Abstract

The aim of Neuroevolution is to find neural networks and convolutional neural network (CNN) architectures automatically through evolutionary algorithms. A crucial problem in neuroevolution is search time, since multiple CNNs must be trained during evolution. This problem has led to fitness acceleration approaches, generating a trade-off between time and fitness fidelity. Also, since search spaces for this problem usually include only a few parameters, this increases the human bias in the search. In this work, we propose a novel two-level genetic algorithm (GA) for addressing the fidelity-time trade-off problem for the fitness computation in CNNs. The first level evaluates many individuals quickly, and the second evaluates only those with the best results more finely. We also propose a search space with few restrictions, and an encoding with unexpressed genes to facilitate the crossover operation. This search space allows CNN architectures to have any sizes, shapes, and skip-connections among nodes. The two-level GA was applied to the pattern recognition problem on seven datasets, five MNIST-Variants, Fashion-MNIST, and CIFAR-10, achieving significantly better results than all those previously published. Our results show an improvement of 39.89% (4.2% error reduction) on the most complex dataset of MNIST (MRDBI), and on average 30.52% (1.35% error reduction) on all the five datasets. Furthermore, we show that our algorithm performed as well as precise-training GA, but took only the time of a fast-training GA. These results can be relevant and useful not only for image classification problems but also for GA-related problems.