A Proximal/Gradient Approach for Computing the Nash Equilibrium in Controllable Markov Games

This paper proposes a new algorithm for computing the Nash equilibrium based on an iterative approach of both the proximal and the gradient method for homogeneous, finite, ergodic and controllable Markov chains. We conceptualize the problem as a poly-linear programming problem. Then, we regularize the poly-linear functional employing a regularization approach over the Lagrange functional for ensuring the method to converge to some of the Nash equilibria of the game. This paper presents two main contributions: The first theoretical result is the proposed iterative approach, which employs both the proximal and the gradient method for computing the Nash equilibria in Markov games. The method transforms the game theory problem in a system of equations, in which each equation itself is an independent optimization problem for which the necessary condition of a minimum is computed employing a nonlinear programming solver. The iterated approach provides a quick rate of convergence to the Nash equilibrium point. The second computational contribution focuses on the analysis of the convergence of the proposed method and computes the rate of convergence of the step-size parameter. These results are interesting within the context of computational and algorithmic game theory. A numerical example illustrates the proposed approach.