Deep infinite mixture of Gaussian processes (DIM-GP) is a unique combination of neural networks and Gaussian processes that combines the advantages of both modeling techniques, overcoming many disadvantages of each one. It is suitable for a variety of machine learning applications such as regression and classification for scalar quantities. DIM-GP is also suitable for any number of samples as well as any number of input parameters. The number of samples loaded into memory at one time can be controlled by setting the batch size. At very high sample numbers, the prediction time may slow down depending on the batch size setting. DIM-GP is also suitable for noisy data, since an additional noise parameter is trained internally. This can also be used to automatically avoid outliers, for example.
Advanced Settings
You can access the DIM-GP advanced settings in the MOP settings dialog box.
Right-click the MOP node and select from the context menu.
Select the Use advanced settings check box.
In the Advanced Settings tab, under Deep Infinite Mixture Gaussian Process (DIM-GP), select the Use check box.
Switch to the Deep Infinite Mixture Gaussian Process (DIM-GP) Settings tab.
Change the settings as required.
Option Default Value Description Batch size 2000 Activates batch training, if the value is non-zero and smaller than the number of samples. Recommended at higher sample sizes. Job Submit Pattern (beta) <jobscript> <arg1> <arg2> <arg3> Specifies the job submit pattern for bash or command prompt for submitting jobs. Will be cast with
subprocess.popen. First argument<arg1>is for submitting jobs, the second argument<arg2>is for defining a submission pattern that allows a list of arguments to be passed to the job. The third argument<arg3>is for defining the directory.If the last entry does not exist, the action will be skipped and if empty, arguments will be submitted without an optional argument command.
Maximum epochs 500 Maximum number of training epochs used. Algorithm messages provide hints for the required number of epochs. Noisy data On Enable this setting if the training data has noise. Only regression On By default, DIM-GP tries to automatically detect classification task and trains an appropriate model. By selecting this setting, the detection can be disabled and regression is enforced. Power Transformation Off Power transformation can be beneficial for outputs where a large variance change is present for specific input parameter combinations Use optiSLang MOP filtering Off When selected, a model is built for each subspace optiSLang suggests. When cleared, only a single model is trained for the global space. To save and close the dialog, click .