A genetic programming approach to WiFi fingerprint meta-distance learning

Created by W.Langdon from gp-bibliography.bib Revision:1.7615

@Article{BRUNELLO:2022:pmcj,

• author = "Andrea Brunello and Angelo Montanari and Nicola Saccomanno",
• title = "A genetic programming approach to {WiFi} fingerprint meta-distance learning",
• journal = "Pervasive and Mobile Computing",
• year = "2022",
• volume = "85",
• pages = "101681",
• keywords = "genetic algorithms, genetic programming, Indoor positioning, Wi-Fi fingerprinting, Metric, Machine learning",
• ISSN = "1574-1192",
• DOI = "doi:10.1016/j.pmcj.2022.101681",
• URL = "https://www.sciencedirect.com/science/article/pii/S1574119222000980",
• abstract = "Driven by the continuous growth in the number of mobile smart devices, location-based services are becoming a fundamental aspect in the ubiquitous computing domain. In this work, we focus on indoor scenarios, where positioning supports tasks such as navigation, logistics, and access management and control. Most indoor positioning solutions are based on WiFi fingerprinting, thanks to its ease of deployment. Such a technique often requires the adoption of a suitable distance metric to compare the currently observed WiFi access points with those pertaining to fingerprints contained in a database, and whose position is already known. Results from the literature make it evident that classical distance functions among WiFi fingerprints do not preserve spatial information in its entirety. Here, we explore the possibility of addressing such a shortcoming by combining a selection of fingerprint distance functions into a meta-distance, using a genetic programming approach to solve a symbolic regression problem. The outcomes of the investigation, based on 16 publicly available datasets, show that a small, but statistically relevant, improvement can be achieved in preserving spatial information, and that the developed meta-distance has a generalization capability no worse than top-performing classical fingerprint distance functions when trained on a dataset and tested on the others. In addition, when used within a k-nearest-neighbor positioning framework, the meta-distance outperforms all the contenders, despite not being expressly designed to support position estimation. This sheds a light on a significant relationship between preservation of spatial information and localization performance. The achieved results pave the way for the development of more advanced metric learning solutions, that go beyond the limitations of currently-employed fingerprint distance functions",

}

Genetic Programming entries for Andrea Brunello Angelo Montanari Nicola Saccomanno

Citations