Efficient Approximation Models of Microwave Devices Through Incremental Modeling

In this paper, an incremental modeling technique is proposed as an efficient substitute for the costly approximation methods used in behavioral modeling of active microwave devices. In this technique, cheap interpolation models are being extracted on a subset of training dataset. The rest of the dataset guides the algorithm which samples to include in the interpolation. The model is built by subsequently adding the samples that minimize the information loss given by Akaike information criterion. This criterion allows to choose a good compromise between the model accuracy and complexity. The results of the modeling of the complex wave quantities of GaN HEMT show that incremental modeling yields comparable and sometimes lower errors to the models interpolating with even order-of-magnitude more samples. The results also show that the incremental modeling greatly reduces overfitting.