Zero-Sum Non-stationary Stochastic Games with the Long-Run Average Criterion

This paper is concerned with the existence and computation of an equilibrium for a non-stationary average stochastic zero-sum game with Borel spaces, in which the payoff functions and transition probabilities are allowed to change over time. First, we present an extension of the span-fixed point theorem for an operator to a sequence of time-dependent operators. Second, we find a new set of conditions, which is the generalization of the ergodicity ones in the existing literature. Using the extension of the span-fixed point theorem and the novel conditions, we prove the existence of a solution to the average-reward game equations (ARGEs). Third, by the ARGEs we establish the existence of the value and the equilibrium for this game. Moreover,by constructing an approximation sequence of the solution to the ARGEs, we provide a rolling horizon algorithm for computing the value and \( \varepsilon \) $\epsilon$ -equilibria, and also prove the convergence of the algorithm. Finally, we illustrate the conditions and results in this paper by several energy management models.