Speaker: Dr Mario Andrés Muñoz – (Optima/Melbourne University)

Bio:

Mario Andrés Muñoz Acosta is a Senior Research Fellow in the School of Computing and Information Systems at the University of Melbourne and a member of the ARC Training Centre in Optimisation Technologies, Integrated Methodologies, and Applications (OPTIMA). He received his PhD from the University of Melbourne in 2014, then held research fellow positions at Monash University and in the University of Melbourne’s School of Mathematics and Statistics before joining CIS and OPTIMA in 2022. His research studies the relationship between algorithm performance and problem structure, with a focus on landscape analysis, instance space analysis, and kernel methods for black-box optimisation. His current work develops distribution-based distances between optimisation problem instances.

Abstract:

Algorithm selection in black-box optimisation relies on describing the structure of the problem instance being solved. Exploratory Landscape Analysis (ELA) has provided the dominant approach: sampled evaluations are compressed into a fixed set of hand-crafted features, which are then used to relate landscape structure to algorithm performance. This feature-based view has been highly influential, but it also has well-known limitations. ELA features can depend on the benchmark suite, the transformation applied to the instance, and the sampling design, and they do not define a principled distance between problem instances.

This talk follows the development of alternatives to this feature-based pipeline. I will begin with early information-theoretic landscape features and their sensitivity to transformations and random evaluation grids. I will then discuss footprint-based analysis for comparing algorithm performance across instance space. The final part introduces LANDMARK, a representation that treats each landscape as a probability distribution and compares instances directly through a kernel-based discrepancy, without feature engineering. On the BBOB suite, this distance recovers meaningful instance-space structure and supports algorithm selection with accuracy comparable to ELA, while remaining interpretable and extensible to constrained and multi-objective settings.

This event is Hybrid:
JOIN VIA ZOOM – MEETING ID: 873 1557 5255; PASSWORD: 778635
Level 8, Room 8109, Melbourne Connect

SEMINAR: WED 30 SEPTEMBER 2026 4.00PM-5.00PM (AEST, Melbourne Time)