PaperlayerView exact revision on arXiv

Source-linked AI summary

DSWM: Decomposed Spatio-Temporal World Model for Demand-Driven UAV Base Station Repositioning

Shengjie Zhong, Zhongliang Zhao, Jingxuan Chen, Xianbin Cao, Xinmei Qiang, Dapeng O. Wu, Tony Q. S. Quek

arXiv:2609.36845v1cs.NIcs.AIcs.LG

TL;DR

UAV-BS repositioning must cover demand that shifts over space and time, while existing approaches either re-solve optimization or learn reactive policies without an explicit demand model. DSWM plans imagined fleet motions in a learned latent space anchored on current observations, and ranks first among non-ablated configurations across three datasets.

  • Problem

    UAV-BS repositioning must cover spatially and temporally shifting demand, while existing optimization and reactive-learning approaches lack a unified way to plan under delayed effects and partial observation.

    Source: §I · p. 1

  • Method

    DSWM learns demand-service dynamics in a recurrent latent model, embeds differentiable service physics in imagined rollouts, and uses observation-anchored CEM planning with an uncertainty penalty.

    Source: §I · p. 2, C. Decomposed Demand-Service Physics… · p. 6, D. Observation-Anchored CEM… · p. 6

  • Results

    DSWM ranks first among non-ablated configurations across three datasets; on Milan weekdays it scores 0.889 ± 0.005, 0.109 above Greedy at 0.780.

    Source: D. Main Results · p. 9, §VI · p. 12

  • Takeaways & Limitations

    Across the evaluated datasets and demand regimes, the evidence supports decision-time use of observations as the central choice for DSWM's demand coverage.

    Source: §VI · p. 12

    YJMob100K measures distinct-user presence with 5% sampling rather than traffic bytes, and energy and fairness discriminate weakly, so claims rest on served ratio.

    Source: L. Limitations · p. 12

Abstract

from arXiv · show

Uncrewed aerial vehicle base stations (UAV-BSs) are expected to cover traffic demand that shifts across space and time, yet most repositioning schemes either re-solve an optimization problem per slot or learn reactive policies without an explicit demand model. We cast demand-driven fleet repositioning as latent-space decision-time planning and propose DSWM, a decomposed spatio-temporal world model: an agentic controller that perceives the demand field through a rolling observation window, retains operational context in a latent recurrent state, reasons about candidate motions by imagined rollouts under an uncertainty penalty, and coordinates the fleet through replanned first actions. DSWM learns a recurrent state-space model shaped by an exponential-moving-average (EMA) based latent predictive objective with variance regularization. It attaches a differentiable service simulator that replays the association, probabilistic line-of-sight channel, and Shannon rate chain inside latent rollouts. Planning uses a cross-entropy method whose imagined demand is anchored on the current observation window with mixing coefficient $\rho=0.95$. On a unified pipeline over three real datasets (Milan CDR (call detail record), Shanghai Telecom, YJMob100K) and 14 methods including five reproduced IEEE baselines, DSWM attains weekday served ratios of 0.889, 0.908, and 0.898, ranking first among non-ablated configurations on every dataset. On Milan it improves over the strongest non-learning baseline (Greedy, 0.780) by 0.109, a margin that comes from decision-time use of observations rather than prediction accuracy.

I. INTRODUCTION

Demand-driven UAV-BS repositioning must balance current and future coverage as demand shifts, but existing optimization, reactive-learning, and forecast-then-plan approaches each leave a gap. DSWM instead plans in a learned latent space anchored on current observations.

I. INTRODUCTION

  • The paper frames fleet control as balancing current and future service under shifting demand, delayed repositioning effects, partial observation, and battery constraints.Charging detours can remove capacity for an hour, while moves take minutes to pay off.Source: §I · p. 1

A. UAV-BS Deployment and Trajectory Design

  • Prior approaches either re-solve placement and trajectory optimization as demand drifts, learn reactive policies without an explicit demand model, or plan against forecasts misaligned with coverage.Optimization-based work includes BCD and convex-relaxation formulations,,,.Source: §I · pp. 1–2, A. UAV-BS Deployment and Trajectory… · p. 3

I. INTRODUCTION

  • DSWM combines a recurrent state-space model with a differentiable service-physics head and plans with observation-anchored CEM rollouts using ρ = 0.95.The service head replays association, probabilistic LoS, Shannon rate, and capped per-cell fulfillment; the planner executes only the first action.Source: §I · p. 2
  • DSWM reaches a Milan weekday served ratio of 0.889, 0.109 above Greedy at 0.780, and ranks first among non-ablated configurations on all three datasets.The evaluation uses a unified pipeline with 14 methods, including five reproduced IEEE baselines, and a leakage audit.Source: §I · p. 2

B. Spatio-Temporal Demand Prediction

The related work distinguishes forecast-based planning from decision-time use of observations, and positions DSWM against existing wireless world models and UAV-BS controllers. The system setup models city-wide fleet coverage with mobility and battery constraints.

