Vibration-Based Indoor Human Sensing Quality Reinforcement via Thompson Sampling

This paper presents an online learning algorithm that recommends sensor locations for structural vibration-based indoor human sensing systems to achieve optimal sensing quality for high learning accuracy. The intuition is that we model the deployment environment features (physics models) to reflect the environmental impact on the signals, which would further impact the sensing applications' performance. We acquire these features by re-purposing the ambient sensing data used for sensing applications. Then we combine these environment features and the application performance acquired as user feedback (data-driven knowledge) to recommend the sensor locations with optimal sensing quality. We formulate the task as a multi-armed bandit problem and develop an online learning algorithm based on Thompson sampling. Real-world datasets are collected to validate the proposed algorithm for both online and offline learning scenarios. The system achieves 56% for R@1 and 90% for R@3, which demonstrate an up to 2.5× improvement compared to baseline approaches.