When Simple Is Enough: Resource-Aware Heuristic versus DDQN for O-RAN Mobility Robustness

When Simple Is Enough: Resource-Aware Heuristic versus DDQN for O-RAN Mobility Robustness Ali Deylamy∗ , and Nima Afraz∗

arXiv:2610.32257v1 [cs.NI] 23 Oct 2026

∗ School of Computer Science, University College Dublin, Dublin, Ireland

Abstract—Open Radio Access Network (O-RAN) enables mobility-management xApps, but a learned controller is not necessarily the most effective or efficient handover policy. This paper compares standard Event A3, a resource- and mobilityaware heuristic, and a Double Deep Q-Network (DDQN) through a common xApp-oriented interface and paired system-level simulation. The heuristic combines radio quality, target-cell resources, UE mobility, and service class, with weights optimized offline using NSGA-II, while DDQN-Multi uses temporal radio and resource observations. All controllers are evaluated on identical UE trajectories, traffic assignments, and ray-traced radio inputs under Base, High-Mobility, and High-UE scenarios. Performance is assessed using service, handover, resource, energy, and fairness KPIs, complemented by CRITIC–TOPSIS ranking and complexity analysis. Results show that the heuristic achieves the lowest outage in all three scenarios and ranks first in the CRITIC–TOPSIS comparison, while retaining linear online complexity and interpretable decisions. DDQN-Multi improves target feasibility under congestion but requires offline training and produces more handovers and ping-pong events. The results indicate that lightweight resource-aware control can provide a favourable performance-complexity trade-off relative to DDQN under the considered scenarios. Index Terms—O-RAN, xApp, mobility management, handover, DDQN, Event A3, near-RT RIC, CRITIC, TOPSIS, ray tracing.

weights, thresholds, and decision rules. Reinforcement learning provides an alternative by treating handover as a sequential decision problem and accounting for longer-term service and mobility outcomes. DDQN-based approaches are widely used for mobility management, but their training requirements, computational complexity, and limited interpretability motivate comparison with simpler resource-aware controllers. A meaningful comparison should also control for differences in mobility, traffic, and propagation conditions. Moreover, average goodput alone may hide important differences in lower-tail performance, outage, handover reliability, admission feasibility, and mobility stability. In O-RAN, controller selection should therefore consider not only service performance but also computational cost, training dependence, and interpretability, particularly when multiple xApps operate on shared RAN parameters [4]. In this work, we investigate whether a lightweight resource- and mobility-aware heuristic can provide a better overall handover trade-off than a more complex DDQN policy. We compare standard A3, the proposed heuristic, and DDQN-Multi through a common xApp-oriented decision interface using paired system-level simulations with common UE trajectories, traffic assignments, gNB configurations, and ray-traced received-power inputs.

I. I NTRODUCTION

The principal contributions of this work are as follows:

Reliable mobility management is essential for maintaining service continuity in fifth-generation (5G) networks. As a user equipment (UE) moves, its serving cell may become unsuitable because of propagation loss, interference, mobility, or resource congestion. A handover controller must therefore select a suitable target in time to avoid service degradation while limiting unnecessary handovers, interruption, signalling overhead, admission failures, and ping-pong events [1]. These objectives are often conflicting: aggressive policies may improve responsiveness but increase handover activity, whereas conservative policies may leave UEs connected to poorly performing cells. 3GPP Event A3 triggers handover reporting when a neighbouring-cell measurement exceeds that of the serving cell by a configured offset for a specified timeto-trigger (TTT) [2]. Although simple and interpretable, its radio-based formulation does not explicitly account for targetcell resources, UE mobility, or service conditions. Multicriteria heuristics can incorporate additional information such as RSRP, RSRQ, SINR, cell load, available PRBs, and UE motion [3], but their performance depends on the selected

