Spectrum Planning In Cellular Network And Radio Resource Management In D2D System Based On Reinforcement Learning

The research of resource allocation has always been one of the core issues in wireless communication system, which is due to the nature of the lack of wireless resource. With the emergency of various types of intelligent terminals and the rapid development of mobile communication technology, user’s business gets more rich and diversified, which puts forward higher demands and requirements to the communication system to get higher service quality and faster transmission rate. So how to use the limited radio resources reasonably and effectively will be the main issue and objective to provide high-speed data transmission for future wireless communication system to meet the needs of all kinds of uses’ business. This thesis focuses on introducing an intelligent method—reinforcement learning into the radio resource management issues to address the problem of allocating the radio resource effectively and dynamically.This thesis first gives a detailed introduction and description to the fundamental and commonly used algorithms of reinforcement learning to lay a theoretical foundation for using it to solve the problems in wireless resource management. Two different applications of the reinforcement learning in cellular network have been proposed as follow:Firstly, for the downlink resource allocation between the LTE cells, based on fractional frequency reuse scheme, a gradient reinforcement learning algorithm has been proposed when the traffic distribution is uneven in the edge region among different cells, which is quite different from the traditional scheme of fixed frequency reuse factor. Through simulation, the scheme proposed in this article performs better in adapting to the edge load distribution in different cells, reducing the dissatisfaction probability of cell-edge user and increasing the average user throughput.Secondly, in cellular network combining the D2 D communication technology, two different methods of resource allocation for orthogonal and non-orthogonal resource sharing scenarios have been proposed, which is different from the traditional instantaneous method. The D2 D cellular system has been modeled as a Markov environment, in which the arrival of user’s business follows Poisson distribution. One important algorithm of reinforcement learning—Q-learning has been used to find the optimal resource allocation strategy in time series. Simulation results show that, using Q-learning can achieve greater global cumulative utility and meet the higher priority of cellular users in the orthogonal resource sharing scenario, and in non-orthogonal resource sharing scenario, greater system capacity can be achieved through Q-learning algorithm.