Measuring the Genotype Diversity of Evolvable Neural Networks

Studying the diversity of genetic algorithms is the most important topic to prevent the problem of premature convergence of the population. In general diversity study can be divided into, measuring diversity of the population and maintaining the diversity of the population. This paper is interested in measuring the diversity. Measuring population diversity can be categorised into genotype diversity measure and phenotype diversity measure. Most of the previous research was devoted to measure diversity of genetic programming and genetic algorithms in general. The present paper introduces a dedicated study to measure and analyse the genotype or structure diversity of evolvable neural networks. A new metric is introduced to measure the genotype diversity of evolvable neural networks based on genetic operations required to transform one neural network to another one. This method is called neuro-edit, it is inspired on edit distance measure and genetic operations used to evolve neural networks. This metric measures the distance between neural networks in terms of connection genes addition, deletion, and substitution.