Resource-Aware Model Selection for Scalable Indoor Localization on HPC Platforms

Resource-Aware Model Selection for Scalable Indoor Localization on HPC Platforms (arXiv:2609.29402v1, published September 24, 2026) introduces a modular framework for WiFi fingerprinting that treats localization as a resource-aware model selection problem rather than a monolithic classification task. The system organizes the environment into a building-floor-spot hierarchy using independent autoencoder models, allowing for scalable, incremental updates without retraining global models.

To overcome the computational costs of evaluating hundreds of local models, the framework employs two execution-pruning strategies: Hierarchical Candidate Pruning (HCP): A coarse-to-fine selection process that reduces model evaluations from 735 to 67 per query while maintaining competitive accuracy. Trajectory-Aware Pruning (TAP): Leverages temporal locality in user movement to restrict inference to spatially plausible neighbors, reducing evaluations to just 10 models per query (a 98.6% reduction in computation).

Validated on the UJIIndoorLoc dataset and scaled on the Fugaku supercomputer, the approach achieves ultra-low inference latency (down to 1.07 ms across 16 CPU nodes) and minimizes memory residency overhead, making it highly suitable for distributed, high-performance indoor localization services.

This material addresses the challenge of deploying indoor localization systems at large, evolving scale—across campuses, smart buildings, factories, and digital-twin environments—where wireless conditions, access-point placements, and spatial layouts are not static. Rather than treating localization as a fixed algorithmic problem, it frames the task as a systems and model-selection problem: as new buildings, floors, rooms, and service areas are added, the localization stack must remain accurate, extendable, and maintainable. The HPC focus suggests that the work is concerned not only with choosing a strong localization model, but with doing so under realistic resource constraints and at a scale that would be impractical on a single workstation or edge node.

Its central insight is that “best” indoor localization should be defined jointly by accuracy and operational cost. In large deployments, the relevant tradeoffs include inference latency, compute and memory usage, calibration effort, resilience to environmental drift, and the ability to onboard new spaces without redesigning the whole system. A resource-aware model-selection approach is therefore valuable because it allows operators to match models, feature pipelines, or processing strategies to the scale and constraints of the deployment, rather than assuming that the most expressive model is always the right choice. In an HPC setting, this can enable more systematic evaluation, parallel processing, and scalable operation across many areas or candidate models.

This matters because indoor positioning is increasingly foundational for location-aware services, building automation, asset tracking, navigation, and digital-twin synchronization, yet it is inherently difficult due to multipath propagation, non-line-of-sight effects, heterogeneous sensor coverage, and frequent physical changes. A practical deployment framework that balances localization quality with resource efficiency and long-term maintainability is therefore more likely to succeed in real infrastructure than a purely accuracy-focused model. The work is relevant to researchers and practitioners working at the intersection of localization, machine learning, and high-performance systems, especially those building platforms that must grow and adapt over time.