Brief

Two new research papers explore novel methods for selecting the best algorithm for continuous black-box optimization tasks. One paper, GeoPAS, uses geometric probing to create 2D slices of the objective landscape, encoding these slices to represent problem instances and then selecting solvers based on a composite score. The other paper, using CNNs, visualizes probed landscapes as contour plots, feeding these images into a convolutional neural network to predict solver performance and guide selection. Both approaches aim to significantly outperform relying on a single best solver, demonstrating improved efficiency and robustness across various optimization scenarios. AI

Researchers have introduced BBO-Pile, a novel open-source dataset containing over 500,000 optimization trajectories across nearly 3,100 black-boxes. This dataset aims to address the limitations of previous work, which relied on non-public or synthetic data, thereby hindering reproducibility and real-world generalization. By using BBO-Pile, foundation models for black-box optimization have been trained at various scales, demonstrating the effectiveness of large-scale pre-training for imitating optimization methods. AI