Neural Network–Based Adaptive Resource Allocation for 5G Heterogeneous Ultra-Dense Networks

Abstract

Increasing spectral bandwidth in 5G networks improves capacity but cannot fully address the heterogeneous and rapidly growing traffic demands. Heterogeneous ultra-dense networks (HUDNs) play a key role in offloading traffic across multi-tier deployments; however, their diverse base-station characteristics and diverse quality-of-service (QoS) requirements make resource allocation highly challenging. Traditional static resource-allocation approaches lack flexibility and often lead to inefficient spectrum utilization in such complex environments. This study aims to develop a joint user association–resource allocation (UA–RA) framework for 5G HUDNs that dynamically adapts to real-time network conditions to improve spectral efficiency and service ratio under high traffic loads. A software-defined networking controller centrally manages the UA–RA process by coordinating inter-cell resource redistribution through the lending of underutilized resource blocks between macro and small cells, mitigating repeated congestion. To further enhance adaptability, a neural network–adaptive resource allocation (NN–ARA) model is trained on UA–RA-driven simulation data to approximate efficient allocation decisions with low computational cost. A real-world evaluation is conducted using the downtown Los Angeles deployment. For performance validation, the proposed NN–ARA approach is compared with two representative baselines from the literature (Bouras et al. and Al-Ali et al.). Results show that NN–ARA achieves up to 20.8% and 11% higher downlink data rates in the macro and small tiers, respectively, and improves spectral efficiency by approximately 20.7% and 11.1%. It additionally reduces the average blocking ratio by up to 55%. These findings demonstrate that NN–ARA provides an adaptive, scalable, and SDN-coordinated solution for efficient spectrum utilization and service continuity in 5G and future 6G HUDNs.