We design a lightweight resource- and mobility-aware handover heuristic that jointly considers radio quality, targetcell load, PRB availability, UE motion, and service class. Its scoring weights are optimized offline using NSGA-II on training runs separate from evaluation, while mobilityadaptive TTT and context-dependent feasibility rules adapt decisions to UE conditions. Once configured, the heuristic requires no online learning and retains linear online complexity and directly interpretable decisions. • We develop a controlled paired evaluation framework for standard A3, the proposed heuristic, and DDQN-Multi under a common handover decision interface. All controllers receive identical UE trajectories, traffic assignments, gNB configurations, and cached ray-traced received-power inputs within each evaluated scenario, thereby controlling for exogenous mobility, traffic, and radio variation. • We evaluate the controllers using a multi-KPI methodology extending beyond mean goodput to lower-tail goodput, outage, degradation duration, handover reliability, ping-pong behaviour, admission performance, energy, and fairness. •

Paired inputs Base — High Mobility — High UE; fixed traces, traffic, and radio cache xApp observations RSRP, RSRQ, SINR, load, PRBs, velocity, and service class Decision engines A3 RSRP, hysteresis TTT

Heuristic radio, resources mobility

DDQN-Multi temporal state masked actions

Common handover simulation admission, execution, resource allocation, and NR goodput KPI assessment service, reliability, stability, resources, fairness; CRITIC–TOPSIS

Fig. 1. Paired evaluation workflow for A3, the resource-aware heuristic, and DDQN-Multi under a common handover simulation and KPI-assessment pipeline.

Raw KPIs are complemented by CRITIC–TOPSIS ranking and equal-weight sensitivity analysis, while a separate complexity and deployability analysis characterizes the trade-off between lightweight deterministic control and learned DDQN-based handover. A. Related Work DDQN-based handover has been widely studied in mobile and heterogeneous networks. Tan et al. [5] applied DDQN to V2N handover and compared it with conventional A3, while Wang and Li [6] considered DDQN-based handover in dual-connectivity mmWave heterogeneous networks. More recently, Mondal et al. [7] incorporated richer mobility context and explicit handover and ping-pong costs, and Liang et al. [8] introduced speed-adaptive DDQN for high-mobility heterogeneous networks. These studies primarily improve the learned handover policy itself. In contrast, this work examines whether the additional learning complexity is justified by comparing a multi-feature DDQN with standard A3 and an offlineoptimized, resource- and mobility-aware heuristic under the same handover interface and matched scenario conditions. II. S YSTEM M ODEL AND H ANDOVER C ONTROLLERS We model handover control as a near-RT RIC xApp, while keeping the evaluation independent of a full E2-connected ORAN deployment. The controller receives UE- and cell-level observations, including RSRP, RSRQ, SINR, UE velocity, cell load, PRB availability, service class, and serving/candidate cell identifiers. It then selects either a target cell or a stay action. This common xApp-oriented interface allows A3, the resource-aware heuristic, and DDQN-Multi to be compared as alternative decision engines for the same mobility-control task [9]. Figure 1 summarizes the paired workflow used to test whether a simple deterministic heuristic can provide a better handover trade-off than a more complex learned policy. All controllers use the same admission and handover procedure. A. Network and Radio Model The simulated network covers the campus of University College Dublin Belfield. Six geographically sourced tower

Fig. 2. The UCD Belfield simulation area, showing the six simulated NR gNBs and the initial positions of 100 UEs.

locations are used as NR gNB sites. The gNBs operate at 3.5 GHz, with 43 dBm conducted transmit power and 25 m antenna height. Carrier bandwidth is 100 MHz with 30 kHz subcarrier spacing and 273 PRBs. Each NR slot of 0.5 ms contains 14 OFDM symbols. Mobility, radio conditions, resource allocation and controller states are updated every 20 ms, i.e. 40 NR slots. Figure 2 shows the study area, six gNBs, and initial UE positions. Directional gNB antenna patterns are defined according to the 3GPP element abstraction with a maximum gain of 8 dBi, a horizontal and vertical beamwidth of 65°, a maximum attenuation of 30 dB, a polarisation slant of +45◦ and an electrical downtilt of 12° [10]. Site specific azimuths are used to orient the antenna boresights to the study area. The geographic propagation scene uses local OpenStreetMap building geometry, additional building-height information, automatic ray-tracer building-material assignment and loam terrain. The shooting-and-bouncing-ray propagation is set up with a small angular separation, up to two reflections and one edge diffraction. These propagation settings are system level modelling assumptions, not parameters imposed by 3GPP. Let Pu,c (t) denote the total ray-traced power received by UE u from cell c at time t. The serving-cell SINR is computed as γu (t) =

