Model‐Free Approximate Dynamic Programming for Stochastic Zero‐Sum Games: Algorithm Design and Analysis

ABSTRACT This paper studies the discrete‐time stochastic zero‐sum games by employing the approximate dynamic programming technique. We present on‐policy and off‐policy policy iteration algorithms to attain the saddle point without using the information of the system dynamics. A comparative analysis of model‐free algorithms and their equivalence relationships is examined. Numerical examples are given to illustrate the efficiency of the proposed algorithms.