Design of experiments by committee of neural networks

In this paper, we present a way of constructing design of experiments for neural networks models such as multi-layer perceptron (MLP). We are trying to solve the problem of modeling a phenomenon with a minimum of measurements and almost no a priori knowledge. Our method is based on query by committee (QBC) which compares the predictions of various models on unsampled locations in order to select the most informative. We compare it to a random selection of samples.