Nu +

P (t) P u,su , ρ c̸=su c (t)Pu,c (t)

(1)

where su is the serving cell, Nu is the thermal-noise power including the UE noise figure, and ρc (t) is the overlapping PRB activity of the interfering cell. Thus, the effective link quality is affected by both propagation and instantaneous utilisation of resources. Only controller-independent ray-traced received powers are cached; attachment, SINR, scheduling, goodput, and handover states are recomputed independently for each controller. Cache metadata prevents reuse across incompatible scenarios or radio configurations.

B. Resource Allocation and Model-Derived NR Goodput The scheduler dynamically allocates 273 PRBs among connected UEs at each system-level step. We consider two QoS-lite service classes: eMBB users request eight PRBs and latency-sensitive users request twelve PRBs. A 10% cell-level admission margin is maintained. All controllers use the same traffic, scheduler, and admission procedure. The primary KPI is model-derived expected NR goodput. The serving-cell SINR is mapped to the highest entry in 3GPP TS 38.214 MCS Table 5.1.3.1-1 whose spectral efficiency does not exceed the efficiency-adjusted Shannon bound [11]. The model then calculates the transport-block size from the allocated PRBs, 12 PDSCH symbols per slot, DM-RS overhead, and one transmission layer. The reported goodput is Gu (t) =

BuTBS (t) (1 − BLERtarget ) . Tslot

(2)

where Tslot = 0.5 ms and BLERtarget = 0.1. This value is calculated exactly within the implemented system-level model, but represents expected NR goodput at the configured target BLER rather than measured application throughput or realized HARQ delivery. C. Standard Radio-Only A3 The first controller is standard radio-only Event A3. For serving cell s and neighbouring cell c, the A3 condition is satisfied when L3 L3 Mu,c (t) > Mu,s (t) + HA3 ,

(3)

where M L3 is the Layer-3-filtered RSRP and HA3 = 3 dB is the configured hysteresis. The condition must remain continuously satisfied for a TTT of 640 ms before a request is issued. These parameters follow valid 3GPP measurementevent configuration choices [2]. The A3 decision stage uses radio measurements without pre-request PRB filtering, providing a transparent standardized baseline. Once triggered, the target proceeds through the common preparation, admission, execution, and completion process applied to every policy. D. Resource-Aware Heuristic The heuristic controller evaluates the radio, resource, load, and mobility suitability of each candidate cell. Its nominal score can be expressed as Su,c = wR R̄u,c + wγ γ̄u,c + wQ Q̄u,c − wL L̄c + wP P̄c , (4) where R̄, γ̄, and Q̄ are the normalized RSRP, SINR, and RSRQ, while L̄ and P̄ denote normalized cell load and available PRBs. The implementation clips all inputs to [0, 1] using RSRP, SINR, and RSRQ ranges of [−120, −60] dBm, [−10, 30] dB, and [−20, −10] dB, respectively. Cell load is normalized by the configured load limit, and available PRBs are normalized by the 273-PRB carrier capacity. Consequently, better radio quality and greater PRB availability

