Training Agents with Neural Networks in Systems with Imperfect Information

The paper deals with multi-agent system that represents trading agents acting in the environment with imperfect information. Fictitious play algorithm, first proposed by Brown in 1951, is a popular theoretical model of training agents. However, it is not applicable to larger systems with imperfect information due to its computational complexity. In this paper we propose a modification of the algorithm. We use neural networks for fast approximate calculation of the best responses. An important feature of the algorithm is the absence of agent's a priori knowledge about the system. Agents' learning goes through trial and error with winning actions being reinforced and entered into the training set and losing actions being cut from the strategy. The proposed algorithm has been used in a small game with imperfect information. And the ability of the algorithm to remove iteratively dominated strategies of agents' behavior has been demonstrated.