Research on Distributed Dynamic Spectrum Access Based on Deep Reinforcement Learning

Dynamic Spectrum Access (DSA) is a critical technology for Cognitive Wireless Sensor Network (CWSN). The main challenge of DSA is how Secondary Users (SUs) can quickly and accurately identify vacant spectrum, while ensuring that the service of the Primary User (PU) is not interrupted. The current DSA solutions do not satisfy the requirements of high throughput, low interference and fast convergence simultaneously for large scale multiple users and multiple channels access scenarios. In this paper, we propose a distributed DSA algorithm based on Deep Reinforcement Learning (DRL). First, we construct a Cognitive Wireless Sensor Network (CWSN) environment with multiple users and multiple channels. Next, based on the spectrum sensing results, each SU provides channel observations to our proposed Deep Q-Network (DQN) model for training in order to learn the optimal spectrum access policy. Finally, using the output of the DQN model, each SU intelligently accesses the appropriate channel. In order to improve the training accuracy and address the performance degradation problem caused by the network depth in deep neural networks, we added the Residual Network (ResNet) structure to the DQN. Simulation results show that the proposed algorithm achieves faster convergence speed, completely avoids collisions between SUs, greatly reduces the interference of SUs to PU, and significantly improves the success rate of channel access.