PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ali Deylamy, Nima Afraz

arXiv:2610.32257v1cs.NI

TL;DR

The paper examines whether a lightweight resource- and mobility-aware handover heuristic can offer a better trade-off than a more complex learned policy. It compares A3, the heuristic, and DDQN-Multi in paired system-level simulations; the heuristic has the lowest outage across the three scenarios and the best overall KPI balance in the reported evaluation.

  • Problem

    The paper asks whether a lightweight resource- and mobility-aware heuristic can provide a better overall handover trade-off than a more complex DDQN policy.

    Source: §I · p. 1

  • Method

    The study compares A3, a heuristic using radio, resource, mobility, and service context, and DDQN-Multi through a common xApp-oriented interface and paired simulations with matched inputs.

    Source: §I · p. 1, §V · p. 6

  • Results

    The heuristic achieves the lowest outage in all three scenarios and the best overall KPI balance in the reported paired evaluation.

    Source: §V · p. 6

  • Takeaways & Limitations

    The results indicate that lightweight resource-aware control is a promising mobility-management xApp candidate, while DDQN-Multi has strong target feasibility but more handover activity and training and computational requirements.

    Source: §V · p. 6

Abstract

from arXiv · show

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 mobility-aware 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.

I. INTRODUCTION

The paper asks whether a lightweight resource- and mobility-aware heuristic can provide a better handover trade-off than a more complex DDQN policy. It motivates a controlled, multi-KPI comparison of A3, the heuristic, and DDQN-Multi, and positions the study against prior work focused mainly on learned-policy improvements.

I. INTRODUCTION

  • Handover objectives can conflict: aggressive policies may increase handover activity, while conservative policies may leave UEs on poorly performing cells.A3 uses a radio-based trigger and does not explicitly account for target-cell resources, UE mobility, or service conditions.Source: §I · p. 1
  • The heuristic jointly considers radio quality, target-cell load and PRB availability, UE motion, and service class, with offline-optimized weights and interpretable, linear-complexity online decisions.The weights are optimized using NSGA-II on training runs separate from evaluation; the configured heuristic requires no online learning.Source: §I · p. 1
  • The study compares A3, a resource- and mobility-aware heuristic, and DDQN-Multi using paired simulations with matched trajectories, traffic, configurations, and ray-traced inputs.The matched inputs control for exogenous mobility, traffic, and radio variation.Source: §I · p. 1
  • The comparison evaluates service, handover, resource, energy, and fairness KPIs, extending beyond mean goodput to lower-tail performance and mobility stability.The paper also complements raw KPIs with CRITIC–TOPSIS ranking, sensitivity analysis, and complexity analysis.Source: §I · pp. 1–2

A. Related Work

  • Prior DDQN handover studies primarily improve learned policies, whereas this work tests whether their additional learning complexity is justified against A3 and an offline-optimized heuristic under matched conditions.The cited examples include DDQN approaches for V2N, dual-connectivity mmWave, richer mobility context, and high-mobility networks.Source: A. Related Work · p. 2

II. SYSTEM MODEL AND HANDOVER CONTROLLERS

The system models handover as a near-RT RIC xApp decision among target-cell selection or staying, using a common interface and handover procedure. Its radio model uses a six-gNB campus scenario, ray-traced received powers, and SINR affected by propagation and interfering-cell resource use.

II. SYSTEM MODEL AND HANDOVER CONTROLLERS

  • A common xApp-oriented interface supplies UE- and cell-level observations and lets A3, the heuristic, and DDQN-Multi choose a target cell or stay under the same admission and handover procedure.The evaluation models the xApp without requiring a full E2-connected O-RAN deployment.Source: §II · p. 2

A. Network and Radio Model

  • The simulated UCD Belfield campus uses six NR gNB sites, with 100 UEs shown at their initial positions in Figure 2.The sites operate at 3.5 GHz with 100 MHz carrier bandwidth; mobility, radio conditions, resource allocation, and controller states update every 20 ms.Source: Figure 2 · p. 2, A. Network and Radio Model · p. 2
  • The radio model uses local building geometry and ray tracing with up to two reflections and one edge diffraction; these propagation settings are system-level assumptions, not 3GPP requirements.Only controller-independent ray-traced received powers are cached; attachment, SINR, scheduling, goodput, and handover states are recomputed for each controller.Source: A. Network and Radio Model · p. 2
  • Serving-cell SINR depends on received power, thermal noise, and overlapping PRB activity from interfering cells, so effective link quality reflects both propagation and resource utilization.The interfering-cell activity is represented by ρc(t).Source: A. Network and Radio Model · p. 2

B. Resource Allocation and Model-Derived NR Goodput

The system allocates PRBs among connected UEs and estimates expected NR goodput from serving-cell SINR and transport-block parameters. It uses radio-only Event A3 as a standardized baseline, with a configured hysteresis and time-to-trigger before a handover request.

