Evaluation of the Use of the irace Tool for the Optimization of Hyperparameters of Convolutional Neural Networks

Deep Learning is a field of Artificial Intelligence that deals with artificial neural networks and has been used in various applications and problems, especially in Computer Vision tasks. Training a neural network, a fundamental step, can take a considerable amount of time and demand a lot of computational effort, depending on the amount of data and the size of the network. Typically, its training also depends on the configuration of its hyperparameter values, and it is often necessary to perform it several times, considering different configurations, in order to choose the one that leads the network to the best performance. This work proposes and evaluates the use of the irace tool, adopted in the area of metaheuristics, in the process of optimizing the hyperparameter values of a convolutional neural network trained for the CIFAR-10 dataset, carrying out computational experiments and comparing the results obtained with those from the application of the Optuna tool.