Automated feature detection of black-box continuous search-landscapes using neural image recognition

Abstract

Extracting search-landscape features of high-dimensional Black-Box Optimization (BBO) problems, by minimal number of queries, is a grand challenge. Knowledge of such features per a given BBO problem-instance offers valuable information in light of the Algorithm Selection and/or Configuration problems. In practice, by solving this challenge, the query complexity of BBO may be reduced when capitalizing on algorithm portfolios. In this study we target the automated recognition of BBO benchmark functions. We propose an identification framework for d-dimensional continuous BBO test-suites, provided as input with a set of N search-points, sampled at random, together with their query values. We address it as a supervised multi-class image recognition problem, by introducing the concept of Landscape Images, and applying the basic LeNet5 Neural Network to classify them. The solution's core lies within the encapsulation of the BBO functions' data as Landscape Images, and the application of neural image recognition to learn their features. The efficacy of our approach is numerically validated on the noiseless COCO/BBOB test-functions, which are demonstrated to be correctly classified at high precision rates (≈90%) when N is in the order of d. This successful learning is another step toward automated feature detection of BBO problems.