Bayesian-Based Automated Model Generation Method for Neural Network Modeling of Microwave Components

Automated model generation (AMG) has become a popular technique for systematically developing neural network models by avoiding manual trial-and-errors. However, when the initial number of hidden neurons is far from the optimal value, the existing AMG methods usually take a relatively large amount of CPU time to find the optimal structure. To deal with this problem, for the first time, Bayesian-based formulation is introduced into the AMG method. The proposed Bayesian-based AMG method can efficiently determine the minimum number of hidden neurons with maximum accuracy during the model development process. The proposed method can greatly reduce CPU time for model generation in comparison with the existing AMG technique. A microwave filter example is used to demonstrate the proposed method.