For Bayesian optimisation performed by the function "bayesobt", how can I change the covariance function? and how to change its hyper-parameters, i.e. Length scale and signal variance?

For Bayesian optimisation performed by the function "bayesobt", how can I change the covariance function? and how to change its hyper-parameters, i.e. Length scale and signal variance?

Answer 1

[Don Mathis]

You can change the KernelFunction used in the GP that models the objective function. To do that, you will need to edit the source code. Make sure you have write permissions on your MATLAB files, or if using Windows, start MATLAB by right-clicking on it and choosing "Run as administrator".

In MATLAB, type

edit BayesianOptimization.fitObjectiveFcnGP

Then modify the 3 calls to 'iFitrgpRobust' in that function to pass something other than 'ardmatern52' as the 'KernelFunction' argument. Refer to the documentation for 'fitrgp' for options.

The hyperparameters of the GP model are determined during fitting of the model. That's what makes GP models Bayesian.