Kernel Based Machine Learning Using Received Signal Strength Indicators For Indoor Localization

With the development of the machine learning,RSSI-based localization problem can be formulated as a machine learning problem.As an important technique of support vector regression(SVR),kernel-based learning algorithm has been widely utilized for localization system.However,the existing SVR algorithms are often based on single kernel.However,whether global kernel or local kernel is not the optimal solution for training data set learning.Thus,in this paper,we study the kernel-based machine learning localization algorithm.The main contributions are as follows:(1)First,the knowledge of fingerprint based localization technique and some common matching approaches are introduced.Then,some concepts of kernel-based machine learning localization technique are studied in which the processing in both off-line phase and on-line phase are described in detail.(2)The hybrid kernel-based machine learning algorithm for indoor localization is proposed.In the off-line phase,after the training set pre-processing by the iterative self-organizing data analysis techniques algorithm(ISODATA),the measurement-label set is utilized for classification learning by c-support vector classification approach.Moreover,each measurement-position training subset is chosen for regressing learning.In the on-line phase,based on the classification result of received RSSI measurement,the corresponding regression function is chosen for target position estimation.Since the hybrid kernel has better interpolation and extrapolation performance than single kernel,it can improve the training ability in the off-line phase and localization performance in the on-line phase.Meanwhile,a v-cross validation-based optimization algorithm is introduced to obtain both optimal kernel parameters and weight parameters of each kernel.Thus,it gives a technical guarantee for target localization.(3)The multiple kernel-based machine learning algorithm for indoor localization is proposed.In the off-line phase,after the training set pre-processing by ISODATA,the measurement-label set is utilized for classification learning by c-support vector classification approach.Moreover,each measurement-position training subset is chosen for regressing learning by SPG-based generalized multiple kernel learning.In the on-line phase,based on the classification result of received RSSI measurement,the corresponding regression function is chosen for target position estimation.Since the size of training set become smaller,the learning performance in the off-line phase will be enhanced and the calculation complexity will be decreased.Compared with other optimizers,SPG can reduce the learning time.This method provides theoretical supports for practical application.