Specifications

In the Specifications step, select methods and define method settings. Once these have been identified, click Apply to accept the changes.

The methods available will depend upon the approach being used, and the method settings available will depend upon the method selected. For more information on settings specific to the different methods, refer to Learn the Concepts.

Modify run data or add additional runs in the Run matrix.

Method	Parameter Screening	Space Filling	Custom	Variable Levels	Continuous	Discrete	Categorical	Basic Parameters	Properties and Comments
Modified Extensible Lattice Sequence				Any				You can either accept the default number of runs or enter a different value.	Use this method when the response surface is highly nonlinear. This method is a better space filler than Latin Hypercube. The default number of runs is 1.1((N+1)(N+2))/2, where N is the number of input variables.
D-Optimal				Any				You can either accept the default number of runs or enter a different value. You can also select the appropriate regression model.	Use this method when the known goal is to build a regression. This method is also useful when corner coverage is important, and you have problems with input variable constraints.
Fractional Factorial				2 or 3				Select the appropriate resolution.	Resolution indicates the level of accuracy of the interactions. Interactions should not be used with Resolution III.
Full Factorial				Any					Requires a high number of simulations and is therefore unsuitable for most studies. Total number of runs should be less than 1,000,000.
Plackett Burman				2				You can either click Apply for AutoSelect or select a table using the Design pull-down menu.	Computationally least expensive. Number of points can be 12, 20, 24, 28 or 36. Selecting Autoselect will pick pbdgn12 if N < 12, where N is the number of input variables; pbdgn20 if 12 <= N < 20, etc. Limited to 35 input variables. Categorical variables must have exactly two levels.
Central Composite Design				5					Use this method when the output responses are known to be quadratic. Limited to 20 input variables.
Box-Behnken				3				You can either click Apply for AutoSelect or select a table using the Design pull-down menu.	Use this method for building quadratic response surfaces if the output responses are known to be quadratic and predictions are not required at the edge of the design space. Number of points can be 13, 25, 41, 49. 57. Selecting Autoselect will pick bbdgn13 if N < 4, where N is the number of input variables; bbdgn25 if N = 4, bbdgn41 if N = 5, etc. Limited to 7 input variables. Discrete variable must have at least 3 levels. Categorical variables must have exactly 3 levels.
Latin Hypercube				Any				You can either accept the default number of runs or enter a different value.	Use this method when the response surface is highly nonlinear. The default number of runs is 1.1((N+1)(N+2))/2, where N is the number of input variables. You must maintain the value of the random seed in order to get repeatable designs.
Hammersley				Any				You can either accept the default number of runs or enter a different value.	Use this method when the response surface is highly nonlinear. This method is a better space filler than Latin Hypercube. The default number of runs is 1.1((N+1)(N+2))/2, where N is the number of input variables.
Taguchi				Varies				You can either choose AutoSelect or a specific design matrix.	The levels of each variable must be set accordingly to ensure compability with a specific design matrix.
User-defined Design				Any				Select the perturb file.	Use this method to create a design matrix using abstract variable levels.
Run Matrix				Any				Select the perturb file.	Use this method to create a design matrix using literal variable values.
None

Filter outliers, duplicate numbers, and bad numbers from run data by selecting the appropriate checkboxes in the Filter tab. For more information on outliers, refer to Distribution - Box Plot.

If you are running a Least Square Regression Fit, select terms to be used in the regression equations in the Regression Terms tab.

Method	Equation Available	Highly Nonlinear	Noisy Output Responses	Accuracy	Efficiency	Basic Parameters	Comments
Least Square Regression						Order should be adjusted and regression terms may be picked.	Noises can be screened out with this method.
Moving Least Squares Method						Order should be adjusted.	The time to build the fit and use the fit (Evaluate From) increases with both the number of runs and the number of input variables in the input matrix. The number of input variables has more influence than the number of runs if order is larger than 1.
HyperKriging							The time to build the fit and use the fit (Evaluate From) increases with both the number of runs and the number of input variables in the input matrix. The number of input variables has more influence than the number of runs if order is larger than 1.
Radial Basis Function							The time to build the fit increases with both the number of runs and the number of input variables in the input matrix. The number of runs has more influence than the number of input variables. The run time for using the fit in another approach (Evaluate From) is very small regardless of the size of the input matrix.

Only the methods that are valid for the problem formulation are enabled and can be selected.

For RRBDO methods (SORA, SORA-ARSM, and SLA) to be enabled, the definition of random properties for one or more input variables and at least one constraint must be random. Robustness is activated by setting Robust Optimization to ON in the Specification step. For a robust minimization problem, SORA minimizes the Robust Min percentile value (95% as default value). For a robust maximization problem, SORA maximizes the Robust Max percentile value (5% as default value).

Method

Continuous

Discrete

Linear

Nonlinear

Single Objective

Multi Objective

Deterministic

Probabilistic

Accuracy

Efficiency

Global

Comments

ARSM

Default method for single objective problems.

GRSM

SOO

GRSM is the default method for multi objective problems and it is also the preferred method when the number of input variables is large. It can start optimizing with just a few numbers of points independent of the number of input variables.

Method	Efficiency	Basic Parameters	Comments
Simple Random		Number of runs	Maintain the value of the random seed to get repeatable designs.
Latin Hypercube		Number of runs	Maintain the value of the random seed to get repeatable designs.
Modified Extensible Lattice Sequence		Number of runs	Maintain the value of the random seed to get repeatable designs.
Hammersley		Number of runs	Similar but slightly different statistical properties than Latin Hypercube.

Specifications

Specifications

Specifications

Modified Extensible Lattice Sequence

D-Optimal

Fractional Factorial

Full Factorial

Plackett Burman

Central Composite Design

Box-Behnken

Latin Hypercube

Hammersley

Taguchi

User-defined Design

Run Matrix

None

Least Square Regression

Moving Least Squares Method

HyperKriging

Radial Basis Function

ARSM

GRSM

SOO

MOO

SQP

MFD

GA

MOGA

SORA

SORA-ARSM

SLA

Xopt

Simple Random

Latin Hypercube

Modified Extensible Lattice Sequence

Hammersley

See Also: