Text

theses

Recent advances in mobile and pervasive computing have enabled accurate location tracking of users wearing wireless devices indoors, where GPS isn’t available. Many indoor WiFi location estimation techniques use received radio signal strength (RSS) values from various access points to track users. In recent years, machine learning techniques have been applied to this application and demonstrated the effectiveness for tracking mobile devices. However, many existing systems suffer from the following problems: (1) lack of labeled data, and (2) non-stationary data distribution. In this thesis, we will describe our kernel learning-based method for solving the these challenges. First, since it can be expensive to collect and label RSS training data in large and complex buildings, we propose a semi-supervised learning approach to learn labelaware base kernels, which are shown to be better aligned to the target comparing to traditional base kernels spanned by the eigenvectors of the kernel matrix (or the graph Laplacian); second, since the data distribution changes constantly as devices change and over different time periods, we propose a transfer learning approach in Reproducing Kernel Hilbert Space(RKHS) to adapt the data distribution changes. Experimental results on real-world benchmark data demonstrate the encouraging performance of our proposed schemes.

ETD_6952

Ph.D.

Includes bibliographical references

by Qiaojun Wang

doi:10.7282/T3GQ70VJ

ETD doctoral

The author owns the copyright to this work.