Reducing a number of full-wave analyses in RBF neural network optimization of complex microwave structures

This paper introduces an original technique of neural network optimization of substantially less computational cost and suitable for viable CAD of complex microwave systems. A number of necessary full-wave simulations are dramatically reduced due to a special objective function and CORS sampling in the dynamic training of the decomposed RBF network. Performance of the algorithm is illustrated by its application to a waveguide band-pass filter and a dielectric resonator antenna; their optimal designs are found from 5-parameter optimizations requiring as little as 167 and 99 analyses, respectively.