Towards Better Integration of Surrogate Models and Optimizers
| Autor: | Martin Zaefferer, Tinkle Chugh, Alma A. M. Rahat, Vanessa Volz |
|---|---|
| Přispěvatelé: | Bartz-Beielstein, Thomas, Filipič, Bogdan, Korošec, Peter, Talbi, El-Ghazali |
| Rok vydání: | 2019 |
| Předmět: |
Mathematical optimization
Optimization problem optimisation Computer science media_common.quotation_subject Testbed Evolutionary algorithm evoluutiolaskenta 02 engineering and technology Benchmarking matemaattinen optimointi mathematical optimisation Surrogate model optimointi evolutionary computation Kriging 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Function (engineering) Global optimization media_common |
| Zdroj: | High-Performance Simulation-Based Optimization ISBN: 9783030187637 High-Performance Simulation-Based Optimization |
| Popis: | Surrogate-Assisted Evolutionary Algorithms (SAEAs) have been proven to be very effective in solving (synthetic and real-world) computationally expensive optimization problems with a limited number of function evaluations. The two main components of SAEAs are: the surrogate model and the evolutionary optimizer, both of which use parameters to control their respective behavior. These parameters are likely to interact closely, and hence the exploitation of any such relationships may lead to the design of an enhanced SAEA. In this chapter, as a first step, we focus on Kriging and the Efficient Global Optimization (EGO) framework. We discuss potentially profitable ways of a better integration of model and optimizer. Furthermore, we investigate in depth how different parameters of the model and the optimizer impact optimization results. In particular, we determine whether there are any interactions between these parameters, and how the problem characteristics impact optimization results. In the experimental study, we use the popular Black-Box Optimization Benchmarking (BBOB) testbed. Interestingly, the analysis finds no evidence for significant interactions between model and optimizer parameters, but independently their performance has a significant interaction with the objective function. Based on our results, we make recommendations on how best to configure EGO. peerReviewed |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|