Abstract
Many analyses show that Artificial Neural Network (ANN) based final solutions are more likely to trapped in local minimums, provide poor generalization and extrapolation behaviors, and are not unique due to the inappropriate values of initial weights. To address these issues, this paper develops and presents accurate small-signal behavioral modelling schemes for gallium nitride high electron mobility transistor by exploiting the global minimums oriented initial weights generated by Ant Lion Optimization (ALO) and the variant of Ant Colony Optimization for continuous domains (ACOR) with ANN. The models are trained and tested on many operating conditions, and simultaneously examined for temperature interpolation and extrapolation tests. Moreover, ANN, ALO-ANN, and ACOR-ANN based models are also examined for numerous conventional regression metrics. We found that ALO-ANN and ACOR-ANN based models provide more accurate behavioral models and better convergence behaviors than ANN based models.