With the increasing shortage of spectrum resources, dynamic spectrum access (DSA) technology is proposed to maximize the spectrum resources utilization. Traditional DSA solutions can no longer meet the requirements of high throughput and low interference in large-scale access scenarios of cognitive radio (CR). Therefore, in this paper, we propose a DSA scheme based on deep reinforcement learning (DRL) combined with multiple access methods to maximize the system throughput. In the DSA network, the access strategy adopted by the secondary user (SU) will directly affect the performance of the entire system, so we introduce DRL to help the SU learn the best access strategy in a dynamic environment. The trained SU can intelligently access the appropriate channel to avoid interference to the primary user (PU) and other SUs. By combining deep $Q$-network (DQN) into two multiple access methods: Frequency Division Multiple Access (FDMA) and Non-orthogonal Multiple Access (NOMA), DQN-based FDMA scheme and DQN-based NOMA scheme are designed, respectively, which can find the best DSA strategy to avoid collisions with PU or other SU and improve system throughput. Simulation results show that the DQN-based NOMA scheme has better performance than the DQN-based FDMA scheme.

随着频谱资源的日益短缺，提出了动态频谱接入（DSA）技术，以最大限度地利用频谱资源。传统的DSA解决方案已经不能满足认知无线电（CR）大规模接入场景中高吞吐量、低干扰的要求。因此，在本文中，我们提出了一种基于深度强化学习 (DRL) 结合多种访问方法的 DSA 方案，以最大限度地提高系统吞吐量。在DSA网络中，次要用户（SU）所采用的访问策略将直接影响整个系统的性能，因此我们引入DRL来帮助SU在动态环境中学习最佳访问策略。经过训练的 SU 可以智能地访问适当的信道，以避免对主用户（PU）和其他 SU 的干扰。通过结合深$问$-将网络（DQN）分为两种多址接入方式：频分多址接入（FDMA）和非正交多址接入（NOMA），分别设计了基于DQN的FDMA方案和基于DQN的NOMA方案，可以找到最佳的DSA避免与 PU 或其他 SU 冲突并提高系统吞吐量的策略。仿真结果表明，基于 DQN 的 NOMA 方案比基于 DQN 的 FDMA 方案具有更好的性能。