Learning Opponents Behavior in Receding Horizon Games

Receding horizon games are MPC-inspired games in which the strategies of the players are sequences of actions over a time horizon, and are implemented employing a receding horizon scheme. In this thesis, we explore the field of receding horizon games with unknown opponents. We consider potential games where the decision making problems of the agents are parameterized by different sets of behavior parameters, representing the preferences of each agent. We study non-cooperative scenarios taking the perspective of the individual agents, whose goal is to learn the behaviors of their opponents in order to play optimally. and approach the problem assuming that agents maintain a local estimate of the underlying game, on which they base the choice of their strategies. After an iteration of the game is played, agents communicate by exchanging various forms of feedback, and use these information to update their knowledge of the game. In order to do so, we propose two online, distributed algorithms based on techniques from inverse optimization and learning in games. We show and numerically analyse our results on an application inspired by the electricity market.