Candidate Set
The Candidate Set report in the Bayesian Optimization platform contains a table of the full set of factor combinations to be considered in the next batch. By default, the number of factor combinations in the table is the minimum of 1000 times the number of factors and 10,000. You can change the default initial number of factor combinations by using the Number of Candidate Set Rows option in the launch window. The candidate set is automatically generated by sampling uniformly over the ranges of the factors, while taking into consideration mixture factors and linear constraint table scripts.
The table is split into two sections: rows to be added to the next batch and rows to be excluded from the next batch. Use the check boxes next to each row to manually select factor combinations to include in the next batch of runs. The rows that are to be added to the next batch are sorted by the order in which they were added. The columns of the table include the values of selected augmented acquisition functions at each factor combination. For a list of possible columns to include in the table, see Candidate Set Report Options.
Note: You can also load a candidate set from an existing data table. See Load Candidate Set from Data Table. If you use this option, the existing candidate set is removed and replaced by the factor combinations in the selected data table. The report title then includes the name of the data table.
The following options are also available in the Candidate Set report:
Autoselect Rows
Automatically selects rows from the candidate set to be included in the next batch of runs. You can specify the number of rows to select and the augmentation method algorithm. The following augmentation methods are available:
Space Filling Exploration
Selects the specified number of points by using the MaxPro space filling criterion. The algorithm focuses on exploring the factor space. This is a model-free method that is useful when the models are not fitting the responses well.
Refine Model
Selects the first point by using the maximum expected improvement criteria. Then, the remaining rows are selected by using the model-based acquisition functions to maximize the Bayesian desirability standard deviation and the multimodel standard deviation criteria. The algorithm balances exploring the factor space and improving the model fit. This method is useful if you want to use the existing model to target areas of sampling improvement instead of random sampling.
Confirm and Challenge Optima
Selects the first run in the batch by finding the factor combination with the largest Bayesian desirability. The remaining runs in the batch are selected by first identifying factor combinations that the Bayesian desirability model indicates are as good as or better than the best run in the training set. From this subset, runs are added to the batch by sequentially maximizing the desirability standard deviation criteria. The algorithm focuses on using the existing model information. This method is useful in the final iterations to directly optimize the model.
Deselect All
Removes all currently selected rows from the current batch.
Undo
Reverses the last change made to the candidate set table.
Redo
Recalls the last change made to the candidate set table.
For additional red triangle menu options, see Candidate Set Report Options.