Direction-aware handover control outperforms fixed thresholds in dense 5G HetNets
Dense 5G heterogeneous networks solve one problem and create another. Ultra-dense small cells increase capacity, coverage, and spectral efficiency, yet they also make mobility management far more fragile, especially when user equipment crosses between macro and small cells with different transmit power, coverage radius, and propagation conditions. In that setting, a fixed handover margin and a fixed time-to-trigger are too blunt for the task. Fast-moving devices need one kind of control; slow-moving devices and directionally asymmetric transitions need another.
The study centers on Event A3 handover, the standard trigger in which a target cell must outperform the serving cell by a prescribed handover margin for a set time-to-trigger interval before a switch occurs. That familiar mechanism remains intact throughout the comparison, along with the finite-state handover flow and radio-link-monitoring framework used to detect deterioration in the link. The practical question is not whether the standard matters. It is how the control parameters should move when mobility state and handover direction both change.
Mobility state and handover direction were built into the control logic
The proposed framework adds a direction-aware layer to handover parameter control. Instead of relying only on instantaneous signal strength, the control state explicitly includes user mobility state and the direction of the transition: macro to small, small to macro, or same-tier handovers. That design matters because handovers in heterogeneous networks are not symmetric. A macro-to-small transition can behave very differently from a small-to-macro transition, and a controller that treats them identically risks either triggering too late or switching too often.
Analytical adaptation and machine learning were both tested under the same standards-compatible structure. Four controllers were compared with identical settings: fixed A3, an analytical adaptive controller, tabular Q-learning, and an improved deep Q-network. The analytical controller optimized the handover margin and time-to-trigger using UE velocity and handover direction. The learning-based controllers used a multi-KPI reward to learn discrete handover-margin and time-to-trigger actions. Because the framework kept the conventional 5G handover machinery in place, the comparison stayed close to real-world network operation rather than drifting into a purely abstract optimization exercise.
The simulation model included 3D geometry and 3GPP propagation assumptions
The evaluation used a repeatable two-tier macro-small-cell deployment with 3D geometry, including user height, and channel modeling based on 3GPP TR 38.901 UMa and UMi propagation scenarios. That choice is important because handover behavior depends on more than distance. Base-station height, line-of-sight conditions, non-line-of-sight transitions, and the geometry of dense urban deployment all influence the signal asymmetry that drives handover decisions. The model therefore reflects the physical reasons a cell-edge user may hesitate, bounce back, or lose radio connection entirely.
Dense networks concentrate the risk at the edges of cell overlap. If the handover margin is too low, ping-pong handovers multiply. If the time-to-trigger is too long, the device may remain attached to the old cell for too long and suffer a late handover, handover failure, or even radio-link failure. The controls tested in this study were designed to sit directly inside that trade-off rather than outside it.
Tabular Q-learning improved robustness, while DQN gave the best overall balance
Across the tested UE velocities, tabular Q-learning improved the principal mobility-robustness KPIs relative to fixed A3. The improved deep Q-network delivered the best observed mean multi-KPI balance among the four controllers. In other words, the most flexible learning-based controller did not merely reduce a single mobility penalty; it produced the most even performance across the set of outcomes the authors tracked.
The article does not present the comparison as a simple winner-takes-all race. Each controller solves the same underlying problem with different trade-offs. The analytical method offers interpretability and direct parameter logic, which can be attractive in network engineering because the handover margin and time-to-trigger remain physically meaningful. Q-learning and DQN, by contrast, can capture more of the interaction between mobility state, handover direction, and observed network conditions, provided the reward design is sensible and the training process converges well.
A single controller cannot remove the asymmetry built into heterogeneous mobility. It can only respond to it more intelligently.
Why the learning result is plausible in the handover setting
The DQN result fits the structure of the problem. Handover control is a sequential decision task with state dependence, delayed consequences, and competing metrics. A setting like this naturally rewards a method that can learn state-action relationships across multiple mobility contexts rather than freezing the same threshold for every user and every direction of travel. By keeping the Event A3 framework and the finite-state logic intact, the learning system can optimize within familiar protocol boundaries instead of replacing them outright.
The multi-KPI reward is also clinically useful in an engineering sense, because mobility optimization is never a single-number problem. Throughput, latency, handover rate, ping-pong frequency, handover failure, radio-link failure, successful handover, and outage can all move in different directions if the controller is tuned too aggressively or too conservatively. A controller that balances those effects is more valuable than one that merely improves one metric while damaging another.
Descriptive certainty, not broad generalization, defines the conclusion
The policy-level comparison relied on three evaluation runs, so the uncertainty measures should be read descriptively rather than as a broad statistical claim. The findings are therefore limited to the tested simulation conditions. The dataset came entirely from the manuscript’s own simulation environment, with no external data source involved, and the authors reported no specific funding and no competing interests.
The practical message is narrow and concrete: dense 5G networks benefit when handover control is aware of both mobility state and handover direction, and when the optimization target remains tied to standard Event A3 behavior rather than to an entirely separate protocol. The analytic controller shows that interpretable adaptation can already improve mobility robustness. The DQN result suggests that a learning system can go further when it is trained on direction-resolved states and evaluated across multiple KPIs. What remains unresolved is how these results will hold under broader deployment conditions, larger policy sweeps, and more exhaustive repeat testing beyond the three runs used here.