increase the score, whereas higher target-cell load reduces it. The implemented nominal weights are (wR , wγ , wQ , wL , wP ) = (0.356, 0.374, 0.058, 0.127, 0.087). (5) The nominal weights were tuned off-line, in a separate NSGA-II campaign, using Base-scenario training runs separate from the evaluation run reported in this paper. NSGA-II used a population size of 16 over six generations to maximise end-to-end handover completion jointly with minimising final handover failure and ping-pong rate [12]. The coefficients selected were then fixed and used as such in all evaluation scenarios. Nominal weights are adapted to the UE context and renormalised before candidate scoring. High-speed and latencysensitive UEs get additional SINR weight and eMBB UEs get additional PRB weight. The top-scoring candidate must beat the serving-cell score by an amount that depends on motion and serving quality, satisfy the radio-improvement gates, and pass the target-resource feasibility check. The nominal TTT is 480 ms and high-speed UEs use 200 ms. E. DDQN Mobility Policies DDQN controller observes the current serving cell and three strongest candidate cells. Each frame contains 64 features: 16 serving/network features and 16 for each of three candidates. Four consecutive frames form the 256-dimensional state. The four actions are either to keep the serving cell or to choose one from the three ranked candidates. Radio- and resource-validity masks discard infeasible selections. The online and target networks have two hidden layers with 128 and 64 fully connected units, respectively. Training is performed using Double Q-learning, prioritised experience replay, a replay buffer of 200,000 transitions, mini-batches of 64, discount factor γ = 0.99, learning rate 10−4 , gradient clipping, and a target-network update every 200 learning steps. The exploration rate is decayed from 0.80 to 0.05 during the training. The reward is driven by successfully delivered NR goodput. Defining Tu , Tth = 1 Mbps, (6) Tu + Tth The service reward is positive when the delivered goodput is above the service threshold, and it is a smooth degradation cost below the service threshold. The highest costs are incurred by final handover failures and ping-pong outcomes. The rewards obtained during preparation, execution and completion are stored in the semi-Markov transition and the agent can take into account the full handover interruption and outcome. DDQN-Multi was trained over eight 60-s episodes, which include two full cycles over Base, Challenging-Radio, HighMobility, and High-UE conditions. Training seeds were 201–204 and 211–214. For held-out seed 9001 (excluded from training), the checkpoints saved at the end of each complete cycle were evaluated over the four conditions. For UT (Tu ) =

TABLE I C ONTROLLER PERFORMANCE AND SCENARIO - SPECIFIC TOPSIS SCORES . P05 Out. Succ. PP PRB CRITIC Equal TP Scen. Ctrl. (Mbps) (Mbps) % HO % Fail.% TOPSIS TOPSIS Rank A3 2.827 0.734 10.34 132/135 25.76 2.22 0.458 0.606 2 1 Base Heuristic 2.834 0.864 8.93 270/270 30.74 0.00 0.775 0.786 DDQN-M. 2.778 0.734 11.77 369/375 36.04 0.27 0.138 0.073 3 A3 3.005 0.576 13.43 291/311 14.09 5.79 0.320 0.748 3 High Heuristic 2.980 0.576 13.38 489/496 22.49 0.00 0.865 0.711 1 Mob. DDQN-M. 2.927 0.432 16.94 734/758 28.61 0.53 0.506 0.157 2 A3 2.154 0.000 29.52 91/406 9.89 77.59 0.481 0.405 2 High Heuristic 2.229 0.432 26.49 253/272 27.67 6.25 0.704 0.789 1 UE DDQN-M. 2.135 0.000 29.38 541/566 39.00 3.36 0.419 0.286 3

these held-out evaluations we chose the checkpoint with the best mean NR goodput. The final paired controller evaluation was performed with seed 1001, independent of the training and checkpoint-validation seeds.

C. CRITIC–TOPSIS Multi-KPI Ranking Mobility policies can generate different trade-offs and, therefore, the raw KPIs are accompanied by a CRITIC–TOPSIS ranking. There are fourteen indicators: mean and fifth percentile goodput, total outage, 95th percentile degradation duration, QoS satisfaction, end-to-end handover success, successful handover interruption, pingpong ratio, handover attempts per UE-minute, post-handover RSRP gain, PRB satisfaction, energy per delivered Mbit, Jain fairness, and orphan/RLF events per UE-minute. For controller i and KPI j, min–max normalization promax duces a benefit-oriented value zij : (xij −xmin −xmin j )/(xj j ) max max min for benefit KPIs and (xj −xij )/(xj −xj ) for cost KPIs. CRITIC assigns greater importance to KPIs that provide both contrast among controllers and non-redundant information [14]:

