EL ROL DEL ALGORITMO DE ENTRENAMIENTO EN LA SELECCIÓN DE MODELOS DE REDES NEURONALES

The Neural net’s fit ability is often affected by the network configuration, particularly the number of hidden neurons and input variables. As the size of these parameters increases, the learning also increases, then the fit of network is better. Theoretically, if parameters are increasing regularly, the error should be reduced systematically, provided that the models are nested for each step of the process. In this work, we validated the hypothesis that the addition of hidden neurons in nested models lead to systematic reductions in error, regardless of the learning algorithm used; to illustrate the discussion we used the number of airline passengers and Sunspots in Box & Jenkins, and RProp and Delta Rule as learning methods. Experimental evidence shows that the evaluated training methods show different behaviors as those theoretically expected, it means, not fulfilling the assumption of error reduction.