Gradient descent for symmetric and asymmetric multiagent reinforcement learning

Abstract

A gradient-based method for both symmetric and asymmetric multiagent reinforcement learning is introduced in this paper. Symmetric multiagent reinforcement learning addresses the problem with agents involved in the learning task having equal information states. Respectively, in asymmetric multiagent reinforcement learning, the information states are not equal, i.e. some agents (leaders) try to encourage agents with less information (followers) to select actions that lead to improved overall utility values for the leaders. In both cases, there are a huge number of parameters to learn and we thus need to use some parametric function approximation methods to represent the value functions of the agents. The method proposed in this paper is based on the VAPS framework that is extended to utilize the theory of Markov games, which is a natural basis of multiagent reinforcement learning.