Multi-criteria optimization of neural networks using multi-objective genetic algorithm
Kaoutar Senhaji, Mohamed Ettaouil · 2017
This paper propose a new multi-objective model optimization allow training the multi-layer perceptron neural network (MLPNN) and optimizing its architecture. More precisely, this model aims to satisfy two objectives: the first one is minimizing the perceptron error (training objective) and the second one is minimizing the sum of the absolute weights (optimizing architecture objective). As known, a multi-objective problem's optimal solution is a set called Pareto set, from which a single weight vector with best performance will correspond to a reduced number of weights; this set is a compromise between the two objectives. To solve the proposed model, we have chosen the NSGA II algorithm (Non-Dominated Sorting Genetic Algorithm II). This algorithm has shown to be very powerful for multi-objective optimization, thus for multi-objective learning.