Today at the COSEAL 2026 workshop in Aachen, Germany, Sara Gjorgjevikj presented the work titled “Learning to Assess the Reliability of Number-of-Runs Estimation in Stochastic Optimization.”

The presented research addresses one of the key challenges in large-scale benchmarking of stochastic optimization algorithms: determining when enough repeated runs have been collected to obtain reliable conclusions without wasting unnecessary computational resources. The work extends a recent empirical online heuristic with learning-based reliability assessment using statistical, shape, stability, and energy-free features.

Using more than 132,000 Nevergrad runs on the COCO benchmark suite across 11 optimizers, the study investigates whether machine learning models can identify potentially unsafe early stopping decisions during benchmarking. Across multiple experimental setups, ensemble models achieved strong minority-class recall, demonstrating promising potential for more trustworthy and computationally efficient benchmarking workflows in stochastic optimization.