Policy regularization for legible behavior

Abstract In this paper we propose a method to augment Reinforcement Learning agent with legibility. This is inspired by the literature in Explainable Planning and allows regularize agent’s policy after training, without requiring modify its learning algorithm. achieved evaluating how optimal may produce observations that would make an observer model infer wrong policy. our formulation, decision boundary introduced legibility impacts states which returns action non-legible because having high likelihood also other policies. these cases, trade-off between such action, legible/sub-optimal made. We tested grid-world environment highlighting policy, gathered both quantitative qualitative results. addition, discuss proposed regularization generalizes over methods functioning goal-driven policies, applicable general policies of are special case.