Simple TS implementation

[taikichi-pc]

Thank you for your previous kind reply.

I'd like to implement TS(Thompson sampling) baseline for comparing with other acquisition function.

How would you implement TS using botorch as a simple(basic) code?

I've already read bellow links.
https://botorch.org/tutorials/turbo_1
https://botorch.org/tutorials/thompson_sampling

BTW COMBO, another BO library, and so on adopt TS as a default algorithm for Multi-objective.
I think It's classical and good(especially in terms of calc. cost) way and not complicated.
But I look forward to refer your code.

Thank you.

[Balandat]

Take a look at MaxPosteriorSampling - this is Thompson sampling. The support for batched models makes it a bit less readable and there a some tensor acrobatics necessary if you want to avoid duplicated samples in that setting, but overall it's pretty straightforward.

Note that this draws samples at some discrete set of points X. Another implementation would be to draw sample paths from the model posterior (using the pathwise sampling feature added in #1463), and then optimize each of them individually using some gradient-based scheme. This is currently not (yet) implemented though.