The Application Of Deep Learning In Wireless Communication

In wireless communications,the scarcity of spectrum resources has caused serious concern.The shortage of spectrum resources does not mean that it has been exhausted,but because the utilization of existing spectrum resources is generally low,resulting in a large number of allocated spectrum resources not being used well.Therefore,dynamic intelligent spectrum sensing and spectrum sharing based on cognitive radio have become a hot research topic.This paper considers the problem of spectrum sharing in a cognitive radio communication system consisting of a primary user and a secondary user.It is hoped that the secondary user can share the spectrum resources with the primary user to improve the spectrum resource utilization without causing harmful interference to the primary user through reasonable design.It is assumed that the primary user and the secondary user work together in a non-cooperative manner,and the primary user updates its transmission power based on a pre-defined power control policy.There is no communication between the primary user and the secondary user,so the secondary user does not have any knowledge about the primary user's transmit power and power control policy.The goal of the secondary user is to learn an efficient power control strategy,so that after several rounds of adjustment,both the primary user and the secondary user can successfully transmit their respective data.The success here is defined as the signal received by the receiver can meet the specified quality of service requirements.In order to interact with the primary user,the secondary user needs to obtain the primary user's information indirectly.This paper considers arranging a set of sensor nodes to collect the received signal strength at different locations in the wireless environment.The received signal strength measured by the sensors can characterize the state information of the system.When the environmental interference is not considered,it can be proved that the power control process of the secondary user is a Markov decision process,so this paper uses the Q-learning-based method.The experimental results show that no matter which power control strategy is adopted by the primary user(this paper takes two different power control strategies as examples),the secondary user can use the Q-learning-based power control algorithm to make the system reach the final state(both primary and secondary users can successfully transmit data)from any initial state in a short time and stay in the final state to achieve efficient and rational use of spectrum resources.However,in the actual communication scenario,the strength of the signal received by the sensors is disturbed by the environment,so the set of states that the secondary user may change from a finite discrete set to an infinite contiguous set,while the Q-learningbased approach cannot handle an infinite number of states.Therefore,this paper proposes a method based on deep reinforcement learning,which replaces the action-value function table(Q table)generated by the Q learning method with a deep neural network.Since the input of the deep neural network can be any value,there is no longer a requirement.The experimental results show that the method based on deep reinforcement learning has good performance under different system parameters.Finally,a comparative experiment with the DCPC optimization method is carried out to further illustrate the advantages of the deep reinforcement learning method.