Direct and Indirect Evolutionary Designs of Artificial Neural Networks

Evolutionary Algorithms (EAs) and other kind of metaheuristics are utilized to either design or optimize the architecture of Artificial Neural Networks (ANNs) in order to adapt them for solving a specific problem; these generated ANNs are known as Evolutionary Artificial Neural Networks (EANNs). Their architecture components, including number of neurons or their kind of transfer functions, connectivity pattern, etc., can be defined through direct or indirect encoding schemes; the former, directly codifies ANN architecture components into the genotype of solutions, meanwhile the last one, requires to transform the solution’s genotype through mapping processes to generate an ANN architecture. In this work, architectures of three-layered feed-forward ANNs are designed by using both encoding schemes to solve several well-known benchmark datasets of supervised classification problems. The results of designed ANNs by using direct and indirect encoding schemes are compared.