关键词: 抗干扰决策, 强化学习, 深度强化学习, 收敛速度, 经验补充

Abstract:

Although classical reinforcement learning algorithms have shown certain effectiveness in the field of anti-jamming decision-making, their convergence speed is typically slow, and it is difficult to generate effective anti-jamming strategies in a short period of time, which is unacceptable in practical applications. To enhance anti-jamming capability while accelerating convergence, a reinforcement learning framework that incorporates an information-regulated long short-term memory is proposed to construct both the policy and value networks. This architecture efficiently models the spectral state under jamming conditions by filtering salient features and suppressing irrelevant information, thereby improving the model’s environmental representation accuracy. Additionally, an experience augmentation mechanism is introduced to accelerate the population of the replay buffer, enabling the algorithm to reach the training-ready state more quickly. Experimental results demonstrate that when the normalized throughput reaches 0.65, the proposed method requires 85.3% fewer episodes than the standard soft actor-critic (SAC) algorithm. Moreover, the normalized throughput after convergence exceeds that of SAC by 26.3%. Experimental results show that the proposed algorithm has high modeling accuracy, high convergence rate, fast normalized throughput and better adaptability and robustness.

Key words: anti-jamming decision-making, reinforcement learning, deep reinforcement learning, convergence speed, experience augmentation