PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Digital Twin-Assisted Multimodal Predictive Handover

Afan Ali, Ali Arshad Nasir, Naveed Iqbal, Daniel Benevides da Costa

arXiv:2610.33833v1eess.SP

TL;DR

Rapid mobility, blockage, and propagation changes can make instantaneous handover decisions unreliable, while existing approaches remain insufficiently connected to site-specific dynamics and event-triggered procedures. The paper proposes a multimodal Agentic DT with three agents that select forecasts, plan handovers, and adapt decisions using execution feedback. It provides the most balanced performance across handover activity, link validity, and target-selection alignment, with approximately 98.2% mean first-step ray-path availability.

  • Problem

    Rapid mobility, blockage, and beam-alignment changes can degrade serving links and complicate target selection, while existing approaches remain insufficiently connected to site-specific handover dynamics and event-triggered procedures.

    Source: §I · p. 1

  • Method

    A site-specific Digital Twin synchronizes mesh-derived geometry, ray-traced multipath, link validity, and UE mobility, while three task-specific agents select forecasts, plan handovers, and adapt actions using execution feedback.

    Source: §I · p. 1, §V · p. 6

  • Results

    The Agentic Twin provides the most balanced performance across handover activity, link validity, and target-selection alignment, with approximately 98.2% mean first-step ray-path availability.

    Source: C. Results and Analysis · pp. 5–6

  • Takeaways & Limitations

    The model-agnostic framework offers a site-specific approach to predictive handover that combines multimodal forecasting with planning and execution-feedback adaptation.

    Source: §I · p. 1, §V · p. 6

Abstract

from arXiv · show

Rapid mobility, blockage, and propagation changes can invalidate instantaneous handover decisions in dense mid-band and millimeter-wave (mmWave) networks. We therefore propose a flexible, model-agnostic \emph{Agentic Digital Twin (DT)} framework for predictive cell handover. The DT generates synchronized multimodal states, while three agents autonomously select forecasts, plan handover actions, and adapt decisions using execution feedback. The proposed Agentic DT provided the best collective balance among handover activity, the reliability of the selected links, and the accuracy of target next-generation Node B (gNB) selection relative to the best available future link. It reduced average handovers from $11.0$ for Event~A3 to $8.0$ per trajectory while maintaining $98.2\%$ ray-path availability.

I. INTRODUCTION

Rapid changes in mobility, blockage, and beam alignment complicate reliable handover, while existing approaches remain insufficiently connected to site-specific dynamics and event-triggered procedures. The paper proposes a multimodal Agentic DT that synchronizes site data and uses agents to select, plan, validate, and adapt handover actions.

I. INTRODUCTION

  • Rapid mobility, blockage, and beam-alignment changes can degrade serving links and complicate target selection; reliable prediction requires wireless measurements and environmental context.Conventional Event A3 uses hysteresis and a time-to-trigger interval [1]1.Source: §I · p. 1
  • Existing predictive handover approaches address only parts of the prediction–decision–action loop, leaving a unified approach beyond fixed hysteresis and TTT insufficiently explored.Prior directions include DT-assisted suitability evaluation with simplified channel and sensing models, agentic Tiny ML, and sensing-assisted feature fusion.Source: §I · p. 1
  • The proposed Agentic DT synchronizes site-specific multimodal data and autonomously selects, plans, validates, and adapts handover actions beyond fixed hysteresis and TTT rules.Its reproducible DT pipeline derives virtual LiDAR and geometric features from site meshes and synchronizes them with ray-traced multipath, link validity, and UE mobility.Source: §I · p. 1
  • The current implementation uses task-specific forecasting, planning, and adaptation agents, with their policies and number designed to expand with site heterogeneity, propagation clutter, and computational resources.The paper describes this implementation as a preliminary step toward more autonomous handover management and physical AI-enabled wireless systems.Source: §I · p. 1
  • The Agentic DT provides the most balanced performance across handover activity, link validity, and target-selection alignment compared with reactive Event A3 and non-agentic fixed policies.The comparison considers the three objectives jointly.Source: §I · p. 1

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system model combines route-level UE mobility, site-mesh geometry, and ray-traced radio features in a synchronized digital-twin observation. It formulates a simplified Event A3 comparison using filtered path power because the simulation does not provide standardized RSRP.

