Players in multiplayer games can now learn near-optimal strategies independently using higher-order prediction, achieving regret bounds that scale polynomially with the number of players rather than exponentially.
This paper presents HOOD, a learning algorithm for multiplayer games that achieves constant regret (O(N³log²K)) without requiring players to coordinate. The algorithm combines higher-order prediction with a clever regularization technique to prevent the chaotic oscillations that plague previous approaches, making it practical for any N-player game.