DQN–MADDPG Coordinating the Multi-agent Cooperation

Abstract

Multi-agent coordinating aims for many agents to finish the same or more mutual goals. To achieve this function, there are two main ideas; the first is making multi-agents communicate with each other and acquire the information for other agents. Another is sharing the same environment, dividing the work, and cooperating to achieve the goal. ‘’Gym-cooking’’ is an excellent model to test the algorithms’ performance in network coordination and game theory; this is a sharing environment. Based on sharing information, the agent have two networks(policy and target) and tries to do something more efficiently. This article will increase the complexity of the environment and use different algorithms to process the experiment. Specially, this paper will use the MADDPG model as the primary model to show its performance in a complex environment and contrast other models like DQN. The MADDPG model in this experiment is unlike the traditional MADDPG; the work trains the MADDPG network to deal with emergencies and accidents.