4.10 Feedforward Neural Network Training: Examples

In this section, three conductor microstripline and physics-based MESFET modeling examples are used to compare the performance of various training algorithms for feedforward neural networks namely, the adaptive back propagation, conjugate-gradient, Quasi-Newton, and Levenberg-Marquardt algorithms [2].

For the microstripline example, there are five input neurons corresponding to conductor width ( w), spacing between conductors ( s ₁, s ₂), substrate height ( h), and relative permittivity ( ? _r), as shown in Figure 4.29. There are six output neurons corresponding to the self-inductance of each conductor ( L ₁₁, L ₂₂, L ₃₃), and the mutual inductance between any two conductors ( L ₁₂, L ₂₃, L ₁₃). A total of 600 training samples and 640 test samples were generated using LINPAR [54]. A three-layer MLP structure with 28 hidden neurons was used, and the training results are shown in Table 4.4. The CPU time was given for a 200 MHz Pentium.

Figure 4.29: A three-conductor microstripline.