A Hybrid Machine Learning Approach for Predictive Handover Management in 5G mmWave Networks Using the ACPT-LB Framework

In the invisible highways of modern communication, data travels as radio waves that must constantly jump from one tower to another as a user moves. This process, known as a handover, is the moment a phone switches its connection from one cell site to the next to maintain a clear signal. In the newest generation of wireless networks, which use extremely high-frequency waves to deliver blistering speeds, these jumps happen far more often and with much greater difficulty. The waves are so short that they cannot bend around corners or pass through walls, leaves, or even a human body. Consequently, a connection can vanish in a split second if something blocks the path, forcing the network to find a new tower instantly. If the network waits too long to switch, the call drops; if it switches too often or to the wrong tower, the connection stutters and wastes energy. For years, engineers have built systems that react to these problems only after they happen, trying to fix a broken link or stop a user from bouncing back and forth between two towers. The challenge has been to move from this reactive state to a proactive one, where the network can foresee a problem and fix it before the user even notices a glitch.

Researchers at The Federal University of Technology in Akure, Nigeria, have taken a significant step toward this goal by developing a new system called PACPT-LB. This framework builds upon an existing method known as ACPT-LB, which already helped towers cooperate to balance the number of users they serve. The new addition layers a machine learning pipeline on top of that foundation, allowing the network to predict three specific types of trouble before they occur: a failed handover, a radio link failure where the connection dies completely, and a "ping-pong" event where a user is forced to switch back and forth between two towers repeatedly. To teach this system, the researchers did not rely on real-world phone data, which can be messy and incomplete. Instead, they created a massive, highly detailed simulation of a 5G network using specialized software. They populated this virtual city with 183 cell towers and simulated the movement of up to 500 users traveling at speeds ranging from a slow walk to 160 kilometers per hour. Over the course of their experiments, they generated a dataset containing approximately 27 million records of network behavior, capturing every signal fluctuation, every blockage, and every decision the network made.

The core of their work involved training computer models to recognize patterns within this vast ocean of data. They tested several different types of algorithms, including simple decision trees that follow a set of rules, probabilistic models that calculate likelihoods, and linear models that look for straight-line relationships. They also combined these algorithms into hybrid teams to see if working together would improve their accuracy. The researchers found that the network's behavior was not uniform; the best approach depended entirely on what specific problem was being predicted. For predicting a failed handover, a combination of a decision tree and a probabilistic model proved to be the most reliable, correctly identifying the issue in nearly 98.5 percent of cases. When it came to predicting a complete loss of connection, a decision tree working alone performed best, achieving a success rate of over 98 percent. Perhaps most surprisingly, for the "ping-pong" problem, a simple probabilistic model alone was able to predict the event with perfect accuracy in many scenarios, because the conditions that cause a user to bounce between towers are distinct and easy to spot.

The study explicitly ruled out the idea that more complex models are always better. While the researchers tested hybrid combinations of three different algorithms, they found that adding more models to the team did not necessarily improve the results. In fact, for predicting radio link failures, combining models sometimes made the predictions slightly worse than using the single best model alone. This suggests that the quality of the individual algorithm matters more than the size of the team. The researchers also discovered that the speed of the user played a critical role in how well the system could learn. The models performed differently depending on whether the users were walking, driving at city speeds, or moving at highway velocities, indicating that a single static rule cannot manage all these conditions. The system learned that at high speeds, the network needs to act faster, while at low speeds, it can afford to wait longer, and the machine learning models successfully adapted to these shifting dynamics.

One of the most important findings of this research is the validation of the simulation itself. Before the machine learning could begin, the researchers had to prove that their virtual world behaved like a real network. They measured how often connections succeeded, how often users bounced between towers, and how stable the links remained. The results showed that while the system was not perfect, it produced realistic patterns of failure and success that varied logically with the number of users and their speed. This realism was crucial; if the simulation had been too perfect, the machine learning models would have had no difficult cases to learn from. By intentionally allowing the simulation to produce a range of outcomes, the researchers ensured their models were trained on data that reflected the messy reality of wireless communication. The final output is a framework that can now look at current network conditions and predict with high confidence whether a connection is about to fail, whether a user is about to get stuck in a loop, or whether a handover is likely to go wrong.

This work does not claim to have solved every problem in wireless networking, nor does it claim that these models work on real hardware yet. The results are based entirely on the simulation, and the researchers acknowledge that real-world factors like unpredictable traffic patterns and hardware delays could change the outcome. They also noted that they only tested three specific types of algorithms, leaving open the possibility that other, more complex models could perform even better. However, the study provides a clear roadmap for how to move from reactive network management to proactive prediction. By demonstrating that a hybrid approach can anticipate failures with high accuracy across a wide range of speeds and user densities, the researchers have shown that the network can soon learn to fix itself before the user ever drops a call. The path forward involves testing these predictions on physical testbeds and refining the models to handle even more complex real-world scenarios, but the foundation for a smarter, more anticipatory 5G network has been firmly laid.