B. Resource Allocation and Model-Derived NR Goodput

  • The scheduler allocates 273 PRBs dynamically, while eMBB and latency-sensitive users request eight and twelve PRBs, respectively, under a 10% cell-level admission margin.All controllers use the same traffic, scheduler, and admission procedure.Source: B. Resource Allocation and… · p. 3
  • Expected NR goodput is derived by mapping SINR to an eligible MCS and calculating transport-block size from allocated PRBs, PDSCH symbols, DM-RS overhead, and one transmission layer.The model assumes a target BLER of 0.1 and a 0.5 ms slot; its goodput is not measured application throughput or realized HARQ delivery.Source: B. Resource Allocation and… · p. 3

C. Standard Radio-Only A3

  • Standard radio-only Event A3 issues a request after its condition remains continuously satisfied for a TTT of 640 ms; the configured hysteresis is HA3 = 3 dB.These parameters follow valid 3GPP measurement-event configuration choices.Source: C. Standard Radio-Only A3 · p. 3
  • A3 uses radio measurements without pre-request PRB filtering, then sends the selected target through the common preparation, admission, execution, and completion process.Source: C. Standard Radio-Only A3 · p. 3

D. Resource-Aware Heuristic

The heuristic scores candidate cells using radio, resource, and load features, with nominal weights tuned offline and then fixed across evaluation scenarios. It adapts weights and TTT to UE context and requires candidate score improvement, radio gates, and target-resource feasibility.

D. Resource-Aware Heuristic

  • The nominal candidate-cell score adds normalized RSRP, SINR, and RSRQ contributions, subtracts target-cell load, and adds available-PRB contribution.The corresponding weights are wR, wγ, wQ, wL, and wP.Source: D. Resource-Aware Heuristic · p. 3
  • The score’s normalized inputs are clipped to [0, 1], and greater radio quality or PRB availability raises the score while higher target-cell load lowers it.Available PRBs are normalized by the 273-PRB carrier capacity.Source: D. Resource-Aware Heuristic · p. 3
  • The nominal weights were tuned offline with NSGA-II on separate Base-scenario training runs, then fixed for all evaluation scenarios.The optimization jointly maximized end-to-end handover completion and minimized final handover failure and ping-pong rate.Source: D. Resource-Aware Heuristic · p. 3
  • Weights are adapted and renormalized for UE context; a candidate must beat the serving-cell score, satisfy radio-improvement gates, and pass target-resource feasibility.High-speed and latency-sensitive UEs receive additional SINR weight, eMBB UEs additional PRB weight, and high-speed UEs use a 200 ms TTT rather than the nominal 480 ms.Source: D. Resource-Aware Heuristic · p. 3

E. DDQN Mobility Policies

DDQN-Multi uses temporal observations to choose among the serving cell and candidate cells, with masks excluding infeasible selections. Its training uses Double Q-learning and replay, while the final paired evaluation uses a seed independent of training and checkpoint validation.

E. DDQN Mobility Policies

  • Four consecutive frames of 64 features form the 256-dimensional state, and four actions select the serving cell or one of three ranked candidates.Radio- and resource-validity masks discard infeasible selections.Source: E. DDQN Mobility Policies · p. 3
  • The online and target networks use two hidden layers of 128 and 64 units, with prioritised experience replay and Double Q-learning.Training uses a 200,000-transition replay buffer, batches of 64, γ = 0.99, and target-network updates every 200 learning steps.Source: E. DDQN Mobility Policies · p. 3
  • The reward favors delivered goodput above the service threshold, degrades smoothly below it, and assigns the highest costs to final handover failures and ping-pong outcomes.Preparation, execution, and completion rewards are stored in the semi-Markov transition so the agent can account for interruption and outcome.Source: E. DDQN Mobility Policies · p. 3
  • The final paired evaluation used seed 1001, independent of the training and checkpoint-validation seeds.DDQN-Multi training comprised eight 60-s episodes across Base, Challenging-Radio, High-Mobility, and High-UE conditions; held-out seed 9001 was used for checkpoint evaluation.Source: E. DDQN Mobility Policies · pp. 3–4

III. EVALUATION METHODOLOGY

The evaluation pairs controllers across three simulated scenarios and a shared mobility-stress cohort. It assesses service, radio, mobility, reliability, resource use, energy, and user-level fairness.

A. Scenarios and Paired Experimental Design

  • The High-Mobility scenario uses 100 UEs at 8–12 m/s, while High-UE increases the population to 150 under normal mobility and per-UE PRB demand.Base uses 100 UEs moving at 0.5–2 m/s; the High-UE scenario creates resource contention.Source: A. Scenarios and Paired Experimental… · p. 4
  • Each 60-s simulation has 3,000 updates, with 20% of UEs in a shared mobility-stress cohort rotating through handover-prone locations.The shared cohort assignment provides repeatable exposure to cell-edge transitions.Source: A. Scenarios and Paired Experimental… · p. 4

B. Performance Indicators

  • Performance indicators cover mean and fifth-percentile goodput, outage, degradation duration, QoS satisfaction, and serving-cell radio measures.Radio measures include RSRP, RSRQ, SINR, and immediate post-handover radio gain.Source: B. Performance Indicators · p. 4
  • Mobility and resource indicators include handover interruption, attempts, ping-pongs, end-to-end success, PRB satisfaction, admission failures, load balance, and energy per delivered Mbit.User-level goodput fairness is measured with Jain’s fairness index.Source: B. Performance Indicators · p. 4

C. CRITIC–TOPSIS Multi-KPI Ranking

CRITIC–TOPSIS combines fourteen benefit-oriented KPI indicators into scenario-specific controller rankings. The scores are descriptive of the observed trade-offs, not statistical tests or universal rankings.

