Prof. Ming Cao, University of Gronigen

Title: Reinforcement learning to solve inverse games

Abstract: In an inverse game, the goal is to find the cost function parameters that explain why the observed game dynamics can be at a Nash equilibrium. By focusing on non-cooperative linear-quadratic output-feedback differential games, I show how the inverse games can be solved using reinforcement learning techniques. Given players’ stabilizing feedback laws, a model-based algorithm can be constructed that finds the cost-function parameters iteratively. The model-free version of the algorithm is more computationally demanding, but realizable using data samples. In addition to explaining the main results in the focused class of games, I also extend the discussion to
broader classes of games.