Abstract: With the application of intelligent technology in future war, multi-agent system, such as intelligent missile group and UAV group, will be put into operation. It is required for agents to have the ability of rapid operational decision-making. Due to the limited computing resources, small memory space and limited data transmission of unmanned systems, the algorithms to realize multi-agent autonomy, cooperation and group decision-making should be lightweight and minimization cost. To face the challenges of multi-agent collaborative decision-making, this paper proposes a reinforcement learning swarm intelligence game model based on deep network, discusses the key technologies involved, and innoratively proposes lightweight ideas from four key links of OODA decision cycle.