Grey-Box Bayesian Optimization for Sensor Placement in Assisted Living Environments

Shadan Golestan, Omid Ardakanian, Pierre BoulangerAAAI 2024 · DOI

Introduction

In assisted living environments, sensors serve as critical infrastructure for fall detection and activity monitoring. However, determining the optimal placement for these sensors is a complex challenge. The search space is mathematically expansive; even in a small apartment, thousands of potential sensor combinations exist. Evaluating a single configuration often requires high-fidelity simulations or weeks of real-world data collection, making exhaustive searches impractical.

Standard Bayesian Optimization (BO) treats this as a "black-box" problem, focusing solely on final accuracy scores. However, this ignores domain knowledge: human activity is not uniformly distributed. Sensors placed in high-traffic areas like kitchens are inherently more valuable than those in rarely used corners.

Distribution-Guided Bayesian Optimization (DGBO)

This paper introduces Distribution-Guided Bayesian Optimization (DGBO), a "grey-box" approach that incorporates domain-specific logic about human activity patterns. Unlike black-box models, DGBO utilizes an internal "information profile" to guide the search.

The DGBO iterative optimization loop, illustrating the calculation of Expected Information Gain and the Information Profile.

The DGBO framework consists of several key components:

  1. SIMsis Simulator: Generates realistic human movement and sensor triggers based on daily routines.
  2. Spatial Credit Assignment: When a configuration performs well, the algorithm assigns "credit" to the specific activation regions covered by those sensors.
  3. Modified Acquisition Function: DGBO prioritizes query locations that overlap with high-activity zones, effectively learning the floor plan's importance during the optimization process.
  4. Surrogate Model: A Probabilistic Random Forest estimates the F1-score of potential placements.

Experimental Setup

The effectiveness of DGBO was validated across three distinct environments:

  • Testbed 1 (T1): A simulated 8x8m one-bedroom suite.
  • Testbed 2 (T2): A simulated 5.2x8m studio apartment with a constrained layout.
  • Aruba Dataset: A real-world dataset from the CASAS project featuring 219 days of data from a home with 31 sensors, providing a "ground truth" for validation.
Layouts of the two simulated testbeds (T1 and T2) used to evaluate the optimization algorithm.

Key Results and Insights

DGBO demonstrates a significant improvement in efficiency over standard black-box methods. By identifying low-activity areas early—often within the first 50 iterations—the algorithm avoids wasting expensive evaluations on unproductive regions.

Efficiency Gains: DGBO reached optimal performance levels with substantially fewer queries than standard BO:

  • 55.4% fewer in T1.
  • 58.9% fewer in T2.
  • 39.6% fewer in the real-world Aruba dataset.

Beyond speed, DGBO consistently found sensor configurations yielding higher macro-averaged F1-scores for activity recognition compared to Genetic Algorithms (GA) and Greedy search methods. The approach proved robust across different grid granularities, maintaining its advantage whether sensor spacing was set at 0.25m or 1m.

Conclusion and Future Work

DGBO effectively bridges the gap between pure black-box optimization and domain-specific heuristics. While the current model starts with zero knowledge for each new environment, future work aims to incorporate transfer learning. By applying knowledge of "kitchen-like" or "bedroom-like" behavior across different floor plans, the setup time for new assisted living installations could be further reduced. Additionally, extending the framework to handle diverse sensor types, such as mmWave radars or thermal cameras, remains a key area for development.

Summary updated

BibTeX
@inproceedings{golestan2024grey,
  author = {Shadan Golestan and Omid Ardakanian and Pierre Boulanger},
  title = {{Grey-Box Bayesian Optimization for Sensor Placement in Assisted Living Environments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i20.30208},
}
Figure