Fairness in two-player zero-sum games with bandit feedback
Researchers have developed a new method to analyze two-player zero-sum games with bandit feedback under fairness constraints. Their approach re-parametrizes the game to transform it into a standard zero-sum game, simplifying the analysis of mixed equilibria. This allows for the derivation of fair minimax values and a dual representation that quantifies the price of fairness, showing it is at most $\alpha(1-1/m)$ and vanishes if the unconstrained equilibrium already has full support. The proposed algorithm achieves an $\widetilde{O}(T^{2/3})$ regret bound for general mixed fair equilibria. AI
IMPACT Introduces a novel theoretical framework for analyzing fair game equilibria, potentially impacting AI agents in competitive or resource-constrained environments.