A study using the robust design of artificial neural networks methodology in neutron spectrometry

In this work, a systematical and methodological ANN optimization process known as robust design of artificial neural networks methodology, based on Taguchi method and Design of Experiments methodology, was applied to the design, training and testing of feed forward artificial neural networks trained with back-propagation training algorithm applied in the neutron spectrometry research area. The methodology was utilized to study the neutron spectrum unfolding problem by using a data set composed by 187 neutron spectra compiled by the International Atomic Energy Agency. In order to study the behavior of the designed neural networks topologies, four cases of grouping the neutron spectra were considered. In the first case 17 neutron spectra subsets were tested. In the second one 6 subsets. In the third case a data set with 53 neutron spectra and finally, in the fourth case, a data set with 187 neutron spectra was tested. For all the subsets, the robust design of artificial neural networks methodology was carried out. Around 1000 different neural network topologies were trained and tested, 36 net topologies for each subset. After all the network topologies were trained and tested, it was observed that the near optimum neural network topology which produced the best results was the fourth case, with the following topology: 7 neurons in the input layer, corresponding to the Bonner spheres readings; 14 neurons in a hidden layer and 31 neurons in the output layer corresponding to the 31 energy bins in which the spectrum is expressed, a learning rate and momentum equal to 0.1. The results obtained reveal that the robust design methodology offer potential benefits in the evaluation of the behavior of the net as well as the ability to examine the interaction of the weights and neurons inside the same one.