Title: Describing algorithm behaviour using recurrence quantification and landscape analyses

Abstract: Differences in performance between algorithms can be attributed to the interaction between their unique rule-sets and the characteristics of the instance’s landscape. However, understanding this interaction can be difficult because algorithms are often composed of multiple elements, and in the worst cases are described using opaque notation and metaphors. In this talk, we present an advance on using landscape analysis and recurrence quantification analysis to visually examine the relationship between a problem instance and the behaviour of an algorithm on it. Using as benchmarks 15 instances from the 24 function COCO set at {2,5,10,20} dimensions, we tested three algorithms: PSO, GWO and SGA. Although tested with default parameters, each algorithm ran with the same random seed, which forced the same initialisation conditions. Landscape analysis metrics were calculated using the state-of-the-art FLACCO library. The results show the instances for which PSO and GWO share strong similar behaviour. Moreover, we examined the results as to find evidence of structural bias.

Bio: Dr Mario Andrés Muñoz’s research focuses primarily on understanding what makes a problem easy or hard for an optimization or machine learning method, through scientific experimentation, visualization, predictive modelling, and statistical inference techniques. With a keen interest in interdisciplinary work, he has published in fields as diverse as Biomechanics, Power Networks, Resources Engineering, Corporate Social Responsibility, and Computational Biology. He received his PhD from The University of Melbourne in 2014. Prior to joining OPTIMA, he was a Research Fellow at the School of Mathematics and Statistics, The University of Melbourne. He has published over 50 papers, including 20 articles in leading journals, and currently co-supervises five PhD students. He is the main developer of the MATILDA computational engine for Instance Space Analysis.

WED 22 FEBRUARY 16:00 – 17:00 PM AEDT MELBOURNE ONLINE

ZOOM MEETING ID: 873 1557 5255; PASSWORD: 778635