Keynote Speakers

Featured image

Carola Doerr

Carola Doerr

Sorbonne Université, France

Title: Surfing at the Interface between Theory and Practice in Evolutionary Computation

Abstract: Understanding evolutionary computation approaches by theoretical means is a challenging task. Problems and algorithms that can be rigorously analyzed are typically not representative of those encountered in practice. We will discuss why it is nevertheless rewarding to study EC methods by mathematical means and how research at the interface between theory, benchmarking, and practical applications helps us obtain a much more complete understanding for which algorithmic ideas to favor for which types of problems. We will also discuss how participants can contribute to better exploit the complementarity of the three mentioned approaches.


Shih-Hsi Liu

Shih-Hsi Liu

California State University, Fresno, USA

Title: Is a Comparison of Metaheuristics on Whole Convergence Graphs More Helpful?

Abstract: For the last few decades, hundreds of evolutionary algorithms have been published to solve complex optimization problems, both benchmark functions and real-world problems. However, many of these publications failed to conduct fair algorithm performance comparisons. Thus, the derived conclusions become questionable, which may continue to propagate and jeopardize future works. The first part of this talk is to offer reminders of some key components of an evolutionary algorithm experiment and comparison process that might affect comparison results in different magnitudes. For example, how different comparison methods and stopping criteria may draw different conclusions, slightly or drastically. With such, a whole converge graph, based on the rating results computed by CRS4EAs (a Chess Rating System for Evolutionary Algorithms), will be introduced. The graph visually expresses the ratings of compared algorithms at different stopping criteria. The rating interval figure collectively computed over different stopping criteria could be depicted to express the overall absolute ranking power of an algorithm. We believe a whole convergence graph may offer more helpful insights on algorithm performance and algorithm selection based on allotted computation resources.