Multiple reinforcement learning agents in a static environment

Abstract

Reinforcement learning agents are used in many domains; however, multiple, benevolent, and communicating reinforcement learning agents are rarely used. This paper explores using multiple reinforcement learning agents in a simple static environment to research methods allowing such agents to communicate during learning. The problem of developing a model whereby reinforcement learning agents can learn faster through the use of communication is explored. This paper develops two agents that are able to communicate with each other and able to learn about their actions using reinforcement learning techniques. The maze environment is used as a testbed. In this environment, the two agents are required to navigate the maze and trade information about the state of each other. This information includes the expected utilities for all actions from that state. To demonstrate the feasibility of the proposed approach, the agents are implemented. In addition, the efficiency of this model is compared with single agents and two independent agents.