This paper presents an advanced algorithm for automated model generation (AMG) using neural networks. AMG trains a neural network in a stage-by-stage manner to obtain a neural network of required accuracy with least amount of training data. In each stage, either the number of data or the size of the neural network is adjusted. The novelty of the proposed algorithm is to incorporate efficient interpolation approaches to make the AMG process much faster. We add an additional procedure to minimize the number of hidden neurons, which makes the final neural-network model more compact compared with the previously published AMG. Examples including automated modeling of MOSFETs and bandpass filters are presented showing the advantage of this technique.