A comparison of NN-based and SVR-based power prediction for mobile DS/CDMA systems

Abstract

We further investigate the performances of our previously proposed technique for received signal power prediction in the direct sequence code division multiple access (DS/CDMA) systems based on support vector regression (SVR.) The scheme is based on one-step ahead prediction using the past values of signal series as the inputs. The predictor parameters are chosen by considering the minimum mean square error (MMSE.) We compare the performances of the proposed predictor to that of the linear and nonlinear neural network-based predictors, i.e., the adaptive linear (Adaline) predictor, multilayer perceptrons (MLP) predictor and the hybrid predictor (Adaline cascade with MLP.) The carrier frequency of 1.8 GHz and a noisy Rayleigh fading channel are considered. The vehicle speeds are set to 5 km/h and 50 km/h. Cross validation is also applied to improve the prediction performance of our technique. The results on the blind test data show that the SVR-based predictor using the five-fold cross validation yields the best prediction performance among the aforementioned predictors.