MaCA: a Multi-agent Reinforcement Learning Platform for Collective Intelligence

Heterogeneous multi-agent cooperative decision-making is one of the kernel problems in collective intelligence field. Reinforcement learning may be an effective technology to improve this research. An appropriate training environment is a necessary condition for intensive training effectively. In this study, a scalable multi-agent reinforcement training platform called MaCA was built to improve reinforcement learning effectiveness for heterogeneous collective cooperative decision making First, a kernel environment aimed electromagnetism military combat background was established as the basis of the platform. Second, a set of reinforcement learning interface was designed for reinforcement learning algorithm adapting. Third, a reinforcement learning agent based on MARL algorithm and a rule-based agent were implemented. Finally, an experiment for training and rivalry was conducted to evaluate the effectiveness of the platform. The experimental results show that after trained in MaCA, the MARL agent shows certain cooperation ability and achieved 100% win rate against the rule-based agent after hundreds of thousands of training iteration. The results demonstrate that MaCA is a suitable and effective environment of multi-agent decision training in heterogeneous reinforcement learning.