Premature Convergence Problem of Gaussian EDA
Problém předčasné konvergence u Gaussovského EDA
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Gaussovský estimation-of-distribution algoritmus (Gaussovský EDA) je populační optimalizační algoritmus, který pro generování nové populace používá odhadnuté normální rozdělení. Trpí však problémem předčasné konvergence, kdy příliš rychle klesá diverzita populace. V této práci jsem prezentoval možné metody řešení tohoto problému. Metody byly testovány na lineární a elipsoidní účelové funkci, čímž bylo zjištěno, které metody problém předčasné konvergence skutečně řeší. Následně jsem porovnal účinnost metod k řešení účelových funkcí pomocí nástroje ,,Comparing Continuous Optimizers`` (COCO). Také byly vyzkoušeny i slibné kombinace metod. Bylo objeveno, že některé z těchto kombinací dokáží na vybraných příkladech konkurovat i v praxi úspěšnému algoritmu CMA-ES.
Gaussian estimation-of-distribution algorithm (Gaussian EDA) is a population-based optimalization algorithm, which uses estimated normal distribution for sampling a new population. However, it suffers from the premature convergence problem, which means a too rapid decline in population diversity. In this thesis, I have presented possible solutions to this problem. These methods were tested on a linear and an ellipsoid objective function, discovering which methods truly solve the problem. Subsequently, I have compared the efficiency of the methonds in solving objective functions with a benchmarking tool "Comparing Continuous Optimizers" (COCO). Also tested were promising combinations of methods. It was discovered, that some of these combinations were in selected cases able to compete with CMA-ES, a successful algorithm in practice.
Gaussian estimation-of-distribution algorithm (Gaussian EDA) is a population-based optimalization algorithm, which uses estimated normal distribution for sampling a new population. However, it suffers from the premature convergence problem, which means a too rapid decline in population diversity. In this thesis, I have presented possible solutions to this problem. These methods were tested on a linear and an ellipsoid objective function, discovering which methods truly solve the problem. Subsequently, I have compared the efficiency of the methonds in solving objective functions with a benchmarking tool "Comparing Continuous Optimizers" (COCO). Also tested were promising combinations of methods. It was discovered, that some of these combinations were in selected cases able to compete with CMA-ES, a successful algorithm in practice.