Solving Football by Exploiting Equilibrium Structure of 2p0s Differential Games with One-Sided Information

A team of researchers has published a breakthrough paper at ICLR 2026 titled 'Solving Football by Exploiting Equilibrium Structure of 2p0s Differential Games with One-Sided Information.' The work addresses a fundamental scalability challenge in imperfect-information extensive-form games (IIEFGs), where traditional solvers struggle with continuous action spaces and long time horizons. The researchers focused on a specific but important class of games where one player (P1) knows the exact payoff while the opponent (P2) only has a belief over I possible payoffs. They proved that under mild conditions, equilibrium strategies concentrate on a small number of action prototypes, dramatically reducing computational complexity.

The key mathematical insight shows that P1's equilibrium strategy concentrates on at most I action prototypes, while P2's concentrates on I+1. When I is much smaller than the action space size U, this collapses the game tree complexity from U^2K to I^K for P1 and (I+1)^K for P2. The researchers implemented this insight in model-free multiagent reinforcement learning and model predictive control frameworks, achieving significant improvements over state-of-the-art IIEFG solvers. Their demonstration solves a 22-player football game with continuous action spaces over K=10 time steps, where the offense team strategically conceals its intended play until a critical moment to exploit its information advantage. This work opens new possibilities for solving complex strategic interactions in sports analytics, cybersecurity, defense planning, and financial markets where information asymmetry plays a crucial role.