Legible Behavior

原文アブストラクト

In Reinforcement Learning interpretability generally means to provide insight into the agent's mechanisms such that its decisions are understandable by an expert upon inspection. This definition, with the resulting methods from the literature, may however fall short for online settings where the fluency of interactions prohibits deep inspections of the decision-making algorithm. To support interpretability in online settings it is useful to borrow from the Explainable Planning literature methods that focus on the legibility of the agent, by making its intention easily discernable in an observer model. As we propose in this paper, injecting legible behavior inside an agent's policy doesn't require modify components of its learning algorithm. Rather, the agent's optimal policy can be regularized for legibility by evaluating how the policy may produce observations that would make an observer infer an incorrect policy. In our formulation, the decision boundary introduced by legibility impacts the states in which the agent's policy returns an action that has high likelihood also in other policies. In these cases, a trade-off between such action, and legible/sub-optimal action is made.