Abstract

monte carlo algorithm. A crucial problem which arises when dealing with Bayesian neural networks is that of determining their most appropriate size, expressed in terms of number of computational units and/or connections. In fact, too small a network may not be able to learn the sample data, whereas one that is too large may give rise to overfitting phenomena and cause poor “generalization ” performance. A few solutions have been proposed in the literature to solve this problem, such as the use of a geometric prior probability on the number of hidden units (Müller and Rios Insua, 1998), thereby favouring smaller-size networks, and a reversible jump algorithm to move between architectures having a different number of hidden units (Rios Insua and Müller, 1998). In this work we propose a variable architecture model where input-to-hidden connections and, therefore, hidden units are selected by using a variant of the Evolutionary Monte Carlo (EMC) algorithm developed by Liang and Wong (2000). The

Citations