Modelos através de redes neurais artificiais sem realimentação para dispositivos e circuitos de RF/micro-ondas.

Abstract

This thesis contributes to the development of neural network methodologies for modeling of RF/micro\vave devices and circuits. The applications of this alternative neurocomputational technique are reviewed. The study of the feedforward neural networks was the main objective. The supervisioned learning strategies, using the steepest descent method, were formulated for these neural networks. A novel neural network, nominated RFS, was proposed with the use of the sample activation function. The conventional method, EM-ANN. was applyed in many RF/microwave modeling problems and in the nonlinear circuit simulation. In the nonlinear circuit transient analysis, unpublished application examples of MLP models for GaAs MESFET transistors in circuit simulators, demonstrate the viability of the use of neural networks, as active device models, for time domain circuit analysis. In the nonlinear circuit steady-state analysis, was applied the Fourier network in conjunction with unpublished method, nominated simplified harmonic balance. Empirical based neural models were developed. A new neural network based space mapping technique, nominated FDPSMN, was proposed. All fomulated algorithms, as well as, the numericals methods used in this subject were implemented in Matlab™ and C++. Suggestions for further research were presented.