Gold and Silver at GECCO 2024: Star Discrepancy and Many-affine BBOB

Two competition wins at GECCO 2024, held July 14–18 in Melbourne, Australia — gold in the Star Discrepancy Competition and silver in the Anytime Algorithms for Many-affine BBOB Functions competition.

Gold — Star Discrepancy Competition

Star discrepancy is a deceptively simple question: given a set of points scattered inside a box, how evenly do they actually cover it? The answer — the L∞ star discrepancy — is a cornerstone measure in quasi-Monte Carlo methods, numerical integration, and experimental design, but it’s notoriously expensive to compute exactly as dimensions grow.

The GECCO 2024 Star Discrepancy Competition, organized by Carola Doerr, Francois Clement, Diederick Vermetten, Jacob de Nobel, Thomas Bäck, Luís Paquete, and Kathrin Klamroth, asked participants to build solvers that push the boundary of what’s tractable — estimating star discrepancy accurately across a range of dimensions and point-set sizes, whether by exploiting the problem’s structure directly (white-box) or through iterative, sampling-based search (grey-box), the territory where metaheuristics live.

The winning entry applied the Andean Condor Algorithm (ACA) — the evolutionary algorithm developed during my doctoral thesis, inspired by the foraging behavior of the Andean condor — to the discrete star discrepancy problem. ACA balances its population between exploration and intensification moves using a performance indicator based on average fitness quality, which is what let it search this space effectively. Source code and experiment data are published on Figshare.

Winning gold here sits alongside the GECCO 2023 high-dimensional track win as validation that the approach generalizes rather than being a one-off result.

Silver — Anytime Algorithms for Many-affine BBOB Functions

Most optimization benchmarks only ask one question: what’s the best solution you found by the end? The Many-affine BBOB competition, organized by Diederick Vermetten, Konstantin Dietrich, Pascal Kerschke, and Carola Doerr, asks a harder one: how good is your solution at every point along the way?

The MA-BBOB suite combines the 24 base functions from the classic BBOB (Black-Box Optimization Benchmarking) benchmark into affine combinations, generating roughly a thousand training instances that stress-test how an algorithm balances exploration against exploitation under a limited budget. Anytime performance is exactly what it sounds like — an algorithm that converges beautifully at generation 10,000 is no use if it’s still hunting blindly at generation 100.

The competition allows up to three entries per team, and I submitted two:

• Progressive Sample Scaling (PSS) — the entry that won silver. PSS is a deterministic algorithm inspired by geological sampling of terrain, adapted to work across dimensions. It runs in two stages: a sampling stage that generates combinations of equidistant points and evaluates their fitness to find the best one, and a scaling stage that increases the density of points around that best solution and refines the search progressively until the budget runs out. That progressive refinement is precisely what rewards it on an anytime metric — there’s always a reasonable answer on the table, and it keeps improving. Source code and experiment data: Figshare.
• Andean Condor Algorithm (ACA) — the second entry, applying the same bio-inspired metaheuristic behind the Star Discrepancy gold above to the Many-affine BBOB suite instead. It didn’t place here — ACA’s exploration/intensification balance is tuned for the kind of geometric search star discrepancy problems reward, not for the early-budget anytime behavior this competition specifically measures. Still a useful data point: the same algorithm doesn’t automatically transfer its strengths across problem types, which is exactly why PSS was designed as a distinct approach rather than reusing ACA here. Source code and experiment data: Figshare.

Two different flavors of black-box optimization, same GECCO 2024 edition, both problems where the metaheuristics built in this workshop had something to say.