Application of Genetic Algorithms to the Construction of Topologies for Multilayer Perceptrons

In this paper we present a new approach for automatic topology optimization of backpropagation networks. It is based on a genetic algorithm. In contrast to other approaches it allows that two networks with different number of units can be crossed to a new valid "child" network. We applied this algorithm to a medical classification task, which is extremely difficult to solve. The results confirm, that optimization make sence, because the generated network outperform all fixed topologies. 1 Introduction As Minsky and Papert [1] have shown the XOR-- problem cannot be solved without a hidden layer. A learning rule which is able to train this kind of networks was developed by Rumelhart et al. [2]. It is known as backpropagation (BP) and it is one of the most often used neural network paradigms. Hornik [3] proved that every function can be approximated by a neural network with just one hidden layer. By adding "enough" hidden units the approximation error can be made as small as required. Th..