Sujit Ghosh, North Carolina State

Title: Are We There Yet? A Probabilistic Journey to Global Optima

Date: Friday, March 13th, 2026
Time: 1:30PM (PDT)
Location: ASB 10900

Abstract: Global optimization over non-convex landscapes is notoriously difficult—akin to a long road trip with no map, no gradients, and lots of local distractions. Traditional algorithms, including many metaheuristics, often struggle with high-dimensionality, sensitivity to initial conditions, or inaccessibility of derivative information—frequently converging to suboptimal solutions (if at all). In this talk, I will introduce ProGO—a Probabilistic Global Optimizer—that leverages a novel integration-based framework for approximating global optima without relying on gradient information. ProGO is underpinned by a mathematically rigorous convergence theory, built on the asymptotic behavior of a so-called nascent optima distribution. To efficiently sample from this evolving distribution, we develop a latent slice sampler that exhibits geometric convergence, thereby addressing the curse of dimensionality without brute-force enumeration. Under mild regularity conditions, we show global convergence of the algorithm in probability. Empirical studies across a suite of challenging benchmark functions with multiple local minima demonstrate that ProGO outperforms several state-of-the-art optimization methods—including gradient-based, zeroth-order, and some Bayesian optimization approaches—in both convergence speed and regret. While not ideal for functions that are prohibitively expensive to evaluate, ProGO offers a robust, scalable, and theoretically sound alternative to navigating the global optimization landscape—no GPS required!