Skip to Main Content
Multilayer perceptrons (MLPs) are a standard tool for establishing relationships between data in many real world problems, in the absence of a parametric model. In the last decade, they have often been used for analyzing data produced by arrays of chemical sensors [electronic noses (e-noses)]. Still, the central issue of controlling the complexity of an MLP for optimal generalization is frequently overlooked by chemical sensors practitioners causing incorrect or suboptimal results (over or underfitting). In this paper, we will: 1) present different ways of controlling the complexity of an MLP (model order selection, early stopping, and regularization); 2) shortly review the literature on complexity control, inside and outside the e-nose community; and 3) give examples of effective complexity control for two e-noses datasets of different size and learning difficulty. It will be shown that, if early stopping or regularization are adopted, overfitting is avoided whatever the number of hidden units (and, hence, network weights). Another issue tackled in this paper is the influence on the generalization error of the number of principal components over which data are projected (before being fed into the MLP). Simulations show that (test set) performance depends strongly on the number of principal components and that even components with less than 1% of the global variance enhance classification.