Source-linked AI summary
Rethinking Radiomap Blind Prediction with Limited Environment and Configuration Representations
Xiaojie Li, Yu Han, Han Fang, Shangqing Liu, Shi Jin, Chao-Kai Wen
TL;DR
Radiomap blind prediction must infer radiomaps from incomplete environment and BS representations, which cannot uniquely determine the target. The paper develops RadioDecomp, using a prior-guided base with deterministic residual refinement, and finds that RadioLSR is especially effective across configurations while improving overall cross-environment performance over a controlled monolithic model.
Problem
Radiomap blind prediction lacks complete environment and BS information, so observable representations cannot uniquely determine the target radiomap.
Method
RadioDecomp uses a prior-guided base predictor and deterministic residual refinement to learn its remaining predictable discrepancy, instantiated as RadioLSR.
Results
RadioLSR is especially effective for cross-configuration generalization and provides overall benefits under cross-environment generalization over a controlled monolithic reference.
Takeaways & Limitations
Structured prior-guided prediction with residual refinement improves generalization under incomplete observation, particularly across BS configurations.
Abstract
from arXiv · showhide
Radiomap blind prediction infers radiomaps from observable representations of the propagation environment and base station (BS) configuration without field measurements. These representations are inherently incomplete and cannot uniquely determine the target radiomap. Under squared loss, we identify the conditional-mean radiomap as the population-optimal deterministic target and decompose domain risk into target-approximation error and irreducible uncertainty. The train-test risk gap motivates propagation priors as cross-domain guidance, although their partial or simplified forms may bias the attainable predictor. We therefore propose RadioDecomp, which treats a prior-guided predictor as a correctable base and uses deterministic residual refinement to learn its remaining predictable discrepancy. We instantiate RadioDecomp as RadioLSR (LoS-Shadow-Residual). Experiments under cross-configuration and cross-environment settings show that RadioLSR is especially effective for cross-configuration generalization and provides overall gains over a controlled monolithic counterpart under cross-environment generalization.
I. INTRODUCTION
Radiomap blind prediction infers spatial radio attributes from incomplete environment and BS representations without field measurements. The paper characterizes the resulting uncertainty and proposes RadioDecomp, instantiated as RadioLSR, to refine prior-guided predictions.
- Motivation: Radiomap blind prediction estimates spatial radio attributes from observable environment and BS representations without field measurements.These radiomaps support wireless network planning, optimization, and predictive radio management.
- Motivation: Existing approaches learn direct mappings or use fully specified ray-tracing, but practical blind prediction provides only partial environment and configuration representations.Measurement-supported radiomap estimation is related but does not directly characterize blind prediction under incomplete observation.
- Proposed Approach: RadioDecomp treats a prior-guided predictor as a base and uses deterministic residual refinement to correct its predictable discrepancy, yielding RadioLSR.RadioLSR is designed for incomplete observation and combines the proposed decomposition with LoS and shadowing structure.
- Problem Formulation: Under squared loss, the population-optimal deterministic predictor is the conditional-mean radiomap rather than a radiomap for one hidden physical realization.Multiple hidden states can match one observable input while producing different radiomaps.
B. Domain-Wise Risk Decomposition and Train-Test Gap
The paper decomposes domain risk into target-approximation error and irreducible uncertainty, clarifying which train-test effects predictor design can address. Propagation priors may improve cross-domain guidance but can introduce structural approximation bias, motivating correctable prior-guided predictors.
- Risk Decomposition: Domain risk separates into predictor-dependent target-approximation error and predictor-independent uncertainty induced by incomplete observation.The decomposition applies separately to training and test domains for deterministic predictors with finite second moment.
- Train-Test Gap: The train-test risk gap equals the domain-wise change in approximation error plus the domain-wise change in conditional uncertainty.Good generalization requires approximation error not to increase substantially across domains, while the uncertainty gap cannot be reduced through predictor design.
- Propagation Priors: Propagation knowledge may help preserve transferable relations and maintain a small approximation gap across domains.The physical propagation mechanism provides guidance because its underlying regularities are shared across environments and BS configurations.
- Propagation Priors: Under incomplete observation, implementable propagation priors describe only selected or simplified aspects of the physical mechanism and therefore induce a restricted hypothesis class.This restriction is represented by the prior-guided class HΦ and its domain-specific structural approximation bias.
- Propagation Priors: When BΦ,d > 0, even the best predictor allowed by the prior-guided base class cannot attain the conditional target.This benefit-bias tradeoff motivates a correctable prior-guided predictor.
C. RadioDecomp as a Correctable Prior-Guided Predictor
RadioDecomp uses a propagation-prior-guided base predictor while adding deterministic residual refinement to correct its predictable discrepancy.
- The transferable interaction component (TIC) captures cross-domain propagation relations but may deviate from the conditional target because of bias, finite-sample learning, or imperfect optimization.
- The residual interaction component (RIC) uses the base prediction as an anchor while exploiting observable relations that TIC does not capture.
- If the residual class contains the zero function, residual refinement can recover the base predictor and therefore cannot worsen the best achievable risk.
- Under squared loss, the pointwise optimal deterministic residual is the conditional mean discrepancy between the target and the fixed base predictor.
- RadioDecomp retains the prior-guided base as a special case while providing a deterministic correction path toward the domain-specific conditional target.The residual branch can correct the predictable approximation bias of the base.
III. A RADIODECOMP INSTANTIATION: RADIOLSR
RadioLSR instantiates RadioDecomp as a LoS-Shadow-Residual model that uses observable-supported structure as a base and refines the remaining discrepancy.
