基于深度迁移学习的动态频谱快速适配抗干扰方法

机器学习逐渐发展成为一种成熟强大的技术工具，并被广泛应用于无线通信抗干扰领域。其中，较为典型的有基于深度强化学习的抗干扰方法，通过与动态、不确定通信环境的不断交互来学习最优用频策略，有效解决动态频谱接入抗干扰的问题。然而，由于外界电磁频谱空间复杂、干扰模式样式动态多变，从头开始学习复杂的抗干扰通信任务往往时效性差，导致学习效率和通信性能显著下降。针对上述问题，提出基于深度迁移学习的动态频谱快速适配抗干扰方法。首先，通过构建预训练模型对已知干扰模式进行学习；其次，使用卷积神经网络提取现实场景下的感知频谱数据，重用过往经验优先启动加速适配；最后，运用微调策略辅助强化学习实施在线抗干扰信道接入。仿真结果表明，相较于传统强化学习算法，所提方法能够有效加快算法收敛速度，提升通信设备抗干扰性能。

Rapid adaption to dynamic spectrum anti-jamming approach based on deep transfer learning

Machine learning has become a mature and powerful technique and has been widely used in the fields of wireless anti-jamming communication. Deep reinforcement learning(DRL), one of the typical anti-jamming approaches, that enables an agent to learn an optimal frequency-using policy by constantly interacting with dynamic and uncertain communications environments, has been proposed as effective tools to solve the problem of dynamic spectrum accessing. However, learning a complex task from scratch often results in poor timeliness due to the complexity of the state space of the external electromagnetic spectrum and the volatile variation for the jamming patterns, which may cause a significant decline of the learning efficiency as well as communication performance instead. For these problems mentioned above, this paper proposes a rapid adaption to dynamic spectrum anti-jamming(DSAL) method based on deep transfer learning(DTL). Firstly, an adequately pre-trained model is established learned from known jamming patterns. Further, convolution neural network(CNN) is used to extract jamming features from sensed spectrum data in real-world scenario and reusing knowledge that comes from previous experience contributes to scale up priority-startup and fast-adaption. In addition, fine-tune strategy is adopted to assist reinforcement learning(RL) algorithm to implement the task of on-line channel accessing for anti-jamming tasks. The simulation results show that, compared with traditional RL algorithm, our improved method can increase the convergence speed and reach better anti-jamming performance.