B. Spatio-Temporal Demand Prediction

  • Forecasting methods produce demand estimates for downstream planners, but per-cell accuracy does not measure demand-weighted coverage.The paper places decision-time use of observations, rather than forecast skill, at the center of its approach.Source: B. Spatio-Temporal Demand Prediction · p. 3

C. World Models and Decision-Time Planning

D. Closest Works and Differentiation

  • Closest-work comparisons distinguish DSWM by placing service physics inside the world model, unlike model-free GA-MATR and simulation-focused MobiWorld [48].HRL-TPRA is reported as last among learning baselines in this setting at 0.485.Source: D. Closest Works and Differentiation · p. 3

A. Scenario and Assumptions

  • The modeled system uses 16 UAV-BSs at 100 m over a gridded city, with 10-minute slots and battery returns below a 15% reserve.A UAV returning to a quadrant station charges for six slots before redeployment.Source: A. Scenario and Assumptions · p. 3

B. Channel Model

The channel model uses UAV and cell geometry to estimate probabilistic LoS and path loss, then computes link rates with a Shannon-rate model.

B. Channel Model

  • The channel chain derives elevation angle from UAV position and cell geometry, models LoS probability, mixes LoS and NLoS path loss, and computes Shannon link rate.The channel chain is deterministic given positions and differentiable almost everywhere.Source: B. Channel Model · pp. 3–4
  • Each UAV-BS has a 20 MHz bandwidth budget, with link rates computed using its allocated bandwidth, transmit power, and noise spectral density.Table I is identified as the main-notation table.Source: B. Channel Model · p. 4, Table I · p. 4

C. Energy Consumption and Battery State Machine

The system models rotor energy use and a battery return-and-charge cycle, then evaluates service through demand-weighted fulfillment across cells.

C. Energy Consumption and Battery State Machine

  • Hovering dominates energy use because repositioning is limited to 1500 m per 10-minute slot, yielding an approximately 3% energy spread across methods.Hover energy is the reference, and the battery state machine governs returns and charging.Source: C. Energy Consumption and Battery… · p. 4
  • A UAV below 15% battery reserve returns to its quadrant station, charges for six slots, and redeploys at full charge.The stated battery capacity is 1,972,800 J (548 Wh).Source: C. Energy Consumption and Battery… · p. 4

D. Demand-Service Model

  • The served ratio aggregates fulfilled demand over the map and is demand-weighted, so serving a hotspot matters more than serving a sparse cell.A cell is marked satisfied at fulfillment ratio Upeak = 0.8 for the Jain fairness statistic.Source: D. Demand-Service Model · p. 4

E. Problem Formulation

The formulation seeks a policy that maximizes cumulative served ratio from a shared observation set. DSWM addresses demand drift, nonconvex service physics, and long-horizon credit assignment through latent dynamics, a differentiable service head, and observation-anchored planning.

E. Problem Formulation

  • The operator seeks a policy mapping observations to actions that maximizes cumulative served ratio over an episode.Source: E. Problem Formulation · p. 4
  • All methods receive six past and the current demand maps, fleet position, battery and charge states, and time-phase encoding.Source: E. Problem Formulation · p. 4
  • Short-window partial observation and time-varying demand make per-slot re-optimization chase a moving target; DSWM carries demand dynamics in latent state.Source: E. Problem Formulation · p. 4
  • The nonconvex channel, association, and battery constraints leave solver gradients unavailable; DSWM uses a decomposed differentiable service head.Source: E. Problem Formulation · p. 4
  • Repositioning pays off across slots through charging and demand drift, motivating CEM planning over 15 slots (2.5 h) anchored on current observations.Source: E. Problem Formulation · p. 4, Figure 2 · p. 5

A. Overview

  • At each slot, an encoder and RSSM update latent state, and CEM imagines candidate action sequences before executing the best sequence's first action.Source: A. Overview · p. 4, Figure 2 · p. 5
  • Offline training updates the model components end to end using sequences from a replay buffer of pre-collected transitions.Source: A. Overview · p. 5

B. RSSM World Model with Latent Predictive Regularization

DSWM uses an RSSM to retain latent context and combines latent prediction with variance regularization rather than reconstructing demand maps. This objective replaced an earlier reconstruction term and raised weekday served ratio from 0.878 in an internal scan.

