Implementing `qGIBBON` (`qLowerBoundMaxValueEntropy`) with Ax about ax

[danielcohenlive]

Hi @NeptuneProjects. Thanks for the question. It looks like _construct_inputs_mc_base() is from an earlier version of botorch, so I'm not exactly sure what it does, but I repro'd without it and it looks like qLowerBoundMaxValueEntropy inherits from DiscreteMaxValueBase which actually doesn't take an objective in init, so you need to make sure "objective" is not a key in the dict returned by construct_inputs_qGIBBON(). The next problem I run into though is that the train_inputs arg is required, which may be because I'm not using _construct_inputs_mc_base().

cc @Balandat

[danielcohenlive]

To follow up @NeptuneProjects , can you confirm what is in the inputs_mc dict, and also try calling del inputs_mc["objective"] to see if that solves the problem? It would also be useful to know what version of botorch and Ax you're using.

[Balandat]

+1 to understanding what versions of Ax/BoTorch you're currently using. On trunk, the input constructor for qMaxValueEntropy (which also subclasses DiscreteMaxValueBase should be a good starting point for q qLowerBoundMaxValueEntropy (no need to define a qGibbon class): https://github.com/pytorch/botorch/blob/841392063a5289377cb3ee46c62827003ff2bc07/botorch/acquisition/input_constructors.py#L1022-L1041 - in fact, unless you want to expose some of the optional args, you should be able to just use that verbatim.

[NeptuneProjects]

I have been using Ax 0.3.1 and BoTorch 0.8.3. I should have upgraded before I asked the question - I will do that now, play with your suggestions, and get back to you.