Attending to Routers Aids Indoor Wireless Localization

Modern machine learning-based wireless localization using Wi-Fi signals continues to face significant challenges in achieving groundbreaking performance across diverse environments. A major limitation is that most existing algorithms do not appropriately weight the information from different routers during aggregation, resulting in suboptimal convergence and reduced accuracy. Motivated by traditional weighted triangulation methods, this paper introduces the concept of attention to routers, ensuring that each router’s contribution is weighted differently when aggregating information from multiple routers for triangulation. We demonstrate, by incorporating attention layers into a standard machine learning localization architecture, that emphasizing the relevance of each router can substantially improve overall performance. We have also shown through evaluation over the open-sourced datasets and demonstrate that Attention to Routers outperforms the benchmark architecture by over 30% in accuracy.

💡 Research Summary

The paper tackles a persistent problem in Wi‑Fi‑based indoor localization: most deep learning models treat all access points (APs) equally, even though the quality of the received signals can vary dramatically across routers due to distance, obstruction, multipath, or hardware differences. This uniform treatment leads to slower convergence and sub‑optimal positioning accuracy, especially in challenging non‑line‑of‑sight (NLoS) or heavily multipath‑rich environments. Inspired by classic weighted triangulation, the authors propose “Attention to Routers,” a lightweight channel‑wise attention mechanism that learns a per‑router importance weight during training and explicitly emphasizes the most informative APs while attenuating noisy ones.

Model Architecture
The baseline follows the DLoc/RLoc paradigm: raw CSI is processed into Angle‑of‑Arrival (AoA) and Time‑of‑Flight (ToF) heatmaps for each AP, which are stacked into a tensor H. An encoder E compresses H into a latent representation (\hat H); a decoder D maps (\hat H) to a spatial likelihood map Y of the client’s position. The authors insert an attention module between E and D. For each router r, the encoder produces a d‑dimensional embedding (\hat h_r). A simple average‑pool over the d dimensions yields a scalar summary (s_r). A two‑layer MLP with ReLU (parameters (W_1, b_1, W_2, b_2)) transforms (s_r) into an unnormalized score (u_r). A softmax across all routers yields normalized attention weights (\alpha_r) that sum to one. The latent embedding is then recalibrated as (\tilde h_r = \alpha_r \cdot \hat h_r) before being fed to the decoder. The attention function is set‑invariant (i.e., permutation‑invariant) so the ordering of routers does not affect the computed weights. Optionally, a global context vector (\bar h) can be concatenated to each (\hat h_r) to enable relative comparisons.

Loss Functions
The training objective combines two terms:

1. Location loss (L_{Loc}) – an L2 loss between the triangulated position (using the decoder’s AoA predictions together with known router coordinates) and the ground‑truth Gaussian heatmap.
2. AoA loss (L_{AoA}) – an L1 loss on the predicted AoA values for each router, weighted by a hyper‑parameter (\lambda).
   The total loss is (L = L_{Loc} + \lambda L_{AoA}). This encourages both accurate final positions and reliable intermediate AoA estimates.

Experimental Evaluation
The authors evaluate on the publicly available DLoc dataset (≈4 k samples). They compare the vanilla encoder‑decoder model (no attention) with the attention‑augmented version. Results show substantial improvements across all statistical measures:

The most striking gains appear for “hard” samples (locations with high baseline error), where the attention model reduces error by roughly 45 %. Medium‑difficulty cases improve by about 26 %, while easy cases show a modest increase in error (interpreted as a deliberate reallocation of model capacity toward harder regions).

Interpretability
Attention weights are visualized per router. AP 3 consistently receives the highest median and mean weight, indicating it provides the most reliable spatial cues in the test environment. AP 1 and AP 2 receive lower, tighter weight distributions, reflecting consistent but less critical contributions. The overall entropy of the weight vector corresponds to ~92 % uniformity, confirming that the model still leverages all routers but preferentially focuses on the most informative ones.

Discussion and Limitations
The paper argues that router‑wise attention brings two key benefits: (1) faster and more stable convergence because the network no longer needs to infer implicit importance through many parameters, and (2) improved robustness to NLoS and multipath because noisy routers are down‑weighted. The learned weights also provide actionable insights for network planning (e.g., identifying which APs to relocate or upgrade).

Limitations include: evaluation on a single dataset and indoor layout, which may not capture the full variability of real‑world deployments; potential scalability issues when the number of APs grows very large, as softmax weights could become overly sparse. The authors suggest future work on sparse or hierarchical attention mechanisms, integration with other modalities (BLE, UWB, vision), and dynamic attention that adapts to real‑time AP health metrics.

Conclusion
“Attention to Routers” demonstrates that a modest, permutation‑invariant attention layer inserted between encoder and decoder can dramatically improve Wi‑Fi‑based indoor localization. By explicitly learning per‑router importance, the model achieves up to 40 % reduction in extreme errors and a 30 % overall accuracy boost, while also offering interpretable diagnostics for network operators. The approach is lightweight, compatible with existing encoder‑decoder pipelines, and opens avenues for more resilient, multi‑modal indoor positioning systems.