- RadioLSR uses LoS-dominant and blockage-dominant shadowing effects as an observable-supported base estimate rather than a complete radiomap decomposition.The remaining discrepancy is refined through a residual part.
- Both RadioLSR predictors operate at the prediction level with corresponding spatial masks incorporated into their outputs.
A. Physics-Aware Input Representation
RadioLSR represents each sample with beam, height, directional-edge, and blockage information, while restricting prediction to valid non-building regions.
- The observable input contains a beam map, height map, two directional edge maps, and a blockage score map on the H × W grid.
- Prediction is restricted to valid non-building cells using a binary valid-region mask.
- The blockage score samples points along each BS-grid line segment and compares ray height with building height to quantify obstruction severity.
- Binary LoS and shadow masks partition the valid region according to a blockage-score threshold.
B. RadioLSR Architecture
RadioLSR uses separate LoS, Shadow, and residual U-Net branches, restricting the base stage to stable structured cues and deferring harder effects to refinement.
- RadioLSR consists of three U-Net branches for LoS prediction, shadow prediction, and residual refinement.The LoS and Shadow branches form the structured base, while the residual branch performs refinement.
- The base branches use restricted inputs to focus on stable, explicitly supported structure, while harder effects are deferred to residual refinement.Reflection-related patterns are deferred because they are more sensitive to fine geometry and surface orientation.
- Base Branch (LoS and Shadow): The LoS branch captures the direct coverage trend induced by the BS configuration using the beam map as its primary support.
- Base Branch (LoS and Shadow): The Shadow branch models blockage-aware attenuation using obstruction cues alongside the beam map.The blockage score supplies obstruction severity, while the beam map preserves configuration-dependent coverage tendencies.
- Base Branch (LoS and Shadow): The structured base predictor is formed by masked aggregation of the LoS and Shadow branch predictions.
2) Residual Refinement:
The residual stage adds a U-Net branch that refines the base prediction's remaining discrepancy in the dB domain. It conditions refinement on the base output and assigns blockage-aware effects to the base stage.
- The residual stage adds one U-Net branch to refine the remaining discrepancy in the dB domain.
- The residual predictor combines observable inputs and the base prediction, then applies the valid-region mask.
- Conditioning on the base prediction makes refinement relative to the current attenuation level rather than an independent full-map prediction.
- The unified U-Net backbone supports the RadioLSR branches used for structured prediction.
C. Training Objective
RadioLSR is trained with a masked final-prediction loss plus auxiliary supervision that encourages specialization of its LoS, shadow, and residual branches.
- RadioLSR uses a masked ℓ1 loss on the final prediction together with auxiliary supervision for the LoS, shadow, and residual branches.
- The residual target is explicitly defined for training the refinement branch.
- The main loss supervises the composed prediction, while auxiliary terms encourage LoS, shadow, and residual specialization.
IV. EXPERIMENTS
Experiments compare RadioLSR with a controlled monolithic counterpart across cross-configuration and cross-environment protocols, multiple train ratios, and matched training settings. The evaluation includes direct comparisons, aggregate error statistics, and an ablation of residual refinement.
- A. Experimental Setup: The study uses 78,400 radiomaps from 800 urban scenes and 98 BS configurations across cross-config and cross-env protocols.Six train/validation/test ratios are evaluated, alongside a monolithic MonoUNet reference using the same input information.
- A. Experimental Setup: The experiments compare RadioLSR-32 with MonoUNet-32 using train MAE and test MAE/RMSE under both split protocols.
- A. Experimental Setup: Aggregated test-sample RMSE is reported with P50, P90, and P95 over all train-ratio settings.
- A. Experimental Setup: The evaluation includes a 6/1/3 ablation comparing full RadioLSR with its TIC-only variant under cross-config and cross-env settings.
- A. Experimental Setup: MonoUNet baselines use widths 32, 48, and 64 and are trained with masked ℓ1 radiomap loss plus an additional Sobel-based structural loss.
- A. Experimental Setup: All models use the stated blockage threshold and loss weights, train for 60 epochs, and select the best validation checkpoint for evaluation.
1) RadioDecomp-Guided Main Comparison:
RadioLSR generally outperforms its matched monolithic counterpart, especially under cross-configuration generalization, while residual refinement improves performance under both split protocols. Qualitative results show a base prediction capturing broad coverage and residual corrections addressing difficult localized regions.
- 1) RadioDecomp-Guided Main Comparison: RadioLSR-32 consistently achieves lower test MAE and RMSE under cross-config and overall better performance under cross-env than MonoUNet-32.It wins on a clear majority of samples in every cross-config setting and exceeds 50% sample-level win rate in every cross-env setting.
- 2) Capacity and Reparameterization Analysis: RadioLSR-32 or RadioLSR-48 achieves the best overall results rather than wider MonoUNet baselines, with lower mean RMSE under both splits and advantages in P90 and P95.The comparison uses identical observable inputs and a large number of test radiomaps.
- 3) Ablation on the Role of the RIC Refinement: Removing RIC increases MAE/RMSE by 1.4221/2.9858 dB under cross-config and 0.9889/2.4713 dB under cross-env at the 6/1/3 split.The ablation indicates that structured base estimation alone is insufficient within the RadioDecomp formulation.
- 4) Qualitative Analysis of the Structured Prediction Path: Qualitative examples show the LoS-Shadow base capturing dominant large-scale coverage while the residual branch adds localized corrections in difficult regions.The reported corrections include reflection-related variations supported by available geometric cues.
- V. CONCLUSION: The paper concludes that RadioLSR is especially effective for cross-config generalization and provides overall benefits under cross-env generalization over a controlled monolithic reference.