The Lagging Anchor Algorithm: Reinforcement Learning in Two-Player Zero-Sum Games with Imperfect Information

摘要

The article describes a gradient search based reinforcement learning algorithm for two-player zero-sum games with imperfect information. Simple gradient search may result in oscillation around solution points, a problem similar to the “Crawford puzzle”. To dampen oscillations, the algorithm uses lagging anchors, drawing the strategy state of the players toward a weighted average of earlier strategy states. The algorithm is applicable to games represented in extensive form. We develop methods for sampling the parameter gradient of a player's performance against an opponent, using temporal-difference learning. The algorithm is used successfully for a simplified poker game with infinite sets of pure strategies, and for the air combat game Campaign, using neural nets. We prove exponential convergence of the algorithm for a subset of matrix games.