Optimizing feed-forward neural network topology by multi-objective evolutionary algorithms: A comparative study on biomedical datasets

The design of robust classifiers, for instance Artificial Neural Networks (ANNs), is a critical aspect in all complex pattern recognition or classification tasks. Poor design choices may undermine the ability of the system to correctly classify the data samples. In this context, evolutionary techniques have proven particularly successful in exploring the complex state-space underlying the design of ANNs. Here, we report an extensive comparative study on the application of several modern Multi-Objective Evolutionary Algorithms to the design and training of an ANN for the classification of samples from two different biomedical datasets. Numerical results show that different algorithms have different strengths and weaknesses, leading to ANNs characterized by different levels of classification accuracy and network complexity.