Non-linear radio frequency model identification using a hybrid genetic optimiser for minimal user intervention

An adaptive reduced-order procedure for developing grey-box models of radio frequency systems is presented. The stochastic model extraction combines a modified genetic algorithm and the Nelder-Mead simplex algorithm to present a user with a range of possible good models from which the user can use intuition to select the most physically realistic candidate. The procedure is ideal for identifying models given incomplete observations, noisy data and inexact model structure. This procedure replaces the commonly used human-in-the-loop ad hoc deterministic approach in which a skilled operator must guide model fitting. Being able to use an expanded range of model architectures beyond the Wiener and Hammerstein family, the stochastic approach, utilisation of incomplete observations and extraction of multiple distinct model candidates, enables the insight and intuition of a skilled operator to be used to advantage. The extraction of a multistage microwave amplifier exhibiting long-term memory effects is used as an example.