More space-filling sampling method than tf.uniform

Describe the feature you’d like

In the Box class the sampling method is tf.random.uniform. I found it less space-filling than for example tf.math.sobol_sample. This will make the metamodel better approximate the function at the initial phase of optimization. it might be good if the user can choose what kind of sampler to use in the Box class. as a comparison, botorch had a set of samplers we can select (link) and in the GPflowOpt we have Design class (link) .

Additional context to give some context what’s available there: link

[satrialoka]

yes. additionally, it might be good to have that sampling customization for generating initial data before the optimization iteration.

[joelberkeley]

I assume you’re particularly looking at how the space is used in the acquisition optimizer. If so, it may also be good to let users customize the acquisition optimizer (which would indirectly allow users to customize the sampling in the optimizer, depending on how that’s done). This is something we intend to do anyway