Automated Multilayer Neural Network Structure Adaptation Method With l₁ Regularization for Microwave Modeling

Abstract

Artificial neural network (ANN) model development for microwave components principally includes two parts of work, i.e., data sampling and model structure adaptation. In existing various ANN modeling methods, the model structure adaptation process mainly focuses on adjusting the number of neurons within each hidden layer of ANN while keeping the number of layers unchanged. To make the ANN modeling process more flexible and efficient, an automated multilayer neural network structure adaptation method with <inline-formula> <tex-math notation="LaTeX">$ {l_{1}}$ </tex-math></inline-formula> regularization is proposed in this letter. We propose a new ANN model structure combining multilayer perceptron (MLP) and additional connections between the output layer and each hidden layer/input layer. A new training scheme with <inline-formula> <tex-math notation="LaTeX">$ {l_{1}}$ </tex-math></inline-formula> regularization is proposed to automatically determine the final model structure with user-desired model accuracy. Using the proposed model structure adaptation method, both the number of layers and the number of neurons within each layer of the final ANN model can be adaptively determined to address different needs for different microwave modeling problems. The proposed method is demonstrated by two microwave filter modeling examples in which the model development process achieves a time saving of at least 40&#x0025; over existing methods.