一种利用优先经验回放深度Q-Learning的频谱接入算法

针对认知无线传感器网络中频谱接入算法的频谱利用率不高、重要经验利用率不足、收敛速度慢等问题，提出了一种采用优先经验回放双深度Q-Learning的动态频谱接入算法。该算法的次用户对经验库进行抽样时，采用基于优先级抽样的方式，以打破样本相关性并充分利用重要的经验样本，并采用一种非排序批量删除方式删除经验库的无用经验样本，以降低能量开销。仿真结果表明，该算法与采用双深度Q-Learning的频谱接入算法相比提高了收敛速度；与传统随机频谱接入算法相比，其阻塞概率降低了6%~10%，吞吐量提高了18%~20%，提高了系统的性能。

A Dynamic Spectrum Access Algorithm Based on Prioritized Experience Replay Deep Q-Learning

To solve the problems of slow convergence speed,low utilization rate of spectrum and important experience in cognitive wireless sensor networks(CWSNs),a dynamic spectrum access algorithm for CWSNs based on prioritized experience replay deep Q-Learning is proposed.When secondary users take samples from the experience bank in the proposed algorithm,a priority sampling method is adopted to break the correlation of samples and make full use of important samples.And a non-ordered batch deletion method is used to delete the useless samples in the experience bank in order to reduce energy cost.The simulation results show that the convergence speed of the proposed algorithm is faster than that of the spectrum access algorithm based on double deep Q-Learning.Compared with that of the random dynamic spectrum access algorithm,the congestion probability and throughput of the proposed algorithm is reduced by 6% to 10% and increased by 18% to 20%,respectively.