Publications
Detailed Information
Indoor localization with application of semi-supervised learning and ensemble approach
Cited 0 time in
Web of Science
Cited 0 time in Scopus
- Authors
- Advisor
- 김현전
- Major
- 공과대학 기계항공공학부
- Issue Date
- 2018-02
- Publisher
- 서울대학교 대학원
- Keywords
- WiFi based indoor localization ; Semi-supervised learning ; Ensemble approach ; K-nearest neighbour ; Gaussian process regression
- Description
- 학위논문 (석사)-- 서울대학교 대학원 : 공과대학 기계항공공학부, 2018. 2. 김현전.
- Abstract
- For indoor localization, WiFi based localization techniques have been widely researched because WiFi infrastructures are pre-deployed in many indoor environments and it is easy to access by smart phone.
We researched the ensemble approach of GPR and KNN for improved localization results. Because each of Gaussian process regression (GPR) and K-nearest neighbour (KNN) has different properties, suitable combination of two techniques is expected to improve the localization accuracy by synergy of them. Firstly, before ensemble approach KNN and GP are handled individually to improve localization accuracy of each for better improvement at ensemble stage. To improve the localization accuracy of KNN, we proposed the feature addition process that strengthens the characteristics before dimension reduction and introduced clustering method to semi-supervised learning that makes the learning performance better by providing more effective training label from clustering ID.
For GPR, various GP models are compared and the best one is selected based on localization error. With improved GP and KNN, ensemble approach using adaptive weights using database from cross-validation is developed. Compared with conventional ensemble approach using the weights only determined from relative accuracies of localization techniques, proposed ensemble approach shows a more improved localization results.
- Language
- English
- Files in This Item:
Item View & Download Count
Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.