Research On Multi-dimensional Resource Allocation Of Wireless Networks Based On Deep Reinforcement Learning

With the increasing demands for higher data rate,lower transmission delay and more reliable services,the integrated wireless networks with communications,caching and computing have become a developing trend for future networks.In the scenarios of complicated network architectures and multi-dimensional network resources,network control,user scheduling and resource allocation etc.are the main challenging technical issues.Introducing machine learning from the field of artificial intelligence into wireless networks,to facilitate network management and control,has become a hot topic recently.This dissertation exploits the current advances of machine learning and deep reinforcement learning algorithm,to tackle the network optimization and resource allocation issues in the integrated wireless networks with communications,caching and computing.The main contributions are as follows.(1)More realistic time-varying wireless channels are considered,where the time-varying wireless channels based on interference alignment are modelled as finite-state Markov channels.A deep-reinforcement-learning-based resource allocation scheme is proposed for cache-enabled opportunistic interference alignment networks.Within the proposed scheme,the formula of network capacity is given with the considerations of both perfect CSI and imperfect CSI.Aiming at maximizing the network capacity,the resource allocation problem is formulated as a reinforcement learning process,and the deep Q learning algorithm is exploited to obtain the optimal strategy.The performances of the proposed scheme with different SNR,different transition probabilities and imperfect CSI are presented via simulations.The simulation results show that the proposed scheme is applicable to complicated time-varying wireless communication scenarios.It can effectly reduce the backhaul link usage and allow more users to access interference alignment networks,and thus the network sum rate and energy efficiency are improved.(2)A software-defined virtulized framework for connected vehicles is designed,which integrates communication,caching and mobile edge computing.A deep reinforcement learning-based resource allocation scheme is proposed for the connected vehicles with the integration of communication,caching and computing.The dynamic change processes of the communication,caching,and computing resources are modeled as Markov chains,respectively.Aiming at maximizing the network operator's total utility,the joint resource allocation problem is formulated as a reinforcement learning process,and the deep Q learning algorithm is used to pursue the optimal solution.Without any assumptions about the objective funtions or any low-complexity preprocessing,the proposed scheme can directly solve the resource allocation problems with large-scale state space.Simulation results verify that the proposed scheme can converge at a fast speed,improve the network operator's total utilities,and possess the ability of resisting perturbation at a certain level.(3)A resource allocation scheme is proposed for future social networks based on the social trust model and a deep reinforcement learning approach.The social trust model utilizes uncertain reasoning to derive the trust values of D2D users,including the trust from direct observations based on Bayesian inference and the trust from indirect observations based on Dempster-Shafer theory.The effects of cache sizes and content types on backhaul link usage are analyzed,and the effectiveness of the proposed social trust scheme is verified via simulations that it can successfully track the variations of D2D users' trust values.The simulation results also show that the proposed scheme can reduce the effects of the malicious users on the network total utility.