Identification of Discrete Non-Linear Dynamics of a Radio-Frequency Power Amplifier Circuit using Symbolic Regression

Abstract

The identification of non-linearities or undesirable dynamic behavior of electrical components is a common problem. Previous modeling forms are largely based on extensive physical knowledge at the semiconductor level, which has produced reliable solutions over the past decades. This however implies the measurement of physical prototypes in laboratories, which can be costly. It is therefore desirable to have reliable software models of the prototypes available to outsource this procedure to simulators. This paper presents a number of solutions from the field of empirical modeling including symbolic regression, which allow to parameterize such models from measured values. As an example we are utilizing time-domain data from a real radio-frequency power amplifier circuit. We compare a Hammerstein-Wiener model with two methods for symbolic regression, and find that the Hammerstein-Wiener model produces the best predictions but has many non-zero coefficients. Both symbolic regression methods produce short linear models with slightly higher prediction error than the HW model.