Resource- and Convergence-Aware Neural Algorithm Selection for 5G/B5G RAN

5G/B5G Radio Access Networks (RAN) deployment is demanding optimal selection mechanisms among the cutting-edge neural algorithms, especially embracing trade-offs between system performance and harsher resource constraints in consumer electronics services. However, current selection methods fail to support neural algorithms, due to their complexity and convergence uncertainty. To fill this gap, this paper proposes RECON-RAN, providing REsource- and COnvergence-aware Neural algorithm selection in 5G/B5G RAN. The mechanism comprehensively models critical deployment metrics, including training time, energy consumption, and memory usage, specifically tailored for neural algorithms in 5G/B5G consumer electronics environments. A novel convergence estimator is proposed for neural algorithms, enabling curvature-aware geometric estimation and empirical loss regression simultaneously. Simulation-based evaluation with a real 5G dataset showcased high reliability of the proposed mechanism, reflected by deviations of ±5% for static metrics and ±12% for dynamic metrics. A testbed with a 5G-enabled edge server validated RECON-RAN in an at-home video streaming scenario, where real-time processing achieved 93.00% accuracy, 4.01 J energy consumption, and 8.10 ms latency per iteration. Therefore, RECON-RAN provides a scalable and interpretable alternative to full-system simulation, supporting deployment-aware neural algorithm selection across heterogeneous, resource-constrained 5G/B5G RAN environments, which is essential for consumer-grade devices and platforms.