Memetic Pareto Differential Evolution for Designing Artificial Neural Networks in Multiclassification Problems Using Cross-Entropy Versus Sensitivity

This work proposes a Multiobjective Differential Evolution algorithm based on dominance Pareto concept for multiclassification problems using multilayer perceptron neural network models. The algorithm include a local search procedure and optimizes two conflicting objectives of multiclassifiers, a high correct classification rate and a high classification rate for each class, of which the latter is not usually optimized in classification. Once the Pareto front is built, we use two automatic selection methodologies of individuals: the best model with respect to accuracy and the best model with respect to sensitivity (extremes in the Pareto front). These strategies are applied to solve six classification benchmark problems obtained from the UCI repository. The models obtained show a high accuracy and a high classification rate for each class.