A Multi-Agent Actor-Critic Based Approach Applied to the Snake Game

Multi-agent reinforcement learning algorithm has been proposed, demonstrated, and improved on tasks which require a team of agents to cooperate or compete to complete. Sometimes the agents are caught in some dilemmas caused by incomplete observation of environmental information, using the existing q-learning algorithms. Thus, our team come up with the idea of introducing an early warning mechanism to tackle this problem. Actor-critic is a special method to achieve early warning mechanism, guiding the actors' follow-up actions through the existence of critic agent. This kind of algorithm is applicable not only in shared rewards settings, but also in individualized settings where the agents are not able to know the global state. Nowadays, it has become a broad consensus to use the game simulation environment as a research and verification environment for artificial intelligence. In this paper, the Snake Game real-time battle platform is selected as the research environment of the multi-agent reinforcement learning algorithm and the performance of the q-learning algorithm and the actor-critic algorithm is compared on the platform. To further improve the actor-critic algorithm, a better way of experience replay is introduced and its performance is also compared in this environment.