Improved Restart Strategies for Unconstrained Radial Basis Function Optimization Method
We develop a new algorithm Surrogate Optimization with Partial Restart Initiation (SOPRI) for unconstrained surrogate-based optimization of high-dimensional expensive black-box functions that employs new restart strategies, based on the DYCORS algorithm. In prior work, the search method restarts completely from scratch, meaning that all previously evaluated values of the objective f(x) are discarded in the hope that a new and better minimum point will be found after the restart. However, these previously evaluated points provide valuable information about the black-box function. This information is of great importance since the objective function values had been obtained at great computational cost before restart was implemented. Thus, we introduce two strategies that use some subsets of the previously evaluated points during the restart. Our first approach will keep the percentage of the worst evaluated points (the percentage points in (x, f(x)) with the highest functioncvalue for minimization problems) in previous search steps and include these points in the surrogate after restart. This helps the algorithm spend less time in the vicinities of function values discovered to be bad before restart. Another strategy considered is to remove the previously evaluated best point and points around the best point. This strategy aims to encourage searching after restart in promising areas without immediately moving to the best decision vector previously found. Combinations of these novel approaches have been numerically compared with the radial basis function (RBF) global optimization algorithm DYnamic COordinate search using Response Surface (DYCORS) models and other optimization models on 21 test problems with 20 dimensions. Experimental results indicate that the new restart strategies find better objective function values using fewer objective function evaluations. The ideas used here for restart with DYCORS surrogate algorithms can potentially benefit other surrogate global optimization algorithm
Year of publication: |
2023
|
---|---|
Authors: | Liu, Nanxi ; Shoemaker, Christine |
Publisher: |
[S.l.] : SSRN |
Saved in:
freely available
Saved in favorites
Similar items by person
-
The Potential Impact of Cotton Insect Control Technology
LaDue, Eddy L., (1979)
-
Müller, Juliane, (2014)
-
Regis, Rommel, (2013)
- More ...