Node Level Crossover Applied to Neural Network Evolution

In this work we present an application of the Confidence Interval Based Crossover using L2 Norm (CIXL2) and BLX-α crossovers to the evolution of neural networks. CIXL2 is a new crossover operator, based on obtaining the statistical features of the best individuals of the population. These features are used as virtual parents for the crossover operator. Due to the permutation problem of neural network coding that negatively affects the crossover operator, we have adopted a novel approach. Crossover is made at node level. Instead of performing crossover over two whole networks, we perform crossover over two nodes. We present here the adaptation of two crossover methods: CIXL2 and BLX-α. All the nodes of the best networks are clustered, by means of a k-means algorithm, and a redefinition of the operators is carried out using the obtained clusters. Each node is mated within the cluster where it belongs.