Inverse Reinforcement Learning for Legibility Automation in Intelligent Agents

When intelligent agents operate in a stochastic environment, they adhere to the principle of maximizing expected rewards to optimize their policies. The maximization of rewards becomes the sole objective when agents’ decision problems are resolved in most cases. However, there are instances where this principle leads to the agent’s behaviors (the optimal policy for solving the decision problems) lacking legibility. In other words, comprehending the agents’ intentions while they execute optimal policies poses a challenge for users, including other agents and even humans. Therefore, it becomes essential to evaluate the legibility of agents’ decision-making processes. Traditionally, domain experts’ insights have been relied upon to define legibility values, but this manual approach often introduces subjectivity and inconsistency, particularly in complex problem domains. Consequently, there is a pressing need for a systematic approach to derive legibility functions. The present study employs inverse reinforcement learning techniques to automate a legibility function in agents’ decision problems. We demonstrate the effectiveness of the inverse reinforcement learning method when considering legibility in a decision problem. We vary problem domains in the performance study and provide empirical evidence to support our findings.