Journal of Electronic Systems Kaoutar Senhaji, Hassan Ramchoun, Mohamed Ettaouil 7 4 2017 http://www.dline.info/jes/fulltext/v7n4/jesv7n4_1.pdf The multi-layer perceptron has proved its efficiencies in several fields as pattern and voice recognition. Unfortunately, the classical training for MLP suffers from a poor generalization. In this respect, we have proposed a new multiobjective training model with constraints which satisfies two objectives. The first one is the learning objective: minimizing the perceptron error and the second is the complexity objective: optimizing number of weights and neurons. The proposed model will provide a balance between the multi-layer perceptron learning and the complexity to get a good generalization. Our model has been solved using an evolutionary approach called the Non-Dominated Sorting Genetic Algorithm (NSGA II). This approach has led to a good representation of the Pareto set for the MLP network, from which an improved generalization performance model is selected.