Promoting training of multi agent systems
Petro Kravets, Vasyl Lytvyn, Victoria Vysotska, Yevhen Burov
Abstract
The problem of incentive training of multi-agent systems in the game formulation for collective decision making under uncertainty is considered. Methods of incentive training do not require a mathematical model of the environment and enable decision making directly in the training process. Markov model of stochastic game is constructed and the criteria for its solution are formulated. An iterative Q-method for solving a stochastic game based on the numerical identification of a characteristic function of a dynamic system in space of state-action is described. Players' current gains are determined by the method of randomization of payment Q-matrix elements. Mixed player strategies are calculated using the Boltzmann method. Pure strategies are determined on the basis of discrete random distributions given by mixed player strategies. The algorithm for stochastic game solving is developed and results of computer implementation of game Q-method are analyzed.