Correlation analysis and performance evaluation of distance measures for evolutionary neural networks

In a genetic algorithm, the search process maintains multiple solutions and their interactions are important to accelerate the evolution. If the pool of solutions is dominated by the single fittest individual in the early generation, there is a risk of premature convergence losing exploration capability. It is necessary to consider not only the fitness of solutions but also the similarity to other individuals. This speciation idea is beneficial to several application domains with evolutionary computation but it requires objective distance measures to calculate the similarity of individuals. It raises a challenging research issue to measure the distance between two evolutionary neural networks (ENN). In this paper, we surveyed several distance measures proposed for ENN and compared their performance for pattern classification problems with two different genetic representations (matrix-based and topology growing (NEAT) approaches). Although there was no dominant distance measure for the pattern classification problems, it showed that the behavioral distance measures outperformed the architectural one for matrix-based representation and. For NEAT, NeuroEdit showed better accuracy against compatibility distance measure.