Multi-Agent Deep Reinforcement Learning in Warehouse Environments

Abstract: This report presents a deep reinforcement algorithm for multi-agent systems based on the classicalDeep Q-Learning algorithm. The method considers a decentralized approach to controlling theagents, by equipping each agent with its own neural network and replay memory. Training isconducted in a shared environment, enabling mutual interaction between agents. We give a fulldescription of our method and test it on a simple model of a warehouse, in which the objective is topick up and deliver packages to specified locations. We show that the algorithm performs well for alow number of agents, but sees a decline in performance once the numbers become high. A solutionto this problem is proposed.