add initial objective values to initial state for sample efficiency

[brunzema]

Motivation

Thank you for providing TuRBO as a small tutorial! In the notebook, however, the initial Y values are ignored in the TurboState which seems data inefficient. As this notebook is the starting point for multiple people if they want to compare against or use TuRBO, I believe this should be fixed.

I document my change below as the diff with notebooks is not super readable. :)

Old implementation (from https://botorch.org/tutorials/turbo_1):

X_turbo = get_initial_points(dim, n_init)
Y_turbo = torch.tensor(
    [eval_objective(x) for x in X_turbo], dtype=dtype, device=device
).unsqueeze(-1)

state = TurboState(dim, batch_size=batch_size)

Crucially, when updating the TurboState, only the most recent Y values (Y_next) are considered and, therefore, the initial points are discarded:

def update_state(state, Y_next):
    ...
    state.best_value = max(state.best_value, max(Y_next).item())

Long story short: I changed the initialization of the state to also include the initial Y values:

X_turbo = get_initial_points(dim, n_init)
Y_turbo = torch.tensor(
    [eval_objective(x) for x in X_turbo], dtype=dtype, device=device
).unsqueeze(-1)

state = TurboState(dim, batch_size=batch_size, best_value=max(Y_turbo).item())

Interestingly, this also (ever so slightly :D ) improves the final performance of TuRBO :)

OLD:

NEW:

---

I also added warning supression to the imports to hide my path--I hope that this is ok (I think I also saw this in other notebooks), but I can of course also change this back

Old first code cell:

import os
import math
from dataclasses import dataclass

import torch
from botorch.acquisition import qExpectedImprovement, qLogExpectedImprovement
from botorch.fit import fit_gpytorch_mll
from botorch.generation import MaxPosteriorSampling
from botorch.models import SingleTaskGP
from botorch.optim import optimize_acqf
from botorch.test_functions import Ackley
from botorch.utils.transforms import unnormalize
from torch.quasirandom import SobolEngine

import gpytorch
from gpytorch.constraints import Interval
from gpytorch.kernels import MaternKernel, ScaleKernel
from gpytorch.likelihoods import GaussianLikelihood
from gpytorch.mlls import ExactMarginalLogLikelihood
from gpytorch.priors import HorseshoePrior


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.double
SMOKE_TEST = os.environ.get("SMOKE_TEST")

New:

import os
import math
import warnings
from dataclasses import dataclass

import torch
from botorch.acquisition import qExpectedImprovement, qLogExpectedImprovement
from botorch.exceptions import BadInitialCandidatesWarning
from botorch.fit import fit_gpytorch_mll
from botorch.generation import MaxPosteriorSampling
from botorch.models import SingleTaskGP
from botorch.optim import optimize_acqf
from botorch.test_functions import Ackley
from botorch.utils.transforms import unnormalize
from torch.quasirandom import SobolEngine

import gpytorch
from gpytorch.constraints import Interval
from gpytorch.kernels import MaternKernel, ScaleKernel
from gpytorch.likelihoods import GaussianLikelihood
from gpytorch.mlls import ExactMarginalLogLikelihood


warnings.filterwarnings("ignore", category=BadInitialCandidatesWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.double
SMOKE_TEST = os.environ.get("SMOKE_TEST")

Have you read the Contributing Guidelines on pull requests?

Yes.

Test Plan

I believe this minor change does not need a test plan.

Related PRs

PR does not change functionality.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.98%. Comparing base (8f25f5b) to head (34f7ee5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2365   +/-   ##
=======================================
  Coverage   99.98%   99.98%           
=======================================
  Files         190      190           
  Lines       16587    16587           
=======================================
  Hits        16584    16584           
  Misses          3        3           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[facebook-github-bot]

@sdaulton has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[sdaulton]

lgtm, thanks!

[facebook-github-bot]

@sdaulton merged this pull request in 8b32670.