An improved multi-objective genetic algorithm for the neural architecture search problem.

Abstract

In recent years, there is a great interest in automating the process of searching for neural network topology. This problem is called Neural Architecture Search (NAS), which can be seen as a 3-gear mechanism: the search space, the error estimation and the search strategy. To guide the selected strategy throughout the search space, we need a metric to help us. The simplest way is to evaluate the error obtained in the validation set, however, due to the long computation times required, alternative methods are being searched for, such as: reducing the training set, reducing the number of epochs , using less filters or using lower resolution images. In this paper, we propose an improved version of the NSGA-Net algorithm, which is a multi-objective genetic algorithm for the NAS problem. One of the drawbacks is the limited diversity that can be generated by the original crossover operator, which generates only one offspring keeping the common genomes, and leaving the rest randomly. In order to avoid this limitation, we proposed a new 2-point crossover restricting the possible cutoff points only to the block limits.