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

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.