Optimization Landscape Classification with Multi-Scale Sampling and Deep Neural Networks

This paper presents an approach to classify the modality of optimization problems that couples multi-scale sam-pling with machine learning techniques. To mitigate the bias of the sample, the multi-scale strategy capitalizes on exploration with power-law random walks and fine-scale exploitation mecha-nisms to sample the optimization landscape across various scales. This set of samples shapes an image that is used to feed a convolutional neural network, which classifies the modality of the optimization landscape. Experimental results show the proposed approach exhibits competitive performance in the classification of noteworthy unseen landscapes. In addition, results suggest the multi-scale strategy would provide more reliable information with respect to uniform random sampling.