An Efficient Batch-Adjustment Algorithm for Artificial Neural Network Structure Adaptation and Applications to Microwave Modeling

Artificial neural network (ANN) is a powerful technique in the microwave modeling area. The ANN structure adaptation in existing automated model generation (AMG) algorithms focuses on adjusting either the number of hidden layers or the number of hidden neurons, or adjusting both simultaneously but in a layer-by-layer manner. In this letter, a novel batch-adjustment algorithm for ANN structure adaptation is proposed to improve the ANN modeling efficiency. We introduce a cascaded multilayer ANN structure and propose a new training algorithm to train it. The ANN structure is automatically adjusted by adding or removing the redundant hidden layers in batch and makes a compromise between the number of hidden layers and the number of hidden neurons simultaneously. Compared to existing algorithms, our proposed algorithm is more flexible and efficient in ANN structure adaptation by performing fewer times of ANN training during the model development. Two microwave modeling examples are used to demonstrate the proposed algorithm.