Resolution-Based Policy Search for Imperfect Information Differential Games

Differential games (DGs), considered as a typical model of game with continuous states and non-linear dynamics, play an important role in control and optimization. Finding optimal/approximate solutions for these game in the imperfect information setting is currently a challenge for mathematicians and computer scientists. This article presents a multi-agent learning approach to this problem. We hence propose a method called resolution-based policy search, which uses a limited non-uniform discretization of a perfect information game version to parameterize policies to learn. We then study the application of this method to an imperfect information zero-sum pursuit-evasion game (PEG). Experimental results demonstrate strong performance of our method and show that it gives better solutions than those given by traditional analytical methods.