Inverse Reinforcement Learning for Stochastic Zero-Sum Games

This paper presents two inverse reinforcement learning algorithms to learn unknown performance index functions for stochastic zero-sum games. The first one is a model-based algorithm that uses known system matrices and applies the Riccati equation to estimate the unknown cost weight matrices in the quadratic index functions. However, in practical scenarios, obtaining accurate system matrices can be challenging. Therefore, a second model-free algorithm is proposed, which calculates the unknown cost weight matrices in the quadratic index function using the target trajectory without requiring knowledge of the system matrices.