Learning to Play General-Sum Games against Multiple Boundedly Rational Agents

We study the problem of training a principal in multi-agent general-sum game using reinforcement learning (RL). Learning robust policy requires anticipating worst possible strategic responses other agents, which is generally NP-hard. However, we show that no-regret dynamics can identify these worst-case poly-time smooth games. propose framework uses this evaluation method for efficiently RL. This be extended to provide robustness boundedly rational agents too. Our motivating application automated mechanism design: empirically demonstrate our learns mechanisms both matrix games and complex spatiotemporal In particular, learn dynamic tax improves welfare simulated trade-and-barter economy by 15%, even when facing previously unseen RL taxpayers.