KGSMS: Knowledge Graph Sample based Multi-agent Simulation
Yue Cheng Deng, Meijuan Xia, Mengyuan Cao, Haocong Ma · 2022 IEEE 2nd International Conference on Electronic Technology, Communication and Information (ICETCI) · 2022
Recent algorithms on multi-agent problems, including StarCraftII macro-management problems and military strategy games simulations, are suffering from the immense computation burden from joint action spaces and the slow convergence speed from state exploring. Multi-agent reinforcement learning methods and knowledge inference within prior artificial experiences are currently the most ways to solve multi-agent problems. In this paper, we propose a new framework that integrates policies learnt from reinforcement learning methods with empirical actions provided by knowledge graphs. Policies improvements and relations between domains are updated synchronously during the training process. Experiments are carried out on two different scenes of coastal air-defense deployment and asymmetrical military attack missions from Mozi simulation system. Results indicate the success of the integration among reinforcement learning algorithms with knowledge graphs and the acceleration of convergence speed during the training with a large number of agents.