The effect of genetic operations on the diversity of evolvable neural networks

Evolutionary algorithms have been used as powerful tool in designing neural networks. Evolutionary algorithms can be used for various tasks, such as connection weights training, structure design, learning rule adaptation, input feature selection, and rule extraction from neural networks. In this paper we study the effect of genetic operators crossover and mutation on the diversity of a population of evolvable neural networks. Diversity is an important aspect of evolvable neural networks. The lack of diversity leads to premature convergence problem and stagnation in local minima. The correlation between population diversity and fitness is also studied and the overall effect of diversity on the evolution of neural networks is investigated. Experimental results show that mutation operator has promoting effect on the diversity more than crossover and there is strong correlation between diversity generated by mutation operator and maximum fitness of population. But the change of maximum fitness associated with crossover operator is greater than one associated with mutation operator. © 2008 IADIS.