ANNE - A new algorithm for evolution of artificial neural network classifier systems

This paper introduces ANNE, a new algorithm for evolution of neural network classifiers. Different from standard divide and conquer approaches, the proposed algorithm evolves simultaneously the input feature vector, the network topology and the weights. The use of the embedded approach is also novel in an evolutionary feature selection paradigm. Tested on seven benchmark problems, ANNE creates compact solutions that achieve accurate and robust learning results. Significant reduction of the input features is obtained in most of the data sets. The performance of ANNE is compared to the performance of five control algorithms that combine different manual and automatic feature selection approaches with different structure design techniques. The tests show that ANNE performs concurrent feature selection and structure design with results that are equal or better than the best results obtained by algorithms specialised only on feature selection or neural network architecture design. Moreover, the proposed approach fully automates the neural network generation process, thus removing the need for time-consuming manual design.