PhD defence of Ahmed Elmeligy – Priority-Aware Random Access Optimization in Massive Machine Type Communications Using Non-Uniform Preamble Selection

Abstract

The use of cellular networks for massive machine-type communications (mMTC) is attractive due to existing infrastructure. However, the large number of user equipments (UEs) in mMTC poses challenges to the random access channel (RACH) in terms of congestion and overloading. Existing RACH designs assume single-priority systems with uniform preamble selection, where each UE randomly selects a preamble from a set. To mitigate congestion, schemes such as UE backoff and preamble partitioning have been proposed, deferring transmissions or assigning subsets of preambles to different UE groups. However, after the backoff period or within a subset, UEs still select preambles uniformly, limiting flexibility. As a result, these methods provide partial relief and fail to meet QoS requirements under heavy network loads. To address this problem, we consider non-uniform preamble selection within each RACH slot and employ a multi-priority RACH system, where UEs are categorized into multiple priority classes with different QoS requirements. The system behavior is captured through access patterns observed at the base station, with non-uniform preamble selection probabilities providing greater flexibility in controlling access rates across classes compared to existing methods. We develop an optimization problem that determines the preamble selection probabilities to maximize the RACH throughput of high-priority UEs while ensuring low-priority UEs achieve a minimum throughput threshold. Since the optimal solution requires network load knowledge, we propose two load estimators based on the probability of observed access patterns over multiple RACH slots. An analytical framework is introduced to compute exact pattern probabilities, followed by a maximum likelihood estimator (MLE) and a reduced-complexity MLE (RCMLE). We then integrate the estimation and optimization frameworks, conducting sensitivity analyses of throughput under estimation errors and investigating the impact of non-uniform preamble selection on estimation accuracy.

Building on this, we reformulate the estimation problem as a multi-armed bandit (MAB) framework, making it suitable for larger networks with stochastic UE behavior. When closed-form expressions of throughput metrics are not required, we extend MAB to the optimization problem of determining preamble selection probabilities, where empirical approximations reduce complexity without requiring network load knowledge. We introduce two action space (AS) formulations and adopt a cross-entropy (CE)-based action selection policy. The framework is further extended to a deep-MAB (D-MAB) model that leverages neural networks for scalability in larger networks. To efficiently explore the resulting AS, we propose a hierarchical AS generation algorithm. Simulations demonstrate that the proposed frameworks achieve superior performance compared to baselines. The non-uniform preamble selection scheme consistently improves H-UE throughput while maintaining L-UE fairness, outperforming uniform preamble selection, access class barring, and preamble partitioning. The proposed estimators achieve high accuracy, with the RCMLE running 46 times faster than the standard MLE while incurring only minor degradation under heavy overloading. Building on these results, the MAB–based estimator demonstrates strong scalability, effectively handling larger networks with stochastic UE behavior while maintaining low mean absolute error and reducing computational complexity. Beyond estimation, the MAB formulation for optimizing preamble selection probabilities perform within 5% of the optimal non-uniform solution while requiring fewer computational resources, and the D-MAB extension further improves scalability by efficiently exploring large ASs. Together, these results confirm that the proposed frameworks outperform existing methods and provide a scalable solution for priority-aware RACH optimization in mMTC scenarios.