Abstract: At present, the method of network attack and defense analysis based on stochastic game adopts the assumption of complete rationality, but in the actual network attack-defense confrontation, it is difficult for both sides of attack and defense to meet the high requirement of complete rationality, which reduces the accuracy and guiding value of the existing methods. Based on the reality of network attack-defense confrontation, the influence of bounded rationality on attack-defense stochastic game is analyzed. Under the constraints of bounded rationality, a stochastic game model is constructed. Aiming at the problem of network state explosion, a method of extracting network state and attack-defense action based on attack-defense graph is proposed, which the game state space is effectively reduced. On this basis, WoLF-PHC algorithm in reinforcement learning is introduced to carry out bounded rational stochastic game analysis and design a defensive decision-making algorithm with online learning ability. By learning, the algorithm can obtain the optimal defense strategy for the current attacker. The obtained strategy is superior to the Nash equilibrium strategy of the existing attack-defense stochastic game model under bounded rationality. By introducing eligibility trace to improve WoLF-PHC, the learning speed of defenders is further improved. The experimental results verify the effectiveness and advancement of the proposed method.