Design and Implementation of a Multi Agent Architecture to Communicate Reinforcement Learning Knowledge and Improve Agents’ Behavior

This research project presents a multi agent architecture which uses reinforcement learning. The goal is to design a system that able the agents to take advantage of its peers’ knowledge. The knowledge of the environment is obtained from the reinforcement learning algorithm, Q-learning. While, the multi agent architecture sets a communication model among the agents of the system. To reach the goal, the present research project takes advantage of the Q-learning characteristic, off-policy, incorporating a condition before the use of ε-greedy. This condition allows the agents not to explore a state that has already been sent by another agent, or itself. In the proposed multi agent architecture the agents work in pairs. Each pair of agents have two different behaviors allowing them to communicate and work on relevant states of the environment. The conditions to send the states depend on the environment, specifically, it depends on the circumstances which the agent obtains a reward from the environment. The results evidences that the number of agent-environment interactions to improve agent’s behavior is reduced by more than 90% through the proposed architecture.