Data-Driven Design of 3GPP Handover Parameters with Bayesian Optimization and Transfer Learning

TL;DR

The paper tackles the challenge of mobility management in dense cellular networks by jointly optimizing 3GPP handover parameters, specifically A3-offset and Time-to-Trigger, using high-dimensional Bayesian optimization (HD-BO). It introduces TuRBO-based HD-BO with Gaussian Process surrogates and Thompson sampling to scale optimization to 60 variables across 30 cells, and extends this with transfer learning to generalize across UE speeds. Empirical results in a real-world London urban deployment show that HD-BO substantially reduces ping-pongs and radio link failures compared to 3GPP benchmarks, while maintaining strong SINR performance; transfer learning further accelerates adaptation with minimal loss. The work demonstrates the practicality of data-driven, site-specific mobility management for large-scale networks and points to future directions in beam-based mobility for mmWave and vertical handovers in integrated networks.

Abstract

Mobility management in dense cellular networks is challenging due to varying user speeds and deployment conditions. Traditional 3GPP handover (HO) schemes, relying on fixed A3-offset and time-to-trigger (TTT) parameters, struggle to balance radio link failures (RLFs) and ping-pongs. We propose a data-driven HO optimization framework based on high-dimensional Bayesian optimization (HD-BO) and enhanced with transfer learning to reduce training time and improve generalization across different user speeds. Evaluations on a real-world deployment show that HD-BO outperforms 3GPP set-1 and set-5 benchmarks, while transfer learning enables rapid adaptation without loss in performance. This highlights the potential of data-driven, site-specific mobility management in large-scale networks.

Figures (6)

• Figure 1: 2D view of the selected urban area showing UE positions (colored dots) and BS sites (triangles). Black lines indicate the five selected streets; colors represent the cell providing the strongest received power.
• Figure 2: RLFs and ping-pongs at different speeds under HD-BO (one-threshold and per-cell) and 3GPP baseline. Lower is better.
• Figure 3: Ping-pongs for UEs at different speeds with HD-BO optimized for a single target speed. Lower is better.
• Figure 4: SINR for UEs at various speeds under HD-BO ($w_{\text{PP}} = 1$, $w_{\text{RLF}} = 9$) and 3GPP baseline.
• Figure 5: Convergence of transfer learning with diverse UE speeds.