Optimization of an artificial neural network for thermal/pressure food processing: Evaluation of training algorithms

The aim of the current paper is to obtain, through a proper selection of the training algorithm, an optimized artificial neural network (ANN) able to predict two parameters of interest for high-pressure (HP) food processing: the maximum or minimum temperature reached in the sample after pressurization and the time needed for thermal re-equilibration in the high-pressure process. To do that, 13 training algorithms belonging to 4 broad classes (gradient descent, conjugate gradient, quasi-Newton algorithms and Bayesian regularization) have been evaluated by training different ANNs. The network trained with the Levenberg–Marquardt algorithm showed the best overall predictive ability. The performance of this network was subsequently optimized by varying the number of nodes in the hidden layer, the learning coefficient and the decrease factor of this coefficient, and selecting the configuration with the highest predictive ability. The optimized ANN was able to make accurate predictions for the variables studied (temperature and time). These predictions were significantly better than those obtained by a previous ANN developed without selection of the training algorithm, that is, assuming the default option of the ANN computational package (gradient descent with a user-defined learning rate). We have shown that a correct selection of the training algorithm allows maximizing the predictive ability of the artificial neural network.