Advanced Use of gpCAM
The advanced use of gpCAM is about communicating domain knowledge in the form of kernel, acquisition and mean functions, and optimization constraints.
Tailored Mean Functions to Communicate Trends
Often times an overall trend of the model is known in absolute terms or in parametric form. In that case, the user may define their own prior mean function following the example below.
Tailored Acquisition Functions for Feature Finding
The acquisition function uses the output of a Gaussian process to steer the experiment or simulation to high-value regions of the search space. You can find an example below.
Tailored Kernel Functions for Hard Constraints on the Posterior Mean
Kernel functions are a tremendously powerful tool to communicate hard constraints to the Gaussian process. Examples include the order of differentiability, periodicity, and symmetry of the model function. The kernel can be defined in the way presented below.
Tailored Cost Functions for Optimizing Data Acquisition when Costs are Present
Cost functions are very useful when the main effort of exploration does not come from the data acquisition itself but from the motion through the search space. gpCAM can use cost and cost update functions. You can find examples for both below. If costs are recorded during data acquisition, gpCAM can use them to update the cost function repeatedly.
This feature is currently in development.