Physics-inspired Gaussian Processes Regression for RSS-based Visible Light Positioning

Visible light positioning (VLP) offers a cost-effective and accurate method for indoor localization. Gaussian processes (GPs), a data-driven method widely used in the received signal strength (RSS)-based VLP systems, face difficulties when training data are scarce or when forced to extrapolate. In this work, we propose a novel hybrid model, PhyGP, which integrates physics-based models into GPs through Bayesian active learning to enhance extrapolation capabilities without decreasing the interpolation accuracy of GPs. Experimental results, validated on real-world data, demonstrate significant improvements in extrapolation accuracy compared to GPs and in computational efficiency compared to physics-based models. Our approach achieves an improvement in extrapolation accuracy ranging from 32% to 83%, reducing the $\mathbf{P 5 0}$ error from 72 cm to 12 cm at its best performance. Additionally, the PhyGP model offers a four-order magnitude gain in computational efficiency compared with physics-based models.