UC Irvine

Every year, network operators spend hundreds of millions of dollars to improve cellular capacities. Capacity improvements typically aim at

adding carriers, frequencies, bands, radios, distributed antenna systems, small cells, and cell towers. In many cases, user traffic and network load associated with special events complicate the issues of tracking, congestion, and degrading quality of service. Often times, operators

handle special events by dedicating human resources to manually solve them instead of deploying automated solutions.

In this thesis, we first use a pair of supervised learning approaches to model Universal Mobile Telecommunication System (UMTS) cellular network capacity measured in terms of total number of users carried and then predict breakpoints of cellular towers as a function of network traffic loading. Similarly, we utilize a supervised deep learning technique to predict the Long Term Evolution (LTE) network loading of connected users and then dynamically predict the congestion threshold of each cellular tower under offered load. Next, we formulate an optimization problem to maximize UMTS network capacity subject to constraints of user quality and predicted breakpoints. For LTE, we use the predicted congestion thresholds together with quality constrains to fine-tune cellular network operating parameters leading to minimizing overall network congestion. We investigate a few algorithmic alternatives including Simulated Annealing (SA), Hill Climbing, Regression, and Genetic Algorithm (GA). We propose a novel variant of simulated annealing referred to as Block Coordinated Descent Simulated Annealing (BCDSA) to solve the problem. Our performance measurements show that BCDSA offers dramatically improved algorithmic success rate and best characteristics in utility, runtime, and confidence range measures compared to alternative solutions for both problems in UMTS and LTE. It is observed that BCDSA is up to an order of magnitude faster than other algorithms and offers success rates twice better than other algorithms in finding best solutions.