Post by Felipe Chegury Viana on Jul 24, 2008 15:45:17 GMT -5
Dear all,
Here it is a reference on use of multiple surrogates:
F. A. C. Viana and R. T. Haftka, "USING MULTIPLE SURROGATES FOR METAMODELING," in: 7th ASMO-UK/ISSMO International Conference on Engineering Design Optimization, Bath, UK, July 7-8, 2008.
The use of surrogates for facilitating optimization and statistical analysis with computationally expensive simulations has become commonplace. However, most applications employ a single surrogate based on past experience or personal preference. Recently there has been interest in using multiple surrogates rather than a single one. This makes sense because no single surrogate works well for all problems and the cost of constructing multiple surrogates is often small compared to the cost of simulations. This paper surveys recent approaches involving multiple surrogates. This includes (i) fitting multiple surrogates and picking one based on cross-validation errors (PRESS in particular), (ii) using cross validation errors to create a weighted surrogate via minimization of the integrated square error; and (iii) performing optimization multiple times by using multiple surrogates. We conclude that with sufficient number of points (i) PRESS is good for filtering out inaccurate surrogates, and identifying the best surrogate of the set; (ii) the cost of using multiple surrogates is low enough to use in real world applications; and (iii) it pays to use multiple surrogates in optimization.
You can find more about it on:
fchegury.110mb.com/publications.html
All the best,
Felipe
Here it is a reference on use of multiple surrogates:
F. A. C. Viana and R. T. Haftka, "USING MULTIPLE SURROGATES FOR METAMODELING," in: 7th ASMO-UK/ISSMO International Conference on Engineering Design Optimization, Bath, UK, July 7-8, 2008.
The use of surrogates for facilitating optimization and statistical analysis with computationally expensive simulations has become commonplace. However, most applications employ a single surrogate based on past experience or personal preference. Recently there has been interest in using multiple surrogates rather than a single one. This makes sense because no single surrogate works well for all problems and the cost of constructing multiple surrogates is often small compared to the cost of simulations. This paper surveys recent approaches involving multiple surrogates. This includes (i) fitting multiple surrogates and picking one based on cross-validation errors (PRESS in particular), (ii) using cross validation errors to create a weighted surrogate via minimization of the integrated square error; and (iii) performing optimization multiple times by using multiple surrogates. We conclude that with sufficient number of points (i) PRESS is good for filtering out inaccurate surrogates, and identifying the best surrogate of the set; (ii) the cost of using multiple surrogates is low enough to use in real world applications; and (iii) it pays to use multiple surrogates in optimization.
You can find more about it on:
fchegury.110mb.com/publications.html
All the best,
Felipe