Research On Load Optimization For Wireless Networks Based On Machine Learning

In recent years,the issue of traffic fluctuation has been largely exacer-bated due to the dramatic growth of connected devices as well as the rise of innovative vertical services with stringent performance requirements.Consequently,load optimization,which aims at improving network per-formance by alleviating the influence from traffic fluctuations,is becoming increasingly important in next-generation wireless networks.Meanwhile,classical information-theoretical methods can no longer provide accurate mathematical models to deal with the exploding network information and dynamic optimization requirements.Therefore,it is necessary to develop new breeds of load optimization techniques for wireless networks.On the other hand,the emerging landscape of big data and machine learning has provisioned an intelligent and data-driven paradigm for the de-sign and optimization of wireless networks.Hence,this dissertation studies the machine learning empowered network architecture and optimization methods for the load optimization in next-generation wireless networks,which can deal with the dynamicity and diversity of wireless environment in a proactive manner.The main contributions of this dissertation are sum-marized as follows.First,this dissertation provides a systematic discussion on the architec-ture design for machine learning empowered wireless network.Specifical-ly,we first propose the data-cognition-empowered wireless network archi-tecture,which specifies how to perform intelligent and data-driven network optimizations by learning from the wireless big data.Then,we propose the scalable wireless network architecture built upon mobile cloud computing and mobile edge computing,which integrates scalable intelligence into the design and operation of next-generation wireless networks.The above two architectures offer the infrastructure foundation for the load optimization methods to be studied in the rest of this dissertation.Second,this dissertation studies the single-node load optimization methods,which aims at solving small-scale traffic prediction and load bal-ancing problems with a single learning machine.Particularly,we propose two types of load optimization schemes with different characteristics.The first is a combinational optimization scheme,which predicts future trends with machine learning models beforehand and then optimize the network with traditional models.Specifically,we first propose a Gaussian process(GP)-based prediction model that can predict future traffic variations.Ex-perimental results show that the proposed GP model can attain up to 97%prediction accuracy on real 4G data,and outperform the state-of-the-art models considerably.Then,we combine the proposed GP-based prediction model with traditional base station on-off models to perform load-aware base station on-off control for energy saving.The second is an all-in-one optimization scheme.Specifically,we propose a deep reinforcement learn-ing(DRL)-based load balancing model which can autonomously learn the optimal load balancing policy by interacting with the wireless environment and learning from the feedbacks.Experimental results show that the pro-posed model outperform existing methods about 20%,in terms of load bal-ancing performance.Third,this dissertation extends the above single-node load optimiza-tion methods to the multi-node load optimization methods,which can solve large-scale traffic prediction and load balancing problems based on parallel computing.Specifically,we first propose a multi-node prediction frame-work with a scalable GP-based prediction model.The proposed framework can improve the prediction speed considerably and outperform other low-complex prediction models about 15%.Then,we propose a multi-node load balancing framework with a scalable DRL-based model to handle large-scale load balancing tasks.The proposed framework can distribute the high computational complexity of large-scale load balancing to a bunch of local nodes.Meanwhile,the proposed scalable DRL-based model also enables joint exploration with multiple behavior policies as well as knowl-edge transfer among the learning agents,which further improve the learn-ing efficiency.Empirical results verify that the proposed multi-node load balancing framework outperforms the single-node load balancing method considerably.Fourth,this dissertation extends the above multi-node load optimiza-tion methods to the multi-agent load optimization methods,where distribut-ed learning agents can make their own decisions and cooperate with each other.Specifically,we propose a voting-based multi-agent reinforcement learning(MARL)method,where distributed learning agents interact with not only the environment but also the other agents.The agents coordinate with each other through voting and follow a proposed distributed primal-dual algorithm to obtain the global optimal solution.We prove that the distributed learning in our proposed method achieves the same sublinear convergence rate as centralized learning.In other words,the distributed decision making does not slow down the process of achieving global con-sensus on optimality.Finally,we verify the theoretical results through nu-merical simulations and conduct a case study of unmanned aerial vehicle assisted large-scale traffic offloading to justify the correctness and effec-tiveness of our proposed MARL method.The complexity of the above optimization problems and learning meth-ods increases step by step,with a growing network intelligence.The pro-posed architectures,frameworks,models and related theories offer solid foundations to approach the blueprint of intelligence-everywhere in next-generation wireless networks.Relevant results have been published in IEEE journals and IEEE conferences and applied for two national patents.