[GENERAL SUPPORT]: Implementation of Evolution-guided BO

[VMLC-PV]

Question

I was thinking of implementing the Evolution-guided BO as described in this paper and I thought that it would make sense to write something with a similar structure to the SEBOAcquisition class and with a .optimize similar to the EGBO implemented in this repo.
To do so, I need to pull back the untransformed X_observed and the corresponding metrics.
I went down the rabbit hole and explored the attributes of many of the objects passed to SEBOAcquisition but could not find what I needed. Any ideas?

I know I could use the original codes here but I was hoping to write something that is more 'Ax-like'.
For context, I print out below the code from the repo implementing the EGBO.

def optimize_st_egbo(acq_func, x, y, batch_size):
	# for st qnehvi
	
	pareto_mask = is_non_dominated(y)
	pareto_x = x[pareto_mask].cpu().numpy()
	
	problem = PymooProblem(n_var=x.shape[1], n_obj=y.shape[1], n_constr=0,
                           xl=np.zeros(x.shape[1]), xu=np.ones(x.shape[1]))
    ref_dirs = get_reference_directions("energy", y.shape[1], batch_size*n_tasks)
    algo = NSGA3(pop_size=raw_samples, ref_dirs=ref_dirs)
    
    algo.setup(problem, termination = NoTermination())
    pop = Population.new("X", x.cpu().numpy())
    pop.set("F", -y.cpu().numpy())
    algo.tell(infills=pop)
    new_pop = algo.ask()

	candidates = torch.tensor(new_pop.get("X"), **tkwargs)
		
	acq_value_list = [acq_func(candidates[i].unsqueeze(dim=0)).detach().item()
					  for i in range(candidates.shape[0])]
	sorted_x = candidates.cpu().numpy()[np.argsort(acq_value_list)]
	
	return torch.tensor(sorted_x[-batch_size:], **tkwargs) # take best BATCH_SIZE samples

Please provide any relevant code snippet if applicable.

No response

Code of Conduct

• I agree to follow this Ax's Code of Conduct

[sdaulton]

If all you need access to is the training data (Xs, Ys), you should be able to get this from Surrogate.training_data (

Ax/ax/models/torch/botorch_modular/surrogate.py

Line 782 in 7bb3fe4

def training_data(self) -> list[SupervisedDataset]:

) and the surrogate is based to the Acquisition.__init__ (

Ax/ax/models/torch/botorch_modular/acquisition.py

Line 92 in 7bb3fe4

surrogate: Surrogate,

), so you could subclass Acquisition similar to SEBO and then store the training data on the Acquisition object, so that you can access it during optimize.