当前位置: X-MOL 学术Swarm Evol. Comput. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Population size in Particle Swarm Optimization
Swarm and Evolutionary Computation ( IF 8.2 ) Pub Date : 2020-05-27 , DOI: 10.1016/j.swevo.2020.100718
Adam P. Piotrowski , Jaroslaw J. Napiorkowski , Agnieszka E. Piotrowska

Particle Swarm Optimization (PSO) is among the most universally applied population-based metaheuristic optimization algorithms. PSO has been successfully used in various scientific fields, ranging from humanities, engineering, chemistry, medicine, to advanced physics. Since its introduction in 1995, the method has been widely investigated, which led to the development of hundreds of PSO versions and numerous theoretical and empirical findings on their convergence and parameterization. However, so far there is no detailed study on the proper choice of PSO swarm size, although it is widely known that population size crucially affects the performance of metaheuristics. In most applications, authors follow the initial suggestion from 1995 and restrict the population size to 20–50 particles. In this study, we relate the performance of eight PSO variants to swarm sizes that range from 3 up to 1000 particles. Tests are performed on sixty 10- to 100-dimensional scalable benchmarks and twenty-two 1- to 216-dimensional real-world problems. Although results do differ for the specific PSO variants, for the majority of considered PSO algorithms the best performance is obtained with swarms composed of 70–500 particles, indicating that the classical choice is often too small. Larger swarms frequently improve efficiency of the method for more difficult problems and practical applications. For unimodal problems slightly lower swarm sizes are recommended for the majority of PSO variants, but some would still perform best with hundreds of particles.



中文翻译:

粒子群优化中的种群大小

粒子群优化(PSO)是应用最广泛的基于种群的元启发式优化算法之一。PSO已成功地应用于从人文,工程,化学,医学到高级物理学的各种科学领域。自1995年推出以来,该方法已得到广泛研究,从而导致开发了数百种PSO版本,并且在其收敛性和参数化方面有许多理论和经验发现。然而,尽管众所周知,种群大小对元启发式算法的性能有至关重要的影响,但到目前为止,还没有关于PSO种群大小的正确选择的详细研究。在大多数应用中,作者遵循1995年的最初建议并将种群大小限制为20-50个粒子。在这个研究中,我们将8种PSO变体的性能与3到1000个粒子的群大小相关联。测试在60个10到100维的可伸缩基准和22个1到216维的实际问题中进行。尽管对于特定的PSO变体,结果的确有所不同,但是对于大多数考虑的PSO算法,使用由70–500个粒子组成的群集可获得最佳性能,这表明经典选择通常太小。对于更困难的问题和实际应用,较大的群体通常会提高方法的效率。对于单峰问题,对于大多数PSO变体,建议使用较小的群大小,但是某些变体在处理数百个粒子时仍然表现最佳。测试是在60个10到100维的可伸缩基准和22个1到216维的实际问题上进行的。尽管对于特定的PSO变体,结果的确有所不同,但是对于大多数考虑的PSO算法,使用由70–500个粒子组成的群集可获得最佳性能,这表明经典选择通常太小。对于更困难的问题和实际应用,较大的群体通常会提高方法的效率。对于单峰问题,对于大多数PSO变体,建议使用较小的群大小,但是某些变体在处理数百个粒子时仍然表现最佳。测试在60个10到100维的可伸缩基准和22个1到216维的实际问题中进行。尽管对于特定的PSO变体,结果的确有所不同,但是对于大多数考虑的PSO算法,使用由70–500个粒子组成的群集可获得最佳性能,这表明经典选择通常太小。对于更困难的问题和实际应用,较大的群体通常会提高方法的效率。对于单峰问题,对于大多数PSO变体,建议使用较小的群大小,但是某些变体在处理数百个粒子时仍然表现最佳。对于大多数已考虑的PSO算法,使用由70-500个粒子组成的群集可获得最佳性能,这表明经典选择通常太小。对于更困难的问题和实际应用,较大的群体通常会提高方法的效率。对于单峰问题,对于大多数PSO变体,建议使用较小的群大小,但是某些变体在处理数百个粒子时仍然表现最佳。对于大多数已考虑的PSO算法,使用由70–500个粒子组成的群集可获得最佳性能,这表明经典选择通常太小。对于更困难的问题和实际应用,较大的群体通常会提高方法的效率。对于单峰问题,对于大多数PSO变体,建议使用较小的群大小,但是某些变体在处理数百个粒子时仍然表现最佳。

更新日期:2020-05-27
down
wechat
bug