Multi-user Dynamic Spectrum Access Algorithm Based On Reinforcement Learning

Because of the rapid development of wireless communication,the future mobile communication system faces challenges.With limited spectrum resources,the key to the future development of mobile communication is how to make the best use of spectrum resources.At present,the traditional static spectrum allocation has been unable to meet the demand.Therefore,dynamic spectrum access algorithm is one of the main topics of future mobile communication research.The goal of this paper is to build a distributed multi-user non-cooperative dynamic spectrum access algorithm model that can adapt to the development of wireless communication by combining reinforcement learning algorithm.The specific research contents are as follows:First,this paper proposes the single-carrier dynamic spectrum access algorithm based on DQN and the single-carrier dynamic spectrum access algorithm based on A3 C.Simulation results show that both algorithms can reduce the probability of conflict and improve the bandwidth utilization.Secondly,considering that the current algorithm does not consider multi-carrier access and Underlay,this paper extends the proposed two single-carrier access algorithms to multi-carrier access and introduces the Underlay access method.The simulation results show that the multi-carrier access algorithm can further improve the bandwidth utilization compared with the single-carrier access algorithm.The DQN algorithm performs better than the A3 C.Finally,this paper proposes an improved algorithm based on optimized sampling and an improved algorithm based on branch network to solve the problem of slow convergence speed of multicarrier algorithm.On the one hand,this paper optimizes the sample sampling rate and introduces the idea of supervised learning and course learning.The improvement of algorithm sample quality improves the training efficiency of the network.On the other hand,according to the difference between single carrier and multi-carrier access,behavioral value network is divided into two networks for simultaneous training.Simulation results show that the algorithm has a fast convergence speed and a lower collision probability after stabilization than the unimproved algorithm,which proves its effectiveness.