On-policy and Off-policy Value Iteration Algorithms for Stochastic Zero-Sum Games

This paper considers the value iteration algorithms of stochastic zero-sum linear quadratic games with unkown dynamics. The model-free on-policy and off-policy learning algorithms are developed, where the system dynamics is not required and the Riccati equation does not need solving. The convergence of the algorithms is shown and relationships between algorithms are illustrated. The effectiveness of model-free algorithms is shown by numerical experiments.