A. Network and Digital Twin Model

  • The site-specific model fixes candidate gNB positions and orientations and represents UE position, velocity, acceleration, and orientation at each route index.Velocity and acceleration use consecutive route samples; UE height follows local terrain.Source: A. Network and Digital Twin Model · p. 2
  • Mesh-derived geometric features provide virtual-LiDAR, range, depth, structural, and line-of-sight context in a portable, extensible DT pipeline.The representation can be extended with real-time sensing and network measurements.Source: A. Network and Digital Twin Model · p. 2

B. Ray-Tracing Channel Model

  • Ray-traced channels are represented across candidate gNBs, route indices, and subcarriers, with path coefficients and delays defining the multipath components.The narrowband approximation evaluates the transmit and receive array responses at the carrier frequency while retaining subcarrier dependence in the delay term.Source: B. Ray-Tracing Channel Model · p. 2
  • The link-validity indicator records whether a candidate has a valid path, while the path-power feature sums path-coefficient power in dB.The power values are clipped at −220 dB for training; this ray-tracing feature is not standardized RSRP or calibrated received power.Source: B. Ray-Tracing Channel Model · p. 2
  • The study uses path power and link validity, reserving full beamwise channel, SINR, and achievable-rate analysis for future beam-aware evaluation.The present model therefore does not evaluate those beam-aware performance measures.Source: B. Ray-Tracing Channel Model · p. 2
  • The observation combines current mobility, radio, geometry, and link feasibility, while Table I summarizes its multimodal blocks.The separate validity vector distinguishes unavailable links from weak but valid links.Source: B. Ray-Tracing Channel Model · pp. 2–3, Table I · p. 3
  • The evaluation uses a simplified Event A3-style rule on filtered path power because simulated ray-traced power is not standardized RSRP.A3 is satisfied when the strongest alternative’s filtered-power advantage exceeds the configured margin throughout the TTT interval.Source: B. Ray-Tracing Channel Model · p. 3

C. Handover Problem Formulation

The handover problem is to choose between retaining the serving gNB and transferring to an eligible candidate using synchronized observations. Because target links may change during execution, the objective balances predicted future quality against handover and execution risks.

C. Handover Problem Formulation

  • An instantaneous strongest-link choice is insufficient because mobility, blockage, and propagation can change the target link during handover preparation and execution.The UE has access to the synchronized observation at each route index.Source: C. Handover Problem Formulation · p. 3
  • The action set includes staying on the serving gNB or handing over to a candidate gNB.The Event A3 net offset is a configured bias in the comparison.Source: C. Handover Problem Formulation · p. 3
  • The controller selects actions by maximizing predicted future quality while penalizing handovers and invalid or failed execution.The objective considers quality over the decision-to-application interval and estimates expected quality and failure probability from future channel predictions.Source: C. Handover Problem Formulation · p. 3

III. PROPOSED AGENTIC DIGITAL TWIN-ASSISTED PREDICTIVE HANDOVER FRAMEWORK

The proposed framework pairs a site-specific digital twin with an Agentic AI decision layer. The DT generates synchronized multimodal states, and the agents forecast, plan, validate, and adapt handover actions using execution feedback.

III. PROPOSED AGENTIC DIGITAL TWIN-ASSISTED PREDICTIVE HANDOVER FRAMEWORK

  • The framework comprises a site-specific DT and an Agentic AI decision layer.The DT generates synchronized multimodal states for the decision layer.Source: §III · p. 3
  • The Agentic AI layer forecasts, plans, and validates handover actions, then adapts decisions using execution feedback.These functions are described as autonomous.Source: §III · p. 3
  • The DT and decision layer form the proposed Agentic DT framework for predictive handover.Source: §III · p. 3

A. Site-Specific DT Representation and Data Generation

The site-specific DT uses a mesh-based map and ray tracing to generate synchronized radio and geometric observations along UE routes. The resulting records provide inputs and reference targets for forecasting future path power and link availability.

A. Site-Specific DT Representation and Data Generation

  • The mesh-based DT models terrain, buildings, materials, fixed gNBs, UE routes, antenna configurations, and propagation interactions.It can generate site-specific radio data and controlled scenarios under varied mobility, blockage, and propagation conditions.Source: A. Site-Specific DT Representation and… · p. 3
  • For each candidate gNB and route sample, ray tracing records valid path counts, delays, complex path coefficients, interaction information, and link validity.Path-power and validity features are derived from these records.Source: A. Site-Specific DT Representation and… · p. 3
  • The observation module synchronizes radio features with UE mobility and mesh-derived virtual-LiDAR and geometric features for forecasting future path power and link availability.Invalid links are retained with validity masks rather than numerical zeros; the study uses one site mesh.Source: A. Site-Specific DT Representation and… · pp. 3–4

B. Agentic AI Autonomous Handover

Three cooperating agents form one flexible decision system, with the forecasting agent selecting a forecaster for its usefulness to the handover objective. Its selected model supplies future link-quality and failure-risk estimates.

B. Agentic AI Autonomous Handover

1) Forecasting Agent:

  • The forecasting agent selects among persistence, deterministic GRU, independent probabilistic GRU, joint probabilistic GRU, and hurdle-aware joint GRU candidates using handover-objective usefulness, not prediction error alone.Source: 1) Forecasting Agent · p. 4
  • The policy-aware score is negative exponentially smoothed first-step masked path-power MAE; the selected forecaster supplies future link-quality and failure-risk estimates.The score is initialized from validation error and updated causally; additional forecasters can be added as resources and data increase.Source: 1) Forecasting Agent · p. 4

2) Planning Agent:

The planning agent evaluates state-dependent action sequences using forecasts, finite-horizon dynamic programming, and TTT persistence. The adaptation agent uses execution outcomes to update shared reliability and risk state for subsequent forecasts and plans.