III. E VALUATION M ETHODOLOGY A. Scenarios and Paired Experimental Design The evaluation considers three system-level simulation scenarios. In the Base scenario 100 UEs are moving according to a random-waypoint model with speeds between 0.5 m/s to 2 m/s. The High-Mobility scenario comprises 100 UEs moving at 8 m/s to 12 m/s with non-synchronized changes in direction, thus increasing the number of boundary crossings and decreasing the residence time in a cell. The High-UE scenario scales the number of UEs to 150, while the normal mobility model and per-UE offered PRB demand are retained, thus creating a resource-contention stress condition. Each simulation is run for 60 s with 3,000 system-level updates. Twenty percent of UEs are assigned to a controlled mobility-stress cohort that rotates around known handoverprone locations. All the policies have the same cohort assignment. This allows for repeatable exposure of cell-edge transitions while the other UEs assume the general mobility distribution of the scenario. B. Performance Indicators We evaluate service, radio, mobility, reliability, resource use, and fairness. The main service metrics are mean and fifthpercentile NR goodput, total service outage, 95th-percentile service-degradation duration, and the QoS-satisfaction ratio. Radio performance is defined by serving-cell RSRP, RSRQ, SINR and immediate post-handover radio gain. A handover attempt is successful if it completes the preparation, execution, completion and path switching. Its end-toend success rate is the number of completions divided by the total number attempted. Mobility stability is quantified by considering a successful-handover interruption, attempts per UE-minute, and ping-pongs per completed handover, where a ping-pong is a return to a recently serving cell within the configured observation interval. Resource efficiency is evaluated by PRB satisfaction, target-admission failures, cell-load balance and energy per delivered Mbit. User-level goodput fairness is evaluated using Jain’s fairness index based on the time-averaged goodput delivered to each UE [13].

Cj = σj

m X

(1 − ρjk ),

k=1

Cj wj = Pm

ℓ=1 Cℓ

,

(7)

where σj and ρjk are the normalized standard deviation and inter-KPI correlation, respectively. The weighted normalized matrix is vij = wj zij . TOPSIS then evaluates the distance of every controller from the empirical ideal and anti-ideal solutions [15]: sX ± Di = (vij − vj± )2 , j

D− Si = + i − , Di + Di

(8) 0 ≤ Si ≤ 1.

where a larger value indicates a policy closer to the ideal multi-KPI operating point. CRITIC–TOPSIS is accompanied by equal-weight TOPSIS to examine sensitivity to the data-derived weights. Because only three primary controllers and one paired evaluation are reported, the scenario-specific scores are interpreted as descriptive summaries of the observed KPI trade-offs. They are not treated as confidence intervals, statistical significance tests, or universal controller rankings. Multi-seed paired bootstrap analysis is reserved for an extended evaluation. IV. R ESULTS AND D ISCUSSION A. Overall Performance Table I compares A3, the heuristic, and DDQN-Multi using the same UE and radio inputs within each scenario. B. Base Scenario Figure 3 shows that the controllers produce broadly similar mean serving SINR and mean goodput, but substantially different mobility behaviour. The heuristic achieves the highest P05 goodput (0.864 Mbps) and lowest outage (8.93%) while avoiding target-admission failures. A3 performs fewer handovers and generally obtains larger post-HO RSRP gains, consistent with its conservative radio-only trigger. DDQN-Multi

Fig. 3. Sample-level KPI distributions across the paired 60-s evaluation.

instead initiates more handovers and exhibits the highest pingpong activity without a corresponding goodput improvement. The heuristic therefore occupies an intermediate operating point: it reacts more actively than A3 but avoids much of the mobility churn observed with DDQN-Multi.

that under congestion, target feasibility is important, but additional handover activity does not necessarily translate into better service continuity.

C. High-Mobility Scenario

Table I reports the scenario-specific multi-KPI scores. Scores are normalized independently within each scenario and are therefore not compared or averaged across scenarios. The heuristic ranks first under both weighting methods in Base and High UE. In High Mobility, CRITIC favours the heuristic’s reliability and admission performance, whereas equal weighting favours A3’s goodput and stability. This difference illustrates the sensitivity of composite rankings to KPI weighting and reinforces the need to interpret the scores together with the underlying raw KPIs.

