Improving next-generation wireless network performance and reliability with deep learning
Access full-text files
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
A rudimentary question whether machine learning in general, or deep learning in particular, could add to the well-established field of wireless communications, which has been evolving for close to a century, is often raised. While the use of deep learning based methods is likely to help build intelligent wireless solutions, this use becomes particularly challenging for the lower layers in the wireless communication stack. The introduction of the fifth generation of wireless communications (5G) has triggered the demand for “network intelligence” to support its promises for very high data rates and extremely low latency. Consequently, 5G wireless operators are faced with the challenges of network complexity, diversification of services, and personalized user experience. Industry standards have created enablers (such as the network data analytics function), but these enablers focus on post-mortem analysis at higher stack layers and have a periodicity in the time scale of seconds (or larger). The goal of this dissertation is to show a solution for these challenges and how a data-driven approach using deep learning could add to the field of wireless communications. In particular, I propose intelligent predictive and prescriptive abilities to boost reliability and eliminate performance bottlenecks in 5G cellular networks and beyond, show contributions that justify the value of deep learning in wireless communications across several different layers, and offer in-depth analysis and comparisons with baselines and industry standards. First, to improve multi-antenna network reliability against wireless impairments with power control and interference coordination for both packetized voice and beamformed data bearers, I propose the use of a joint beamforming, power control, and interference coordination algorithm based on deep reinforcement learning. This algorithm uses a string of bits and logic operations to enable simultaneous actions to be performed by the reinforcement learning agent. Consequently, a joint reward function is also proposed. I compare the performance of my proposed algorithm with the brute force approach and show that similar performance is achievable but with faster run-time as the number of transmit antennas increases. Second, in enhancing the performance of coordinated multipoint, I propose the use of deep learning binary classification to learn a surrogate function to trigger a second transmission stream instead of depending on the popular signal to interference plus noise measurement quantity. This surrogate function improves the users' sum-rate through focusing on pre-logarithmic terms in the sum-rate formula, which have larger impact on this rate. Third, performance of band switching can be improved without the need for a full channel estimation. My proposal of using deep learning to classify the quality of two frequency bands prior to granting the band switching leads to a significant improvement in users' throughput. This is due to the elimination of the industry standard measurement gap requirement—a period of silence where no data is sent to the users so they could measure the frequency bands before switching. In this dissertation, a group of algorithms for wireless network performance and reliability for downlink are proposed. My results show that the introduction of user coordinates enhance the accuracy of the predictions made with deep learning. Also, the choice of signal to interference plus noise ratio as the optimization objective may not always be the best choice to improve user throughput rates. Further, exploiting the spatial correlation of channels in different frequency bands can improve certain network procedures without the need for perfect knowledge of the per-band channel state information. Hence, an understanding of these results help develop novel solutions to enhancing these wireless networks at a much smaller time scale compared to the industry standards today