Traditional network situation assessment methods mostly focus on theoretical analysis, it is difficult to apply to the actual network environment; the data collected by sensors is complex, which easily causes excessive storage space load. In response to the above problems, combining reinforcement learning and game theory algorithms, proposing a network attack-defense dynamic perception model based on reinforcement learning and game theory to analyze network situation security and predict attack paths. First, designing the hierarchy analytic process with priority relation matrix to calculate system losses and the security situation; Secondly, using the Boltzmann probability distribution method to calculated the mixed strategy Nash equilibrium to find the best strategy. Finally, improving the Q-learning method and game theory to dynamic analyze the network state transition, to accurately predict the attack path and assist the defender choose the best defense through network simulation experiments, it is verified that the model is effective and feasible.