Effective optimization techniques for RF circuits

The best gradient and direct-search algorithms for constrained optimization of RF circuits are compared. Two complete, similar programs were written in the QuickBASIC language for personal computers to compare the Nelder-Mead simplex direct-search minimizer to a Gauss-Newton minimizer with Levenberg interpolation, automatic scaling, and finite-differenced derivatives. Powell´s method for nonlinear bounds on circuit responses was used in both programs. It is concluded that the gradient method with approximate derivatives was only slightly better than the direct-search method. Neither program was anywhere near as fast and accurate as one that used exact partial derivatives with the same techniques. Clearly, the complicated arrangements for computing exact gradients are worthwhile for both speed and effectiveness