The invention relates to the technical field of confrontation agents, and relates to a multi-agent
reinforcement learning-based red and blue confrontation intrusion monitoring collaborative decision-making method, which comprises the steps of S1, acquiring
environmental data of a target area; constructing a dynamic environment model fused with
terrain concealment; s2, on the basis of the dynamic environment model, constructing a random trace generation and attenuation model of a blue-party agent and a limited
perception model of a red-party agent; s3, constructing and updating a risk thermodynamic diagram used for representing the existence probability of the blue square in real time; s4, constructing a multi-agent game model of red and blue parties under the condition of information
asymmetry; carrying out adversarial training through multi-agent
reinforcement learning until the strategy is converged; and S5, deploying the trained strategy model, and carrying out red-blue adversarial
simulation and collaborative decision. According to the invention, by introducing the
terrain concealment index, the concealment identification capability of intrusion monitoring is effectively improved; by combining the dynamic environment model and the agent modeling, the accurate
simulation of the behavior
modes of the red and blue parties is realized, the
false alarm rate can be effectively reduced, and the decision timeliness can be improved.