Behavior Learning Based on a Policy Gradient Method: Separation of Environmental Dynamics and State Values in Policies

Abstract

Policy gradient methods are very useful approaches in reinforcement learning. In our policy gradient approach to behavior learning of agents, we define an agent’s decision problem at each time step as a problem of minimizing an objective function. In this paper, we give an objective function that consists of two types of parameters representing environmental dynamics and state-value functions. We derive separate learning rules for the two types of parameters so that the two sets of parameters can be learned independently. Separating these two types of parameters will make it possible to reuse state-value functions for agents in other different environmental dynamics, even if the dynamics is stochastic. Our simulation experiments on learning hunter-agent policies in pursuit problems show the effectiveness of our method.