AI-assisted Signal Prediction Enables Reliable Handover in 6G Multi-RAT Networks with Fast Channel Dynamics

Future wireless networks, particularly those envisioned for 6G, demand seamless connectivity as users move between different radio access technologies, such as cellular and WiFi. Maria Lamprini A. Bartsioka, Anastasios Giannopoulos, and Sotirios Spantideas, all from Four Dot Infinity, present a new approach to managing these transitions, known as handover, by proactively predicting signal quality. Their research introduces a Machine Learning-assisted Predictive Conditional Handover framework that anticipates connectivity issues before they arise, rather than reacting to them as they happen. This predictive capability, achieved through carefully trained forecasting models, significantly reduces handover failures and the frustrating “ping-pong” effect where devices constantly switch between networks, paving the way for more reliable and efficient wireless experiences in future 6G deployments.

The system is designed to function in heterogeneous networks, combining 5G, Wi-Fi, and potentially other technologies, with the aim of improving Quality of Service (QoS), reducing latency, and enhancing the overall user experience. Scientists developed a system architecture, trained ML models for prediction, and defined a handover decision-making process. The core innovation lies in moving away from reactive handover, which waits for a signal to weaken before switching networks, to a proactive approach that anticipates network changes.

The system relies on ML models, including ARIMA and XGBoost, to predict key network metrics such as signal strength, channel quality, and user mobility. It also incorporates context-awareness, considering factors like user location, speed, and network conditions to make informed handover decisions. Researchers engineered a system that forecasts signal quality, enabling smoother transitions for mobile users. The study employed a hybrid Cellular/WiFi HetNet simulator to generate realistic datasets, modelling path loss, shadowing, fading, and interference as a user traverses dense deployments. The team trained specialized Long Short-Term Memory (LSTM) networks to predict signal-to-interference-plus-noise ratio (SINR) along randomized user trajectories, utilizing bidirectional LSTM models for cellular base stations and lightweight LSTM models for WiFi access points.

These models capture temporal dependencies in radio dynamics, providing accurate predictions of user equipment measurements essential for mobility control. A RAT Steering Controller combines user equipment reports, applies pre-trained SINR prediction models, and translates forecasts into actionable handover execution conditions. Experiments demonstrate consistent improvements in handover success rates and reduced latency, showcasing the effectiveness of the proactive approach. This work addresses the limitations of current handover methods, which react to signal changes rather than anticipating them, leading to dropped connections and service interruptions. The core of the P-CHO framework is a standardized workflow orchestrated by a RAT Steering Controller, which collects data, predicts signal quality for each available network, and makes handover decisions based on hysteresis-based conditions.

Researchers trained Long Short Term Memory (LSTM) networks to forecast signal quality indicators along randomized user trajectories, evaluating performance under various channel conditions for both cellular and WiFi networks. Comparative tests revealed that LSTM predictors maintain comparable error rates for low numbers of user paths and significantly outperform XGBoost predictors as the number of paths increases, demonstrating their effectiveness in handling more complex mobility patterns. Further analysis compared a “Soft P-CHO” approach, which triggers handovers based on immediate predicted gains, with a “Hysteresis-enabled P-CHO” variant that requires consistent predicted gains over multiple steps before executing a handover. Results demonstrate that the hysteresis-enabled scheme consistently yields fewer handovers than the soft P-CHO, improving stability and reducing signaling overhead. Experiments along a testing trajectory revealed that the hysteresis-enabled P-CHO maintains longer, contiguous segments on selected networks, suppressing rapid back-and-forth switching around cell borders. The team developed a system that proactively anticipates signal quality changes, enabling smoother transitions between networks before a connection is lost. This is achieved through the use of Long Short Term Memory networks, trained to forecast signal strength along predicted user trajectories, and integrated into a standardized handover workflow managed by a central controller. The results demonstrate that this predictive approach improves handover stability and reduces unnecessary network switching, particularly when incorporating a hysteresis mechanism to filter out short-term signal fluctuations.

Compared to traditional handover methods and baseline prediction models like ARIMA and XGBoost, the LSTM-based predictors exhibit greater accuracy and resilience to complex data. The team quantified the impact of various system parameters, finding that appropriate prediction windowing and minimizing trajectory overlap enhance prediction accuracy. Future work will focus on integrating this framework into Open Radio Access Networks, incorporating online learning to adapt to changing conditions, and exploring joint optimization of prediction and handover policies. Further research will also investigate model compression techniques to enable low-latency performance at the network edge and extend the prediction horizon for different user mobility patterns.