Data-Efficient Training of Gaussian Process Regression Models for Indoor Visible Light Positioning.

A data-efficient training method, namely Q-AL-GPR, is proposed for visible light positioning (VLP) systems with Gaussian process regression (GPR). The proposed method employs the methodology of active learning (AL) to progressively update the effective training dataset with data of low similarity to the existing one. A detailed explanation of the principle of the proposed methods is given. The experimental study is carried out in a three-dimensional GPR-VLP system. The results show the superiority of the proposed method over both the conventional training method based on random draw and a previously proposed line-based AL training method. The impacts of the parameter of active learning on the performance of the GPR-VLP are also presented via experimental investigation, which shows that (1) the proposed training method outperforms the conventional one regardless of the number of final effective training data (E), especially for a small/moderate effective training dataset, (2) a moderate step size () should be chosen for updating the effective training dataset to balance the positioning accuracy and computational complexity, and (3) due to the interplay of the reliability of the initialized GPR model and the flexibility in reshaping such a model via active learning, the number of initial effective training data () should be optimized. In terms of data efficiency in training, the required number of training data can be reduced by ~27.8% by Q-AL-GPR for a mean positioning accuracy of 3 cm when compared with GPR. The CDF analysis shows that with the proposed training method, the 97th percentile positioning error of GPR-VLP with 300 training data is reduced from 11.8 cm to 7.5 cm, which corresponds to a ~36.4% improvement in positioning accuracy.