High mobility highlights a different trade-off. A3 achieves the highest mean goodput (3.005 Mbps) and lowest pingpong ratio, while the heuristic achieves nearly identical mean goodput (2.980 Mbps) and outage (13.38% versus 13.43%) with higher handover completion and no admission failures. DDQN-Multi generates substantially more handover attempts and ping-pong events, while obtaining lower goodput and higher outage. Thus, when mobility rather than resource congestion dominates, conservative A3 remains a strong operating point, while resource awareness primarily improves target feasibility rather than radio performance. D. High-UE Scenario The High-UE scenario exposes the effect of resource awareness most clearly. Although A3 continues to select targets with relatively large post-HO RSRP gains, 77.59% of its handover requests fail target admission, reducing its end-to-end completion rate to 22.41%. The heuristic and DDQN-Multi reduce admission failure to 6.25% and 3.36%, respectively. However, DDQN-Multi requires 566 attempts and exhibits the highest ping-pong activity, whereas the heuristic requires only 272 attempts while achieving 93.01% completion. Consequently, the heuristic obtains the lowest outage (26.49%) and the only non-zero P05 goodput (0.432 Mbps). These results indicate

E. Multi-KPI Ranking

F. Computational Complexity and Deployability A3 is the least complex in online evaluations since the relative radio measurements are evaluated and the TTT state is maintained. Also the designed heuristic evaluates radio quality, resources, mobility and service class but its complexity is still O(C) since the feature set is fixed. DDQN-Multi requires a 256-dimensional temporal state, action masking and a 256–128–64–4 neural-network forward pass, which is 41, 216 multiply–accumulates per UE decision, excluding feature construction and masking. This comparison, along with the observed KPI results, places the heuristic as a lightweight and interpretable operating point between radio-only A3 and learned DDQN control. The trade-off is summarised in Table II.

TABLE II C ONTROLLER COMPLEXITY, OFFLINE PREPARATION , AND DEPLOYMENT CHARACTERISTICS . Controller

Decision information

A3

Serving- and candidate-cell RSRP, hys- O(C); no learned model parame- None; hysteresis and TTT are config- Lowest complexity and direct interesis, and TTT ters ured before operation terpretation of every decision

Online computation per UE

Offline preparation

Deployment characteristics

Resource-aware heuristic

RSRP, RSRQ, SINR, cell load, available O(CF ), equivalent to O(C) for Five score weights tuned by NSGA-II Lightweight execution without a PRBs, UE mobility, and service class the fixed feature set using a population of 16 over six gen- neural-network runtime; decision erations; coefficients are fixed during scores and feasibility gates remain evaluation interpretable

DDQN-Multi

Four stacked 64-feature frames, producing 41,412 trainable parameters and Eight 60-s training episodes followed Requires a trained checkpoint and a 256-dimensional state; three candidate 41,216 multiply–accumulates per by checkpoint evaluation using held- neural-network inference; supports cells and an action-validity mask forward pass, excluding activation, out seed 9001 temporal decisions but has lower feature-construction, and masking direct interpretability costs

C denotes the number of candidate cells and F the fixed number of features evaluated per candidate. All controllers operate at the common 20-ms system-level decision interval.