C. Decomposed Demand-Service Physics Head

The decomposed head predicts demand and differentiably replays service physics to estimate served ratio. Ensemble disagreement supplies an uncertainty penalty that is described as insurance rather than a gain source.

D. Observation-Anchored CEM Decision-Time Planning

DSWM uses observation-anchored CEM to evaluate multi-slot consequences while repeatedly grounding imagined demand in the current observation. Offline training uses replayed sequences, and planning has fixed per-slot cost independent of solver convergence.

D. Observation-Anchored CEM Decision-Time Planning

F. Complexity Analysis

  • Planning costs O(NcemIH) imagined forwards per slot, independent of solver convergence; at K = 16, JTORATC solves exceeded 8 s per slot and were reduced to one solve per six slots.Source: F. Complexity Analysis · p. 7

E. Training Procedure

  • Offline training pre-collects 60,000 training-day transitions and runs 30,000 gradient steps over replayed sequences.Source: F. Complexity Analysis · p. 7

V. EXPERIMENTS

The experiments evaluate demand-driven repositioning on three real datasets with dataset-specific demand definitions, spatial scales, and temporal splits. Milan exhibits a commute bimodal pattern and more dispersed demand than YJMob100K.

A. Datasets and Preprocessing

  • Milan uses 20 × 20 cells at 235 m, with 10 training days and a 12-day test spanning Christmas–New Year; its top 5% of cells carry 13% of demand.The demand has a commute bimodal pattern.Source: A. Datasets and Preprocessing · p. 7, Figure 5 · p. 8
  • Shanghai Telecom contains 6.15M cleaned records, uses 470 m cells, and tests on November after training from June through October.Internet-session demand is genuinely sparse, with an 85.3% zero-value rate.Source: A. Datasets and Preprocessing · p. 7
  • Milan demand is more dispersed than YJMob100K: its top 5% of cells carry 13% of demand, compared with 22% for YJMob100K.The figure marks the temporal train/test split and Christmas–New Year OOD week alongside daily demand cycles and total drift.Source: Figure 5 · p. 8
  • YJMob100K demand counts distinct user IDs per cell and slot, an approximately 5% sampled mobility proxy rather than traffic bytes.Its native 30-minute slots are divided into 10-minute slots, and its 19-day test follows 56 training days.Source: A. Datasets and Preprocessing · p. 8

B. Setup and Evaluation Protocol

The protocol gives learning methods a shared observation and action space and evaluates them on fixed temporal test days. On Milan, DSWM leads non-ablated configurations in served ratio, while energy and fairness distinguish methods weakly.

B. Setup and Evaluation Protocol

  • All learning methods share the past six demand maps, current demand, fleet position, battery and charge states, and time-phase encoding; actions contain two movement coordinates and a bandwidth logit per UAV.The action components are bounded in [−1, 1].Source: B. Setup and Evaluation Protocol · p. 8
  • Evaluation uses temporal splits, fixed weekday ID and holiday OOD days, five Milan seeds, and three seeds per other dataset.Random train/test splits are forbidden; per-day variability is shown for the test period.Source: B. Setup and Evaluation Protocol · p. 8

D. Main Results

  • 0.889 ± 0.005 on weekdays and 0.903 on holidays put DSWM first among non-ablated configurations, 0.109 above Greedy (0.780).Served ratio is the primary discriminating metric; energy spans 1210–1247 kJ, about a 3% spread.Source: Table III · p. 9, D. Main Results · p. 9

C. Baseline Lineup

  • DSWM and two internal ablations, w/o unc and w/o dec, complete the method lineup described for the experiments.Source: C. Baseline Lineup · p. 9

D. Main Results

  • Greedy, GA-MATR, and TD-MPC score 0.769–0.780 on weekdays, while INS-WOA (0.744) falls between Static (0.736) and Greedy.The passage associates INS-WOA’s placement with its Voronoi pre-deployment inheriting K-means’ sensitivity to demand shape.Source: D. Main Results · p. 9
  • EMORL-TCTO (0.614) and HRL-TPRA (0.485) trail the learning-method results reported for Milan.The paper points to a later diagnostic chain for these negative results.Source: D. Main Results · p. 9

E. OOD Analysis

The experiments examine holiday distribution shift, structural ablations, sample efficiency and fleet scaling, and world-model quality. DSWM remains strong under holiday shift, while the service head accounts for the larger ablation effect.

