This node allows you to open a optiSLang monitoring database file (*.omdb) and modify data and filter settings.

There are settings that define the method to deal with the given data in order to train and test the possible approximation models and there are settings that define the method to find the optimal subspace for every output parameter. You can deselect all filters, then the algorithm will search for the optimal metamodel using the full parameter model.

You can export a MOP as a Functional Mock-Up Unit for use in a variety of simulation environments.

Further information about methods of metamodelling used in optiSLang can be found here.

To access the following tabs and options, double-click MOP on the Scenery pane.

Basic Settings

The MOP dialog provides a decision tree which helps to find a suitable model with good and proofed settings. Some information about the metamodels to test and filtering of variables is required.

Advanced Settings

If you select the Use advanced settings check box, a property tree is displayed with all of the available settings.

Option	Description
Testing type	Select the method of the test point determination.
Approximation type	Select either smoothing models (Polynomials, MLS with exponential weights or Kriging) or use interpolating models (MLS with regularized weights).
Resampling for single CoPs	The single Coefficient of Prognosis (CoPs) are estimated by varianced based sensitivity analysis on the final metamodel. If you select Empirical distribution from data, the marginal distributions are generated by using the empirical distribution functions. Otherwise uniform distributions within the lower and upper bounds are assumed.
Max. responses in parallel	Sets the maximum number of responses to be calculated in parallel.
Use incomplete designs	When selected, designs with missed but at least one calculated responses are also used for creating the metamodel.
Export FMU after MOP creation	When selected, exports the MOP as a Functional Mock-Up Unit.
Adapt bounds	Adapts the boundaries of the resulting parameters to the sampling. An additional percentage factor of safety for the adapted boundaries can be added. The new boundaries are calculated by:
CoP tolerance	Specifies the range for the CoP values in which a simpler model should be selected. The variable tolerance is used to automatically remove all variables below this tolerance from the best model. If Use adjusted CoP is selected, the number of model coefficients is considered to penalize the CoP in order to prefer simpler models with similar prognosis quality.
Transformation	Enables or disables Box Cox transformations.
Models
Select which metamodel types are tested. At least one model type has to be selected. Polynomial regression with linear and quadratic basis functions (order 1 and 2) and Moving Least Squares with linear and quadratic basis functions are available as well as ordinary Kriging with isotropic and anisotropic kernel functions. By means of the coefficient factor, the minimum number of samples for a specific model is defined. If the number of model coefficients times the coefficient factor exceeds the available number of samples, than this model candidate is not considered. An extra feature is using external models. Currently ASCMO is available on Windows platforms. This model can only be selected if Testing type is set to Leave one out. Custom Script Surrogates are supported.
Filter
Significance Filter	Select the minimal quantile for the significance filters. This means that all possible significance filter configurations, defined by quantiles which are varied from 99% down to this minimal value, will be tested.
Importance Filter	Define an upper bound for the COI limits. The threshold minimal COI value is varied from 1% up to this bound and that defines another filter criterion for the input parameters.
Correlation Filter	Choose the number of steps in which the correlation coefficients of the remaining input parameters (after the significance filter) are divided. Each of the correlation coefficients determined that way defines a lower bound for a correlation filter.
CoD Filter	Define an upper bound for the CoD limit. The threshold minimal CoD value is varied from 1% up to this bound and that defines another filter criterion for the input parameters.
Input Correlations Filter	The input correlation filter removes strongly correlated input variables from the parameter subspace before the other filters are applied. Only if the Check value is exceeded by at least one input correlation, this filter is executed. Only correlations larger as the Minimum value are considered. The filter removes all variables having a multi-dimensional correlation larger than the Maximum value.
Algorithm messages	Show algorithm messages in project log or write to a separate file including the function strings for polynomials.
Crossvalidation values	Define the usage and naming of crossvalidated values.

Show Postprocessing

When selected, opens postprocessing.

Inputs

For parameters, you can specify the importance manually when either the Variable reduction setting is set to User defined or the Use advanced settings check box is selected. Possible values are:

Outputs

For responses, you can specify whether or not to generate a metamodel. Select or clear the Use check box.

Additional Options

To access the options shown in the following table, in any tab, click Show additional options.

Slots

Slot Name		Slot type		Data type	Description
		In	Out
MDBPath		X			MOP result file of type *_MOP.bin, *.omdb, *.ascmo.
Designs		X			Used support points of MOP.
ParameterManager		X			All parameters that are used for MOP with corresponding properties.
ShowPPWhenAvailable		X			Show Postprocessing? yes/no
MDBPath			X		The MOP results file which can be used for MOPsolver.
CoDadj			X		The adjusted coefficient of determination of all responses.
COP			X		The coefficient of prognosis of all responses (predictive coefficient of determination for Leave one out).
FoundModel4Response			X		The information for each response if a model was found.
ParameterManager			X		Filtered parameters found by MOP.
SingleCOP			X		The coefficient of prognosis for single inputs (predictive coefficient of determination for Leave one out).

Available Models for Scalar Outputs

For scalar outputs, the following models are available within the MOP competition:

The polynomial model is the simplest but the fastest available model. It can be applied for small, medium, and even large number of samples. The approximation of the model with the MOP solver and within the exported Functional Mock-Up Unit (FMU) is also faster than all other available models. Moving Least Squares and Kriging are more time consuming for the model training, especially the anisotropic Kriging, which limits an efficient application up to 1000 samples. You can export these modules as FMU and are fast in the approximation. MOP filtering is applied to reduced the active variables to the important ones.

Genetic Aggregation Response Surface (GARS) and Support Vector Regression (SVR) are metamodels from DesignXplorer. GARS is similar to the anisotropic Kriging limited to smaller data sets due to exponential increasing training time. Support Vector Regression (SVR) can be applied even for larger data sets. Both models use the MOP filtering for variable selection. You currently cannot export FMU for these models.

The Deep Feed Forward Network (DFFN) is a deep learning neural network with automatic feature and variable filtering (smart layout). The training of this model becomes efficient for large data sets with more than 1000 samples, where the training of other models such as MLS, Kriging, or GARS becomes inefficient.

Deep Infinite Mixture Gaussian Process (DIM-GP) is a further development of the Kriging approach, using a more flexible covariance matrix description represented by a neural network approximation. The DIM-GP model can be efficiently trained up to 2000 samples. Currently, variable filtering is not available within the MOP node for the DIM-GP model. Similar to GARS and SVR,you cannot export the DIM-GP and the DFFN models as FMU.

The following table describes the properties and a suggested application for each model.

Looking at possible use cases, you can distinguish between the following

The following table describes which models should be used for each use case, and provides suggestions for the number of training samples.