New d-ISBRAG algorithm enables robust Nash equilibrium under data uncertainty

Researchers Nirabhra Mandal and Sonia Martínez from UC San Diego have developed a new framework for solving stochastic games where agents face unknown probability distributions. Their paper, published on arXiv, addresses a critical gap: existing Nash equilibrium methods assume full knowledge of randomness, but real-world agents often have only a finite set of samples. The solution is distributionally robust optimization—each agent maximizes the worst-case expected utility over a Wasserstein ball centered on the empirical distribution. This yields a set of distributionally robust Nash equilibria (DRoNE) that are provably close to the true stochastic game equilibria.

The key contribution is the d-ISBRAG algorithm (distributed inertial supported better response ascending supergradient dynamics). It operates under partial observations: agents estimate others’ strategies via a dynamic consensus subroutine over a directed communication network. The algorithm handles both shared sample sets and individual samples (with simplifying assumptions). By reformulating the robust optimization problem in a tractable way, d-ISBRAG computes the required supergradients in a distributed manner. Simulations demonstrate convergence, making this a practical tool for multi-agent systems in robotics, economics, and networked control where data is scarce and communication is limited.