E. OOD Analysis

  • DSWM scores 0.903 on holidays versus 0.889 on weekdays, while JTORATC falls to 0.495 from 0.583 under the Christmas–New Year shift.GA-MATR and TD-MPC are the only learning baselines whose holiday scores exceed Greedy’s 0.792.Source: E. OOD Analysis · p. 9

F. Ablation

  • Removing the decomposed service head drops weekday served ratio from 0.889 to 0.769 (−0.120); removing uncertainty instead raises it by 0.004 to 0.894.Under demand scaling s ∈ {0.5, 2}, DSWM stays near its s=1 reference and degrades less than w/o dec.Source: F. Ablation · pp. 9–10, Figure 8 · p. 10

G. Sample Efficiency and Scalability

H. World Model Quality

  • Demand RMSE is 0.128 Mbps at one imagination step and 0.119 Mbps at 24 slots, while imagined-return rankings correlate with realized returns at rs = 0.520 ± 0.087.Each return-correlation point is one seed–evaluation-day pair.Source: H. World Model Quality · p. 11, Figure 12 · p. 11

I. Cross-Dataset Generalization

DSWM ranks first among non-ablated configurations across the three datasets without per-dataset tuning. The discussion links coverage results to demand dispersion and spatial redistribution, while qualifying what the served ratios measure.

J. Discussion: Lessons Learned

  • Cross-agent checks found GA-MATR’s first re-implementation collapsed (cross-agent std = 0), recovering 0.770 after variance and symmetry-breaking fixes; HRL-TPRA ranked last among learning baselines at 0.485 despite a 5.7× speedup.Identical code rerun on a different day deviated by ∼10−4 due to BLAS/CPU variation.Source: J. Discussion: Lessons Learned · p. 11
  • YJMob preprocessing uniformly splits native 30-min counts into 10-min slots and crops the densest 9×9 cells to align with Milan’s 4.5-km scale and slot physics.The passage says no information is injected by the uniform de-aggregation.Source: J. Discussion: Lessons Learned · p. 11

I. Cross-Dataset Generalization

  • 0.908 on Shanghai (+0.033 over TD-MPC), 0.898 on YJMob100K (+0.019 over PPO), and 0.889 on Milan (+0.109 over Greedy); DSWM ranks first among non-ablated configurations with no per-dataset tuning.The comparison covers three datasets under one production configuration.Source: §I · p. 11, Figure 13 · p. 12

J. Discussion: Lessons Learned

  • Static (K-means) falls to 0.437 on 85%-sparse Shanghai; GA-MATR shows robustness under shift, while INS-WOA inherits K-means demand-shape sensitivity.The discussion reports these as lessons from the benchmark.Source: J. Discussion: Lessons Learned · p. 12

K. Qualitative Visualization

  • The visualizations show weekday–weekend OOD shifts as spatial redistribution rather than uniform intensity change, while Milan prediction errors cluster at cell boundaries during peak transitions.Milan predictions track commute-driven hotspot migration across the city grid.Source: K. Qualitative Visualization · p. 12

L. Limitations

  • YJMob100K served ratios measure distinct-user presence at 5% sampling, not traffic bytes; energy and fairness discriminate weakly, so claims rest on served ratio.The reported energy spread is ∼3%, and Jain values range from 0.888 to 1.000.Source: L. Limitations · p. 12

VI. CONCLUSION

  • The conclusion identifies decision-time use of current observations—not forecast accuracy—as the binding coverage constraint, and says the service head trains on operator-valued coverage.Observation-anchored CEM grounds imagined rollouts in the current observation rather than an open-loop forecast.Source: §VI · p. 12
  • Future work targets byte-accurate demand semantics on mobility datasets, heterogeneous fleets, and energy-aware altitude control.Source: §VI · p. 12

Related papers

Loading 2609.36845v1…