Research On Energy Efficient Resource Management Algorithms For Ultra-Dense Networks

With the rapid development of mobile interconnection technology and ecological evolution of mobile applications,a large number of new mobile services are emerging and the number and types of smart mobile terminals are increasing,which result in ubiquitous access demand and exponential growth of wireless traffic for wireless networks.On the other hand,the energy consumption and related pollution problems of information and communication technology industry are have now become a key pillar of social and economic concerns.As one of the key technologies of the next generation cellular mobile communication networks,the ultra-dense networks(UDNs),which consists of substantial low power and small coverage access point(SAPs)with different types overlaid with the macro base stations(MBSs),brings the users closer to their association access points with inter-site distances as short as tens of meters,so as to achieve the goal of improving network capacity and reducing the overall energy consumption of the system.However,the dense deployment of SAPs make the network suffers a complex topology and serious interference,which brings challenges to the resource management and performance optimization.This dissertation studies the energy-efficient UDNs resource management technology.The main research contents and contributions can be summarized as the following three aspects:First,aiming at the maximizing the downlink energy efficiency of the UDNs,the dissertation studies the energy efficient resource allocation algorithm for the UDNs.Firstly,an energy efficiency maximization problem has been proposed according to the frameworks of UDNs.Three resource allocation algorithms,i.e.,centralized,distributed,robust distributed algorithms,are proposed from the perspective of optimization system performance,complexity and adaptability to environment.To cope with the non-convexity of the original problem,the original problem is transformed into a convex form according to the fractional programming theory,and a centralized iterative algorithm is proposed based on DinkelBach's algorithm.In improve the efficiency and reduce the computational complexity,based on the direction method of multipliers(ADMM),a distributed iteration resource allocation algorithm has been proposed.The distributed algorithm decomposes the original problem into smaller sub-problem,and orignal varables are divided into local variables,dual variables and global variables.Local variables and dual variables are solved by local access points in parallel,and global variables are solved by information interaction among access points.Moreover,in order to improve the noise tolerance of information interaction between access points,a robust distributed algorithm is proposed to reduce the complexity of the centralized algorithm and improve the reliability of the distributed algorithm.The convergence of the proposed robust distributed algorithm is proved theoretically.Finally,the effectiveness of the three algorithms is proved by simulation.The results show that,compared with the original distributed algorithm,the robust distributed algorithm can effectively eliminate the influence of information interaction errors and coverage to the optimal value with a slightly higher complexity.Even so,the robust distributed algorithm has much lower computational complexity compared with the centralized algorithm.Second,aiming at maximizing the system energy efficiency of cache-aided ultra-dense networks(CUDNs),this dissertation studies the cache content placement strategies under the cache-aided ultra-dense networks(CUDNs)based on deep reinforcement learning(DRL).The intergration of caching and ultra-dense networks cannot only improve the efficiency of content retrieval by reducing duplicate content transmissions,but also improve the network throughput and system energy efficiency.Unlike the existing cache content placement strategies,this dissertation considers the dynamic and unknown nature of the cache content popularity,and a deep reinforcement learning based cache content placement algorithm has been proposed.This algorithm uses neural networks to select the best cache combination,and use the try-and-observe procedure to train the parameters of the neural networks according to the predefined reward function.Furthermore,the structure and corresponding parameters of the DRL based cache content placement algorithm have been optimized according to the latest findings in the field of DRL and deep learning(DL).The simulation results show that the improved cached content placement algorithm not only converges to the optimal energy efficiency value of the system at a faster speed,but also has better convergence stability,and it also avoids the prior information of the cached content popularity and meets the practical application requirements.Third,the dissertation studies the secure and reliable communication of the UDNs with respect to both the dynamic smart jammers to improve the system reliable energy efficiency.To cope with the influence of the smart jammers,we propose a deep reinforcement learning based anti-jamming strategy.Unlike other DRL based anti-jamming strategies,the proposed anti-jamming strategy is a deterministic policy gradient approach,which make deterministic anti-jamming action from the continuous action space,which avoids the discretization noise.Furthermore,we present our strategy in an actor-critic framework,for which a convoution neural network(CNN)is adopted as the actor for action selection,and a deep neural network(DNN)is used as the critic for value estimation.Simulation results show that the system performance can be significantly improved by the proposed strategy.Fourth,the dissertation studies the secure and reliable communication of UDNs and propose a optimization assistant training DRL approach to improve system secure and reliable energy efficiency.Firstly,the dissertation proposes an improved semi-definite relaxation(SDR)based algorithm,which improves the optimal convergence performance and reduces the computational complexity cost.And then,the dissertation proposes a continuous action strategy space decision algorithm based on the deterministic policy gradient(DPG),which overcomes the performance loss caused by the discrete operation of the existing DRL-basedalgorithm.Lastly,the dissertation combines the advantages of the two approaches,and an advanced DRL-based algorithm is proposed with the help of prior knowledge from the convex optimization approach to accelerate the training of the adopted neural networks.Simulation results show that,the advanced DRL-based algorithm overcomes the dynamic effects of both eavesdroppers and jammers and accelerates the neural network training while guaranteeing the SREE performance.Finally,the above mentioned theoretical results and algorithms can provide guidelines and potentially promoting the evolution of next generation mobile cellular network.