Search

In the case of time-consuming simulation models or other so-called black-box functions, we determine a metamodel which approximates the relation between the input- and output-variables of the simulation model. To solve multi-objective optimization problems, we approximate the Pareto set, i.e. the set of Pareto optimal solutions for which it is not possible to improve one objective without deteriorating another.

In the field of design of computer experiments (DoCE), Latin hypercube designs are frequently used for the approximation and optimization of black-boxes. In certain situations, we need a special type of designs consisting of two separate designs, one being a subset of the other. These nested designs can be used to deal with training and test sets, models with different levels of accuracy, linking parameters, and sequential evaluations. In this paper, we construct nested maximin Latin hypercube designs for up to ten dimensions. We show that different types of grids should be considered when constructing nested designs and discuss how to determine which grid to use for a specific application. To determine nested maximin designs for dimensions higher than two, four different variants of the ESE-algorithm of Jin et al. (2005) are introduced and compared. In the appendix, maximin distances for different numbers of points are provided; the corresponding nested maximin designs can be found on the website http://www.spacefillingdesigns.nl.

In many fields, we come across problems where we want to optimize several conflicting objectives simultaneously. To find a good solution for such multi-objective optimization problems, an approximation of the Pareto set is often generated. In this paper, we con- sider the approximation of Pareto sets for problems with three or more convex objectives and with convex constraints. For these problems, sandwich algorithms can be used to de- termine an inner and outer approximation between which the Pareto set is 'sandwiched'. Using these two approximations, we can calculate an upper bound on the approximation error. This upper bound can be used to determine which parts of the approximations must be improved and to provide a quality guarantee to the decision maker. In this paper, we extend higher dimensional sandwich algorithms in three different ways. Firstly, we introduce the new concept of adding dummy points to the inner approx- imation of a Pareto set. By using these dummy points, we can determine accurate inner and outer approximations more e±ciently, i.e., using less time-consuming optimizations. Secondly, we introduce a new method for the calculation of an error measure which is easy to interpret. The combination of easy calculation and easy interpretation makes this measure very suitable for sandwich algorithms. Thirdly, we show how transforming cer- tain objective functions can improve the results of sandwich algorithms and extend their applicability to certain non-convex problems. The calculation of the introduced error measure when using transformations will also be discussed. To show the effect of these enhancements, we make a numerical comparison using four test cases, including a four-dimensional case from the field of intensity-modulated radiation therapy (IMRT). The results of the different cases show that we can indeed achieve an accurate approximation using significantly fewer optimizations by using the enhancements.

In the area of computer simulation, Latin hypercube designs play an important role. In this paper the classes of maximin and Audze-Eglais Latin hypercube designs are considered. Up to now only several two-dimensional designs and a few higher dimensional designs for these classes have been published. Using periodic designs and the Enhanced Stochastic Evolutionary algorithm of Jin et al. (2005), we obtain new results which we compare to existing results. We thus construct a database of approximate maximin and Audze-Eglais Latin hypercube designs for up to ten dimensions and for up to 300 design points. All these designs can be downloaded from the website http://www.spacefillingdesigns.nl.

When building a Kriging model, the general intuition is that using more data will always result in a better model. However, we show that when we have a large non-uniform dataset, using a uniform subset can have several advantages. Reducing the time necessary to fit the model, avoiding numerical inaccuracies and improving the robustness with respect to errors in the output data are some aspects which can be improved by using a uniform subset. We furthermore describe several new and current methods for selecting a uniform subset. These methods are tested and compared on several artificial datasets and one real life dataset. The comparison shows how the selected subsets affect different aspects of the resulting Kriging model. As none of the subset selection methods performs best on all criteria, the best method to choose depends on how the different aspects are valued. The comparison made in this paper can be used to facilitate the user in making a good choice.

Latin hypercube designs (LHDs) play an important role when approximating computer simula- tion models. To obtain good space-filling properties, the maximin criterion is frequently used. Unfortunately, constructing maximin LHDs can be quite time-consuming when the number of dimensions and design points increase. In these cases, we can use approximate maximin LHDs. In this paper, we construct bounds for the separation distance of certain classes of maximin LHDs. These bounds are useful for assessing the quality of approximate maximin LHDs. Until now only upper bounds are known for the separation distance of certain classes of unrestricted maximin designs, i.e. for maximin designs without a Latin hypercube struc- ture. The separation distance of maximin LHDs also satisfies these “unrestricted” bounds. By using some of the special properties of LHDs, we are able to find new and tighter bounds for maximin LHDs. Within the different methods used to determine the upper bounds, a vari- ety of combinatorial optimization techniques are employed. Mixed Integer Programming, the Travelling Salesman Problem, and the Graph Covering Problem are among the formulations used to obtain the bounds. Besides these bounds, also a construction method is described for generating LHDs that meet Baer’s bound for the ℓ1 distance measure for certain values of n.

The subject of this paper is a new approach to Symbolic Regression.Other publications on Symbolic Regression use Genetic Programming.This paper describes an alternative method based on Pareto Simulated Annealing.Our method is based on linear regression for the estimation of constants.Interval arithmetic is applied to ensure the consistency of a model.In order to prevent over-fitting, we merit a model not only on predictions in the data points, but also on the complexity of a model.For the complexity we introduce a new measure.We compare our new method with the Kriging meta-model and against a Symbolic Regression meta-model based on Genetic Programming.We conclude that Pareto Simulated Annealing based Symbolic Regression is very competitive compared to the other meta-model approaches

In the area of computer simulation Latin hypercube designs play an important role.In this paper the class of maximin Latin hypercube designs is considered.Up to now only several two-dimensional designs and designs for some small number of points are known for this class.Using periodic designs and simulated annealing we extend the known results and construct approximate maximin Latin hypercube designs for up to ten dimensions and for up to 100 design points.All these designs can be downloaded from the website http://www.spacefillingdesigns.nl