Bayesian Optimization Batch Customizer
The Bayesian Optimization Batch Customize report enables you to customize the selection of new experimental runs. The report contains two prediction profilers and a set of options for optimizing the factor settings under various conditions. This report enables you to use the Gaussian process models that are fit to each response to evaluate acquisition functions and manually select iterative sample points to be included in the next batch of experiments (or experimental runs). You can use the profiler shortcut options to find factor combinations for different optimization goals.
Add Current Profiler Settings to Batch
Adds the current profiler factor settings to the candidate set and includes them as a run in the current augmentation batch.
Tip: Use this option after updating the profilers manually or by using the Profiler Shortcuts buttons.
Profiler Shortcuts
Provides options for finding the optimal factor settings under various conditions. Each option corresponds to one of the possible responses in the Augmented Acquisition Functions Profiler. All of the acquisition functions are based on the augmented model. Use the Add Current Profiler Settings to Batch button to add an identified run to the current batch of runs.
Maximize Bayesian Desirability
Finds the factor settings that maximize the posterior mean of the joint desirability function. This option finds the best settings given the current model and observed responses. Use this option to find the optimal settings that are suggested by the model at the end of the last iteration.
Maximize Bayesian Desirability Std Dev
Finds the factor settings that maximize the posterior standard deviation of the desirability distribution that incorporates the training data and the current batch. This option explores the factor space sequentially. Use this option when the batch size is larger than one and the models fit the responses well.
Maximize Multimodel Std Dev
Finds the factor settings that maximize the overall prediction standard deviation for all of the models. This option is a model-based exploration step that results in factor settings that minimize the worst-case model prediction uncertainty at the next iteration. Use this option when the batch size is larger than one and you want to explore the factor space in such a way that points at the extremes of the factor ranges are included.
Maximize MaxPro Criterion
Finds factor settings that maximize the Maximum Projection (MaxPro) criterion. This option is a model-free exploration of the factor space that avoids replication of any of the factor settings in both the training data and the current batch. Use this option for any batch size when one or more of the models is not fitting well.
Maximize Expected Improvement
Finds the factor settings with the largest expected improvement of the Bayesian desirability model that uses both the training data and the current batch. This option balances exploring the factor space and using information from the current model. Use this option for any batch size once the model fits are above a minimum acceptable model goodness-of-fit threshold. The automated algorithms in the platforms use a threshold of 0.25 by default.
Tip: It is recommended to use only one Maximum Expected Improvement run per batch.
Maximize Upper Confidence Bound
Finds the factor settings with the largest upper confidence bound of the Bayesian desirability model that uses both the training data and the current batch. This option balances exploring the factor space and using information from the current models. Use this option for any batch size once the model fits are above a minimum acceptable model goodness-of-fit threshold. The automated algorithms in the platforms use a threshold of 0.25 by default. This is often called the UCB criterion.
Restore Best Training Point
Returns factor settings to those of the training row with the highest observed desirability. In the initial report, these factor settings are shown in the Augmented Acquisitions Functions Profiler. Use this option after any changes are made to the profiler to return to the settings of the best run that was observed in the training set.
Augmented Acquisition Functions Profiler
This section contains a prediction profiler for the augmented acquisition functions. By default, only the Augmented Bayesian Desirability response is shown. This is the mean of the posterior distribution of the desirability of the Gaussian Process models. The intervals around the mean take into consideration the model that is fit to the training data as well as the reduction in variability that would result after adding the runs in the current batch. As you select options in the Profiler Shortcuts section, the response that is shown on the profiler is updated. To see the profilers for all augmented acquisition functions, use the Toggle Acquisition Functions option. For more information about the prediction profiler, see “Profiler” in Profilers.
Augmented Prediction Profiler
This section contains a prediction profiler for each response variable. To view the standard deviations, use the Toggle Prediction Standard Deviations option. For more information about the prediction profiler, see “Profiler” in Profilers.
Note: The Augmented Acquisition Functions Profiler and the Augmented Prediction Profiler are linked. Changes in the factor settings in one profiler are also applied to the other profiler.
For additional red triangle menu options, see Bayesian Optimization Batch Customizer Report Options.