Sensitivity analysis of reinforcement learning for massive-multi-agent systems
Zhiyuan Liu, Tianqi Shen, Yupei Li · 2nd International Conference on Artificial Intelligence, Automation, and High-Performance Computing (AIAHPC 2022) · 2022
Over the past few years, reinforcement learning has become one of the most popular topics in the field of Machine Leaning. Its nature of unsupervised learning has made it rather powerful and convenient for solving specific tasks. Unlike decision trees, reinforcement learning models do not need a pre-configured policy tree to conduct actions, which makes it possible for machinery to solve extreme complex tasks such as playing GO or other games. One of the most powerful networks is the Deep-Q network. DQN picks the optimal action under current situation to achieve the largest reward. However, even though DQN has been tested to work well on single agent environments, there lack experiments of its performance on large massive multi-agent environment trainings. This paper conducts several experiments in Petting Zoo to test the performance and other factors that affect the performance of the DQN in a massive-multi-agent system. Results demonstrates that lager number of agents and experience replay size improve the performance of DQN in a "Battle" environment.