Cost-effective GRNN-based modeling of microwave transistors with a reduced number of measurements

Abstract

In this article, a simple, accurate, fast, and reliable black-box modeling is proposed for the scattering (S)-parameters and noise (N)-parameters of microwave transistors using the general regression neural network (GRNN) with the substantially reduced measurements and computational cost. In this modeling method, GRNN is employed as a nonlinear extrapolator to generalize the S-data and N-data belonging to only a single bias voltage in the middle region into the entire device operation domain of the bias condition (V-DS/V-CE, I-DS/I-C, f) within the shortened human effort. The proposed method is implemented to the modeling of the two transistors BFP640 and ATF-551M4 as study cases. Thus, comparisons are made with the multilayer perceptrons, trained by the two standard backward propagation algorithms, which are the Levenberg-Marquardt, Bayesian regularization and the 10 data mining methods recently published in the literature using the chosen training data sets in both nterpolation and extrapolation types of generalization. All the comparisons are achieved using four criteria commonly used in the literature. It can be concluded that GRNN is found to be a fast and accurate modeling method that extrapolates the reduced amount of training data consisting of measured S-parameters and N-parameters at the typical currents of the middle bias voltage to the wide operating range. Copyright (c) 2015 John Wiley & Sons, Ltd.