On the Impact of the Large Population on Evolutionary Algorithm

• Published in: 2024 20th International Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD) pp. 1 - 10
• Main Authors: Jin, Chen, Liu, Gang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.07.2024
• Subjects: Closed box; Distance measurement; EAs; Evolutionary Algorithms; Evolutionary computation; Fuzzy systems; GPU Computing; High-Performance Computing; HPC; Iron; Knowledge discovery; Large Popu-lation Size; Optimization; Parallel EAs; Parallel processing; Runtime; Search problems
• DOI: 10.1109/ICNC-FSKD64080.2024.10702201

As population-based searching techniques, Evolutionary algorithms (EAs) are ideal for solving real-world or black-box optimization problems. Due to the stochastic nature, there is a common viewpoint that a large population benefits solving difficult or complex problems if given enough runtime. To figure out the necessity of large populations, many theoretical works on sizing population have been proposed since 1990s regardless of the computing efforts. However, an interesting phenomenon is observed that two opposite views exist meanwhile. Most theoret-ical work admitted the benefits of a large population on solving difficult problems with a sufficient computing budget. While a few works claimed that a large population is unhelpful and it is only a waste of computing efforts. Therefore, a comprehensively experimental study that practically examines the benefits of a large population is necessary but still lacking due to the backward computing facilities in the past. However, benefiting from the modern parallel computing platforms, the computing capability is no longer the bottleneck and processing a great number of fitness evaluations (FEs) in a reason time is now practical. In this work, we systemically investigate the performance of EAs with a large population in terms of solution quality and computing speed on a GPU-enabled parallel computing platform. In order to comprehensively evaluate the impacts of a large population, two state-of-the-art and three generic algorithms are tested with population sizes ranging from small (64) to large (4096) on eight difficult problems for three dimensions. Results illustrate that not only EAs with a large population can achieve solutions with better quality but also run faster on a GPU-enabled parallel computing platform.