Analysis of Coordination Structures of Partially Observing Cooperative Agents by Multi-agent Deep Q-Learning

Abstract

We compare the coordination structures of agents using different types of inputs for their deep Q-networks (DQNs) by having agents play a distributed task execution game. The efficiency and performance of many multi-agent systems can be significantly affected by the coordination structures formed by agents. One important factor that may affect these structures is the information provided to an agent’s DQN. In this study, we analyze the differences in coordination structures in an environment involving walls to obstruct visibility and movement. Additionally, we introduce a new DQN input, which performs better than past inputs in a dynamic setting. Experimental results show that agents with their absolute locations in their DQN input indicate a granular level of labor division in some settings, and that the consistency of the starting locations of agents significantly affects the coordination structures and performances of agents.