A New Distance Metric Based on Bias-Variance Tradeoff for Wireless Fingerprinting

— This article proposes a new distance metric for \nindoor positioning based on wireless fingerprinting, referred \nto as the weighted bias-variance (WBV) metric. The WBV \nmetric is parameterized by a single parameter that is adjusted \nbased on the concept of bias-variance tradeoff, making it highly \neffective in scenarios where there is a network reconfiguration \nor change in the environmental layout. The optimal setting \nwas found to be minimally affected by the number of nearest \nneighbors in the estimation algorithm as well as the access point \n(AP) selection schemes. In an experiment at a university library, \nthe switch from the Manhattan metric to the WBV metric after \na network reconfiguration achieved a reduction in the mean \nposition error (MPE), standard deviation and median of the \nerror distribution of 34.9%, 22.1%, and 41.5%, respectively, \nwith negligible changes to the floor accuracy (FA). When the \nnumber of APs was reduced, the corresponding reduction was \n42.7%, 35.5%, and 46.3%, respectively. In another experiment \nat a university faculty building, the switch from the Manhattan \nmetric to the WBV metric after a layout change attained a \nreduction in the MPE, standard deviation, and median of the \nerror distribution of 55.1%, 62.6%, and 23.1%, respectively. The \nFA improved by 22.2%. The drastic improvement in accuracy \nhas the potential to significantly advance indoor positioning for \nthe ubiquitous Internet of Things