Function Approximation for Solving Stackelberg Equilibrium in Large Perfect Information Games

Function approximation (FA) has been a critical component in solving large zero-sum games. Yet, little attention given towards FA general-sum extensive-form games, despite them being widely regarded as computationally more challenging than their fully competitive or cooperative counterparts. A key challenge is that for many equilibria no simple analogue to the state value function used Markov Decision Processes and games exists. In this paper, we propose learning Enforceable Payoff Frontier (EPF)---a generalization of We approximate optimal Stackelberg correlated equilibrium by representing EPFs with neural networks training using appropriate backup operations loss functions. This first method applies setting, allowing us scale much larger while still enjoying performance guarantees based on error. Additionally, our proposed incentive compatibility easy evaluate without having depend self-play best-response oracles.