Neural input space mapping optimization based on nonlinear two-layer perceptrons with optimized nonlinearity

Abstract

A neural space mapping optimization algorithm based on nonlinear two layer perceptrons (2LP) is described in this article. This work is an improved version of the Neural Space-Mapping (NSM) algorithm that uses three layer perceptrons (3LP) to implement a nonlinear input mapping function at each iteration. The new version uses a nonlinear 2LP whose nonlinearity is automatically regulated with classical optimization algorithms. Additionally, the new algorithm uses a different optimization method to train the SM-based neuromodel and a more efficient manner to predict the next iterate. With these improvements, we obtain a more efficient and faster algorithm. To verify the algorithm performance, we design some synthetic circuits, as well as a stopband microstrip filter with quarter-wave resonant opens stubs, a bandpass microstrip filter, and a microstrip notch filter with mitered bends. The last three cases use commercially available full-wave electromagnetic simulators.

Accurate large-signal modeling using neuro-space mapping for power transistors

“…This technique uses neural networks to map the voltage or current signals of the existing device model into that of the device data. The Neuro-SM method can be applied to not only the simple DC and S-parameters modeling of nonlinear devices, but also the complex large-signal modeling [6,7,8]. In [9], a dynamic neural network is used as the mapping network for the Neuro-SM model of power transistors.…”

Robust microwave design optimization using adjoint sensitivity and trust regions

“…Moreover, an extra simulator must be involved in the process and linked to the optimization algorithm. Another problem is that performance of SM algorithm depends heavily on the selection of the SM transformations used to construct the surrogate [21].…”

Response correction techniques for surrogate-based design optimization of microwave structures

“…A number of techniques for modeling and simulationdriven design of microwave structures have emerged over the recent years, including the methods that exploit artificial neural networks [8,[9][10][11][12], fuzzy systems [13], kriging [14], or multidimensional Cauchy approximation [15], as well as surrogate-based techniques such as space mapping (SM) [16][17][18][19][20][21][22][23]24], simulation-based tuning [25,26], and combination of both [27,28]. The last three approaches offer computationally efficient design optimization where, under certain circumstances, a satisfactory design can be obtained after a few high-fidelity (or fine) EM simulations of the structure of interest [16].…”