Development of an Adaptive Clustering Algorithm for Localization of Wireless Signals

Abstract

The algorithm selection problem is an idea John Rice proposed in the 1970’s in which he suggests that selecting the proper algorithm to be used to solve a problem is a problem itself [1]. Since then, there have been several ways proposed to address how to best select an algorithm for a given dataset [4], and more specifically for the use case of multiple target localization [3], which clustering algorithm will give the best results. One popular approach to solving the algorithm selection problem problem is with meta-learning, meta-learning operates off of heuristics in the data set and how it performs with a given algorithm to provide feedback for later iterations. The method described in this thesis foregoes the analysis of the data set going into the clustering algorithm and instead proposes an a priori method that uses the characteristics of the scenario that forms the data set to decide which algorithm is best suited resulting in less overall computation on a large dataset. This thesis covers a particle filter based multi target tracking algorithm developed around wireless signal localization. The Adaptive Clustering Engine (ACE) is introduced as a sub-algorithm to the particle filter that decides the clustering to be used. Algorithm selection is driven by a set of scenario characteristic estimates to determine quantity of targets, observability of targets, and proximity between targets. Simulation tests were run to develop a selection map that correlates algorithm performance to localization scenario characteristics. Additional simulation testing is performed to validate the system. Validation testing shows ACE is capable of selecting a clustering algorithm that correctly estimated number of targets 88.09% of validation runs.