Design Of a multi-agent architecture using deep reinforcement learning to decrease the number of interactions between agents and environment by communicating knowledge.

Abstract

The purpose of this integration is to create a communication system where agents can learn from the environment using a deep learning algorithm and communicate the knowledge obtained. Through communication, agents can take advantage of the knowledge of its peers in a way that the number of incorrect actions taken in the environment is reduced. Apart from the use of a pair-based multi-agent architecture with well-defined roles, some adjustments were performed in the deep q learning algorithm. One of them is the addition of information to the observations stored in the experience replay; this additional information is a flag which allows the agent to recognize a relevant state so the value of the rewards can be adjusted during the training of the network. In addition, the use of epsilon-greedy is limited to prevent an agent from exploring states that have already been explored and reported by other agents. These states have been identified as having errors and are therefore not worthy of further exploration. These adjustments proved to be effective since the agent reduces the number of episodes where errors are made in around 86%.