2) Planning Agent:

  • Finite-horizon dynamic programming with TTT persistence evaluates state-dependent action sequences using forecasts of future link quality and execution risk.Source: 2) Planning Agent · p. 4
  • Only the first action in the maximizing sequence is executed; the sequence is recomputed at the next epoch with new observations and execution feedback.The feasible action set depends on preceding serving-cell transitions and the TTT-persistence state.Source: 2) Planning Agent · p. 4

3) Adaptation Agent:

  • Execution feedback updates shared state, including selected-gNB reliability, forecasting scores, failure state, and recent handover activity, shaping subsequent model selection and action planning.Successful actions reinforce reliability, whereas invalid actions reduce reliability and increase corresponding planning risk.Source: 3) Adaptation Agent · p. 4

IV. SIMULATION SETUP

The simulation represents the KFUPM campus with a site-specific radio-environment digital twin and evaluates six separate UE trajectories using synchronized records generated from ray tracing.

A. Site-Specific Deployment

  • The site-specific deployment uses the KFUPM campus terrain, building meshes, materials, antenna locations, and route geometry in its radio-environment digital twin.Source: A. Site-Specific Deployment · p. 4
  • The deployment description notes that 3DP compares successive future stay and handover choices and selects the sequence with the best cumulative predicted outcome.Source: A. Site-Specific Deployment · p. 4
  • Table II covers the five-gNB deployment on the site-specific map.Source: Table II · p. 5
  • Fig. 4's six slightly shifted trajectories are separate leave-one-trajectory-out folds, not simultaneous UEs following one path.Source: A. Site-Specific Deployment · p. 5
  • Ray tracing generated 1,440 synchronized records from six routes with 240 samples each, after an eight-point CUDA pilot verified the solver and blockage behavior.Table III summarizes route and radio configuration.Source: A. Site-Specific Deployment · p. 5, Table III · p. 5

B. Generated Dataset and Benchmarks

The section establishes a site-specific, synchronized dataset and benchmark setup, then evaluates handover methods along routes and across normalized end-to-end metrics. Agentic DT provides the most balanced overall performance, although fixed policies perform better on individual objectives.

B. Generated Dataset and Benchmarks

  • The site-specific pipeline generates synchronized RF, mobility, and mesh-derived LiDAR records; invalid RF links are masked rather than treated as zero-power observations.Stored RF power is uncalibrated dB, so absolute dBm thresholds await a calibrated rerun.Source: B. Generated Dataset and Benchmarks · p. 5, Figure 4 · p. 5, Table IV · p. 5
  • The six LOTO folds each hold out one trajectory; forecast-based benchmarks use the same DT observations, input windows, and six-fold LOTO protocol.Best usable-gNB counts sum to 1,431 rather than 1,440 because nine route samples have no usable gNB. Additional baselines are reactive Event A3 and standalone persistence.Source: B. Generated Dataset and Benchmarks · p. 5

C. Results and Analysis

  • The Agentic Twin reduces cumulative handovers versus Event A3, while fixed independent and joint policies produce even fewer.Handover count alone does not capture link validity or target-selection quality.Source: C. Results and Analysis · p. 5, Figure 5 · p. 6
  • Approximately 98.2% mean first-step ray-path availability is maintained across six folds, with small reductions near route transitions.This measures a valid ray-traced path to the selected gNB, not successful communication or completed handover.Source: C. Results and Analysis · p. 5, Figure 5 · p. 6
  • The Agentic Twin remains close to the best valid target over most route segments, with occasional oracle-gap peaks during rapid propagation changes.The usable oracle gap compares selected-target future power with the best valid target's future power.Source: C. Results and Analysis · pp. 5–6, Figure 5 · p. 6
  • The three-agent Agentic Twin provides the most balanced performance across handover activity, link validity, and target-selection alignment, despite fixed policies leading individual objectives.Fig. 6 normalizes the plotted methods across the three metrics.Source: Figure 6 · p. 6, C. Results and Analysis · p. 6

V. CONCLUSION

The conclusion presents a flexible, model-agnostic Agentic DT for predictive handover, combining a site-specific DT, multimodal forecasting, and three agents. It reports a favorable trade-off and identifies greater agent autonomy through LLM-based reasoning as future work.

V. CONCLUSION

  • The framework combines a site-specific DT, multimodal forecasting, and three agents for planning and execution-feedback adaptation.Source: §V · p. 6
  • The Agentic Twin provides a favorable trade-off among handover activity, link validity, and target-selection alignment.Source: §V · p. 6
  • Future work may incorporate LLM-based general-purpose reasoning for goal interpretation, forecasting, and planning strategy selection.Source: §V · p. 6
Loading 2610.33833v1…