Research map / Probabilistic and scientific modeling / Bayesian optimization
Grey-Box Bayesian Optimization for Sensor Placement in Assisted Living Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DGBO, a grey-box Bayesian optimization method that uses learned spatial activity distributions in the acquisition function to find sensor placements for activity recognition, evaluated in two simulated suites and the Aruba dataset.
Problem
Finding sample-efficient motion sensor placements in indoor assisted-living spaces that maximize downstream activity recognition F1, where each evaluation is expensive.
Approach
Treats sensor placement as a grey-box BO problem: each candidate sensor's activation region receives a share of observed performance (spatial credit assignment), yielding an information-gain estimate that is added to expected improvement in the acquisition function.
Key results
- DGBO attains macro F1 of 77.6 vs BO 76.9 in T1 at epsilon=0.25 with 9 sensors, with BO and DGBO both significantly outperforming GA and greedy.
- DGBO requires on average 55.4% fewer queries than BO in T1, 58.9% fewer in T2, and 39% fewer on Aruba to reach BO-level performance.
- On Aruba, DGBO reaches 76.3 F1 with 9 sensors versus 75.7 for BO and 60.2 for GA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGBO instead of BO for sensor placement | Activity recognition sensor placement · T1 | macro-averaged F1: 75.6±2.0 -> 77.5±0.1 | Helped |
| Use DGBO instead of BO for sensor placement | Activity recognition sensor placement · T2 | macro-averaged F1: 67.7±0.3 -> 68.7±2.6; 72.3±1.6 -> 72.3±1.8 | Mixed |
| Use DGBO instead of BO for sensor placement | Activity recognition sensor placement · Aruba | macro-averaged F1: 75.7±0.2 -> 76.3±0.2 | Helped |
| Use DGBO instead of BO | Sensor placement optimization · T1 | function queries to reach the target performance: -55.4% | Helped |
| Use DGBO instead of BO | Sensor placement optimization · T2 | function queries to reach the target performance: -58.9% | Helped |
| Use DGBO instead of BO | Sensor placement optimization · Aruba | function queries to reach the target performance: -39% | Helped |
Surprising findings
- The original 31-sensor manual placement in Aruba achieves only 67.8 F1, lower than optimized placements, showing the placement problem is non-trivial.
- The greedy algorithm often exhausts the 1000-query budget without finding a placement for small epsilon, because it needs many queries per added sensor.
Limitations
- Evaluated only with a single omnidirectional ceiling-mounted motion sensor type in simulated suites and one real home.
- Greedy and small-epsilon settings are query-expensive; method performance depends on the activation-region credit assignment heuristic, with no theoretical guarantees.
- Number of sensors must be fixed in advance for BO and DGBO, sweeping values from 5 to 15.
Experimental setup
Datasets: Aruba SIM_SIS simulated traces (T1, T2)
Benchmarks: Aruba T1 simulated testbed T2 simulated testbed
Metrics: macro-averaged F1 number of function queries to reach BO performance
Baselines: Bayesian optimization genetic algorithm greedy algorithm Deshwal et al. dictionary-based BO
Builds on: Bayesian optimization expected improvement probabilistic random forest SIM_SIS simulator
Closest papers by meaning
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.79)
- Optimizing Vital Sign Monitoring in Resource-Constrained Maternal Care: An RL-Based Restless Bandit Approach (AAAI 2025 · similarity 0.79)
- Evidential Active Recognition: Intelligent and Prudent Open-World Embodied Perception (CVPR 2024 · similarity 0.79)
- Scalable Bayesian Optimization via Focalized Sparse Gaussian Processes (NeurIPS 2024 · similarity 0.79)
- LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality (ICLR 2025 · similarity 0.79)
- Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions (NeurIPS 2026 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.