Modular neural networks (MNNs) are increasingly popular models for dealing with complex problems constituted by a number of dependent subtasks. An important problem on MNNs is finding the optimal aggregation of the neural modules, each of them dealing with one of the subproblems. In this paper, we present a functional network approach, based on the minimum description length quality measure, to the problem of finding optimal modular network architectures for specific problems. Examples of function approximation and nonlinear time series prediction are used to illustrate the performance of these models when compared with standard functional and neural networks.Eje: Sistemas inteligentesRed de Universidades con Carreras en Informática (RedUNCI)