Optimizing over discrete parameter domains

[avimit]

Can BoTorch be used over discrete parameter domains?
(if so, than this is a feature inquiry, and not a feature "request")

We have a use case of domains which are partly continuous, partly discrete, like:
[{"name": "param1", "type": "continuous", "domain": [-5, 10]}, {"name": "param2", "type": "continuous", "domain": [1, 15]}, {"name": "param3", "type": "discrete", "domain": [1, 1.5, 2, 2.5, 3, 3.5, 4]}]

The functions under "botorch/optim/optimize.py" accept an argument called "bounds", which you define as : "bounds: A 2 x d tensor of lower and upper bounds for each column of X".

These are obviously bounds for a continuous search space. Can BoTorch be used for searching over discrete spaces?

Thank you so much for the package!
Avi

[Balandat]

Hi @avimit, BoTorch itself at this point doesn't directly deal with discrete parameters, but If your goal is simply to run BayesOpt (rather than developing new BayesOpt primitives), you can easily do this using the higher-level Ax interface. Ax will perform appropriate transformations (e.g. rounding for ordinal parameters, or one-hot encoding for categorical ones) when talking to BoTorch under the hood.

We are also working on improving support for discrete parameters directly in BoTorch (e.g. by using a concrete relaxation to still enable gradient-based optimization).

Finally, if you're willing to write your own Kernels or even models that comply with our basic API, you can do anything you like directly with discrete parameters and still use BoTorch's acquisition functions. Note, however, that optimizing over mixed domains will then also require using a different optimization algorithm, i.e., you can't use the functions in optim/optimize.py directly.

Hope this helps!

[avimit]

Hi @Balandat,

Thank you for the answer!

I will have a look at Ax, but I feel that we are in need of the more "low level" character of BoTorch.

Explanation: we have have our own small platform for distributing & managing jobs (mostly DL training) on the cloud, and on top of it we sometimes run HPO. Our HPO framework treats the training of models as a "black box" function, launching train jobs as resources are freed, and collecting evaluations when ready. Our use case for BoTorch is as a supplier of next points for evaluation; I managed to successfully integrate it (and enjoyed the process!) as long as I don't use discrete parameters.

Thanks,
Avi

[Balandat]

Glad you were able to use BoTorch directly for this.

That said, Ax is designed for exactly your use case, and provides a number of different APIs (including lower-level ask/tell style ones) that should make that integration quite straightforward. You can also stick your own BoTorch components into these APIs if you want to retain the flexibility of customizing your own models/acquisition functions. See the tutorial on this.

As I mentioned, we are working on improving support for discrete parameters directly in Botorch. For now, what you could do is perform the parameter transformations that Ax does in this module yourself directly in torch.

[avimit]

I will sure give it a try!
Thanks a lot,
Avi

[eytan]

@avimit , based on your description, I think that Ax is the appropriate tool to plug into. For this ordered, categorical parameter (param3), you'll want to use a ChoiceParameter (as described here) with the keyword argument is_ordered=True. Ax will take care of all the necessary encoding, normalization of the search space and outcomes, etc.

Your setup sounds kind of like what one might do with Slurm or Ray Tune. We have forthcoming interfaces between Ax and these platforms, and they were quite straightforward to implement.

[eytan]

You might want to look at the source code for the Ray Tune interface as an example.

[avimit]

@eytan, ChoiceParameter does look like what I need, and the is_ordered switch covers all options; the example in Ray Tune in indeed instructive. Thanks.

Taking a parameter transformation from Ax transforms is also a viable option, @Balandat; thank you too.

Now it's time for me to do the work ;}

[avimit]

Can you please direct me to a function in Ax which returns next candidates for evaluation, similar to joint_optimize() in BoTorch?

(I have been doing something like this:

model = botorch.models.SingleTaskGP(X, Y)
mll = gpytorch.mlls.ExactMarginalLogLikelihood(model.likelihood, model)
botorch.fit.fit_gpytorch_model(mll)

acquisition_function = botorch.acquisition.qNoisyExpectedImprovement(model, X_baseline)
X_candidates_tensor = botorch.optim.joint_optimize(acquisition_function, bounds=bounds, 
                                                                   q=batch_size, num_restarts=1, raw_samples=len(X))

)

Thanks!

[Balandat]

If you use Ax's service api see API examples here, get_next_trial does what you want.

You can also use the developer api (next tab on the above page) - you can access the generated candidates ("arms") by accessing th arms property of the GeneratorRun returned by the gen function.

[avimit]

Thanks! I was there, but never noticed the tabs...

Aren't you supposed to be resting on 4th of July? ;)

[avimit]

Hi,

I must admit that I find implementing Ax for my task much harder compared to BoTorch.

The five schematic BoTorch lines I quote above do what I want: supply X new candidates for evaluation, given the known+pending evaluations.

Haven't yet found a way to do it with ax....
I've been trying to use the BotorchModel (for continuous params) via developer api:

Do I have to state an evaluation function when defining an "experiment"? Our "evaluation" is a training & testing of an ML model done on a cloud sever. I just want to feed the results to the model.

Also, I couldn't find how to load the known+pending evaluations to the model (and are the objective_weights, that the gen() fonction of BotorchModel requires, factors for low/high-fidelity evals?)

Thanks,
Avi

[Balandat]

Our "evaluation" is a training & testing of an ML model done on a cloud sever

If you want to manage that manually you could just use the service API - calling parameters, trial_index = ax.get_next_trial() will give you a new set of parameters to evaluate. You can then communicate with your cloud server to run the ML evaluation and get some results back in however way you want. Then calling ax.complete_trial(trial_index=trial_index, raw_data=results) will include that data into the experiment state. Then calling ax.get_next_trial() gives you the next set of parameters.

The benefit of using Ax for this is that it keeps track of all the metadata, so you can for instance store a complete history of your experiment in a sql database (if you hook Ax up to it with the appropriate tables).

The developer API gives you even more flexibility. For instance you can write a Runner that implements the logic for talking to the ML training job on the cloud server. For more information on this see https://ax.dev/docs/runner.html

I hope this helps. If you have specific questions about Ax, we should probably move the discussion to an issue in the Ax repo. There you'll also have folks like @lena-kashtelyan, @amurthy1, @sdsingh who understand the Ax APIs much better than I do.