New Algorithm Reveals Hidden Payoff Sets in Game Theory with Optimal Accuracy

A team of researchers led by Annalisa Barbara has tackled a fundamental challenge in inverse game theory: given only observed actions of players (believed to be playing a Nash equilibrium), how accurately can we infer the set of possible payoff functions that could have produced that behavior? Their paper, 'Optimal Rates for Feasible Payoff Set Estimation in Games,' provides the first minimax-optimal rates for this estimation problem.

Unlike traditional inverse reinforcement learning that produces a single payoff estimate, this work focuses on estimating the entire set of feasible payoffs consistent with observed equilibrium play. The research covers both exact Nash equilibria and approximate equilibria, zero-sum games and general-sum bimatrix games. The key theoretical contribution is establishing tight bounds on the Hausdorff distance between the true feasible payoff set and any estimator, achieving optimal rates up to constant factors. This has direct implications for counterfactual analysis and mechanism design in high-stakes scenarios like auction design, dynamic pricing, and security games, where understanding the range of possible player valuations is critical.