Mecanismo de clasificación de paisajes de optimización basado en muestreo multiescala y aprendizaje automático

Abstract

This work presents an approach to classify the modality in optimization landscapes, combining multiscale sampling with machine learning techniques. A set of optimization functions was selected and labeled according to the modality definition proposed by Kanemitsu et al. To minimize sample bias, a multiscale sampling algorithm was developed, based on the behavior of a random walk guided by a power law, complemented with fine-scale exploitation mechanisms, to both explore and exploit the optimization landscapes. The obtained samples are represented as an image, which is used as input to a convolutional neural network responsible for classifying the landscape modality. Experimental results show that the proposed approach achieves competitive performance in classifying previously unseen landscapes. Furthermore, the results suggest that the multiscale strategy provides more reliable information compared to random sampling, which is the standard technique in optimization landscape analysis. It is important to highlight that this work leads to the assertion that the problem of understanding optimization problems could be viewed as a pattern recognition problem.