A Multi-agent Reinforcement Learning Modeling and Transfer Technique for Confrontation Game
Multi-agent game confrontation problems involve cooperations among agents.The traditional solutions based on game theory are not suitable to game confrontation problems in the complex scenarios.The multi-agent cooperative training mechanism is a research hotspot in recent years.For the multi-agent game confrontation problems published by China Electronics Technology Group Corporation,a deep multi-agent deep reinforcement learning method is designed based on the value decomposition.A network model for each agent is built indepen-dently.Each agent is connected by introducing the hybrid network.During training,the hybrid network is used to guide the network update of each agent.During the execution,each agent network runs independently to realize the training mode of centralized learning and decentralized execution.As for the isomorphic and heterogeneous scenarios,an efficient transfer training method is designed to improve the efficiency of fast training of multi-agent reinforcement learning method in different scenarios.Lastly,the experiments on the isomorphic and heterogeneous game confrontation problems are carried out respectively.The experiment results show that multi-agent reinforcement learning method and the transfer technique based on value decomposition can effectively improve the cooperative behaviors and the training efficiency of agents.
NETL
NSTL
万方数据
