Handover reduction in 5G mobile networks using ensemble learning method

The dense deployment of small cells in fifth-generation (5G) networks, while enhancing capacity, introduces frequent handovers that increase signaling overhead and degrade user experience. This paper presents an ensemble deep reinforcement learning (DRL) framework that reduces unnecessary handovers by integrating user mobility profiling with a Deep Q‑Network (DQN) agent. The proposed method first extracts high‑level features—including serving cell clusters and a handover‑benefit probability—using an unsupervised ensemble of K‑Means clustering, principal component analysis, and a decision tree. These features, together with real‑time signal measurements (RSRP, CINR, RSRQ), form a compact state representation. The DQN agent is pre‑trained offline on historical drive‑test data and employs a confidence margin to trigger handovers only when the expected gain exceeds a threshold. Evaluated on both real Brazilian drive‑test measurements and large‑scale 3GPP Urban Micro (UMi) realistic simulations (up to 60 cells and 3,000 users), the proposed method reduces handovers by 82% compared to the standard A3 event and by 45–54% compared to a conventional Q‑learning baseline, while maintaining comparable failure rates (< 3%) and latency (< 0.2 s). Throughput improves by up to 78% relative to Q‑learning. The inference latency stays below 1 ms per decision, confirming its suitability for real‑time deployment. This work demonstrates that combining unsupervised feature extraction with DRL offers a practical, data‑driven path toward efficient handover management in dense 5G small‑cell networks.