Analysis of Sequential Parameter Optimization for Computer Simulation Optimization

Abstract

The main goal of this work is to investigate how the Sequential Parameter Optimization (SPO) framework performs if compared with non-sequential approaches. This is of great importance when dealing with expensive black-box functions as is often the case with regards to simulation optimization. For comparison purposes, well-established metaheuristic methods were applied to classical test functions, which, although being very fast to evaluate, pose challenges for their optimization. The idea was to examine whether SPO would outperform the other methods when a smaller amount of function evaluations was utilized, what could be imposed by budget constraints. Black Hole Optimization, Differential Evolution, and Particle Swarm Optimization were the metaheuristics selected in our study, being all populational methods that may require a large number of iterations in order to achieve good results. The conclusion from the conducted analysis shows that SPO, despite being a more expensive method in the computational standpoint, indeed gets to better results when there is a very limited number of function evaluations available.

@misc{costa2023analysis,
  author = {Costa, Andre N. and Dantas, Joao P. A.},
  title = {{Analysis of Sequential Parameter Optimization for Computer Simulation Optimization}},
  howpublished = {SSRN eLibrary},
  year = {2023},
  doi = {10.2139/ssrn.4328180},
  url = {https://doi.org/10.2139/ssrn.4328180}
}