V. C ONCLUSION In this paper, we design and evaluate a resource-aware heuristic and a service-aware DDQN mobility policy for an ORAN xApp, taking radio-only 3GPP Event A3 as the reference baseline. A paired system-level framework ensured that each controller saw the same UE trajectories, traffic assignments, and cached ray-traced radio conditions. The results show that similar average goodput can hide large differences in outage, handover reliability, admission feasibility and mobility stability. The designed resource-aware heuristic achieves the best overall KPI balance and lowest outage in all three scenarios in the reported paired evaluation whilst maintaining linear online complexity, fixed coefficients and directly traceable decisions. A3 is still the simplest controller, but it is not resource-aware under congestion; DDQNMulti has strong target feasibility but generates more handover activities, is training-dependent and computationally complex. The results thus indicate lightweight resource-aware control as a promising candidate for mobility-management xApps. Future work will extend the evaluation to multiple paired seeds and an E2-connected near-RT RIC. VI. ACKNOWLEDGMENTS Funding for this research is provided by the European Union’s Horizon Europe research and innovation program through the Marie Skłodowska-Curie SE grant under the agreement RE-ROUTE No 101086343 and Research Ireland CONNECT-2 grant 13/RC/2077-P2. R EFERENCES [1] A. Haghrah, M. P. Abdollahi, H. Azarhava, and J. M. Niya, “A survey on the handover management in 5G-NR cellular networks: Aspects, approaches and challenges,” EURASIP Journal on Wireless Communications and Networking, vol. 2023, no. 52, 2023. [2] 3GPP, “NR; Radio Resource Control (RRC); Protocol specification,” 3rd Generation Partnership Project, TS 38.331, 2025, release 18. [3] M. R. Palas, M. R. Islam, P. Roy, M. A. Razzaque, A. Alsanad, S. A. AlQahtani, and M. M. Hassan, “Multi-criteria handover mobility management in 5G cellular network,” Computer Communications, vol. 174, pp. 81–91, Jun. 2021. [4] A. Wadud, F. Golpayegani, and N. Afraz, “AI-powered conflict management in Open RAN: Detection, classification, and mitigation,” Computer Networks, vol. 286, p. 112432, 2026.

[5] K. Tan, D. Bremner, J. L. Kernec, Y. A. Sambo, L. Zhang, and M. A. Imran, “Intelligent handover algorithm for vehicle-to-network communications with double-deep q-learning,” IEEE Transactions on Vehicular Technology, vol. 71, no. 7, pp. 7848–7862, 2022. [6] H. Wang and B. Li, “Double-deep q-learning-based handover management in mmwave heterogeneous networks with dual connectivity,” Transactions on Emerging Telecommunications Technologies, vol. 35, no. 1, p. e4907, 2024. [7] S. Mondal, S. Aslam, D. Khan, and K. Chang, “Intelligent handover orchestration in beyond 5G and urban V2X dual connectivity networks: A deep reinforcement learning approach,” IEEE Access, vol. 13, pp. 196 509–196 525, 2025. [8] G. Liang, F. Su, B. Lu, and H. Yu, “A speed-adaptive handover decision algorithm for ultra-dense heterogeneous wireless networks,” Ad Hoc Networks, vol. 192, p. 104313, 2026. [9] M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Understanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023. [10] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project, TR 38.901, 2025, release 19. [11] ——, “NR; Physical layer procedures for data,” 3rd Generation Partnership Project, TS 38.214, 2025, release 19. [12] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary Computation, vol. 6, no. 2, pp. 182–197, Apr. 2002. [13] R. Jain, D.-M. Chiu, and W. R. Hawe, “A quantitative measure of fairness and discrimination for resource allocation in shared computer systems,” Digital Equipment Corporation, Hudson, MA, USA, Tech. Rep. DEC-TR-301, 1984. [14] D. Diakoulaki, G. Mavrotas, and L. Papayannakis, “Determining objective weights in multiple criteria problems: The CRITIC method,” European Journal of Operational Research, vol. 55, no. 2, pp. 763– 770, 1995. [15] C.-L. Hwang and K. Yoon, Multiple Attribute Decision Making: Methods and Applications. Berlin, Germany: Springer-Verlag, 1981.

Related documents

• When Simple Is Enough: Resource-Aware Heuristic versus DDQN for O-RAN Mobility Robustness#1447624
• Network Mobility Support Terminology#500871
• Reinforcement Learning for Intensity Control: An Application to Choice-Based Network Revenue Management#139074
• Signalling Connection Control Part User Adaptation Layer (SUA)#132793
• AUSO: Action-Level Unified Skill Optimization from Internalization to Utilization#480085