C. CRITIC–TOPSIS Multi-KPI Ranking

  • The ranking uses fourteen indicators spanning goodput, outage, degradation, QoS, handover reliability and stability, radio gain, PRB satisfaction, energy, fairness, and orphan/RLF events.CRITIC–TOPSIS accompanies the raw KPIs because mobility policies can involve different trade-offs.Source: C. CRITIC–TOPSIS Multi-KPI Ranking · p. 4
  • CRITIC weights indicators for both contrast among controllers and non-redundant information, using normalized standard deviation and inter-KPI correlation.The weighted normalized values are then evaluated by TOPSIS against empirical ideal and anti-ideal solutions; a larger score means closer to the ideal operating point.Source: C. CRITIC–TOPSIS Multi-KPI Ranking · p. 4
  • Only three primary controllers and one paired evaluation are reported, so scenario-specific scores are descriptive rather than confidence intervals, significance tests, or universal rankings.Equal-weight TOPSIS examines sensitivity to data-derived weights; multi-seed paired bootstrap analysis is reserved for extended evaluation.Source: C. CRITIC–TOPSIS Multi-KPI Ranking · p. 4

IV. RESULTS AND DISCUSSION

In the Base scenario, the heuristic records the highest fifth-percentile goodput and lowest outage while avoiding target-admission failures. Its mobility behavior lies between conservative A3 and the higher churn of DDQN-Multi.

A. Overall Performance

  • Table I compares A3, the heuristic, and DDQN-Multi using the same UE and radio inputs within each scenario.Its multi-KPI scores are scenario-specific, not compared or averaged across scenarios.Source: A. Overall Performance · p. 4

B. Base Scenario

  • 0.864 Mbps P05 goodput and 8.93% outage: the heuristic leads on these Base-scenario service outcomes while avoiding target-admission failures.Figure 3 shows broadly similar mean serving SINR and mean goodput, but substantially different mobility behavior.Source: B. Base Scenario · p. 4, Figure 3 · p. 5
  • A3 performs fewer handovers and generally has larger post-handover RSRP gains, while DDQN-Multi produces more handovers and the highest ping-pong activity without corresponding goodput improvement.The heuristic reacts more actively than A3 while avoiding much of DDQN-Multi’s mobility churn.Source: B. Base Scenario · p. 5

C. High-Mobility Scenario

The High-UE results show that the heuristic combines high completion with the lowest outage, while the multi-KPI ranking generally favours it. Under high mobility, A3 remains competitive on goodput and ping-pong stability.

C. High-Mobility Scenario

  • In High Mobility, A3 has the highest mean goodput (3.005 Mbps) and lowest ping-pong ratio, while the heuristic delivers 2.980 Mbps mean goodput and 13.38% outage versus 13.43%.The heuristic also has higher handover completion and no admission failures; DDQN-Multi has lower goodput and higher outage.Source: C. High-Mobility Scenario · p. 5

D. High-UE Scenario

  • In High UE, the heuristic achieves 93.01% completion, the lowest outage (26.49%), and the only non-zero P05 goodput (0.432 Mbps).DDQN-Multi reduces admission failure to 3.36% versus 6.25% for the heuristic, but requires 566 attempts compared with the heuristic’s 272.Source: D. High-UE Scenario · p. 5
  • Under congestion, target feasibility matters, but more handover activity does not necessarily improve service continuity.Source: D. High-UE Scenario · p. 5

E. Multi-KPI Ranking

  • The heuristic ranks first under both weighting methods in Base and High UE, while High-Mobility rankings differ with KPI weighting.CRITIC favours the heuristic’s reliability and admission performance in High Mobility, whereas equal weighting favours A3’s goodput and stability.Source: E. Multi-KPI Ranking · p. 5

F. Computational Complexity and Deployability

The heuristic is presented as a lightweight, interpretable operating point between radio-only A3 and learned DDQN control. In the paired evaluation, it achieves the lowest outage in all three scenarios while retaining linear online complexity.

F. Computational Complexity and Deployability

  • The heuristic’s fixed feature set gives it O(C) online complexity, between A3 and DDQN-Multi’s 41,216 multiply–accumulates per UE decision.The DDQN-Multi count excludes feature construction and masking; its decision uses a 256-dimensional temporal state and a 256–128–64–4 network.Source: F. Computational Complexity and… · p. 5

V. CONCLUSION

  • The paired framework evaluates controllers on identical UE trajectories, traffic assignments, and cached ray-traced radio conditions.Source: §V · p. 6
  • The heuristic achieves the best overall KPI balance and lowest outage in all three scenarios while retaining linear online complexity and directly traceable decisions.A3 is simplest, while DDQN-Multi has strong target feasibility but more handover activity, training dependence, and greater computational complexity.Source: §V · p. 6
Loading 2610.32257v1…