Improved MSVR-Based Range-Free Localization Using a Rational Quadratic Kernel Function

Wireless sensor networks have been suggested for various applications such as smart structures, monitoring, routing and many others. Most of these applications are location-dependent which makes localization problem be one of the most challenging issues for such networks. Many range-free localization algorithms based on hop proximity suffer from low accuracy in both isotropic and anisotropic networks. In this paper, the range-free localization problem using machine learning (ML) is formulized as a regression or classification problem. Then, support vector regression (SVR) or support vector machine (SVM) is used in order to train. In most of the existing localization algorithms, radial basis functions (RBF) are considered for both training and testing stages in ML. Since the accuracy of localization is very important, using these functions as basis functions do not lead to desirable results in all defined topologies. We find that rational quadratic kernel (RQK) is a good alternative for kernel function in (SVR). The value of constant parameter used in desired kernel function is obtained through data evaluation method. Simulation results show that the proposed kernel function has the smallest error among all existing methods. Hence, it could be replaced with Gaussian kernels during the process of training and testing for localization problems.