This tutorial shows how to use the Sparse Axis-Aligned Subspace Bayesian Optimization (SAASBO) method for high-dimensional Bayesian optimization [1]. SAASBO places strong priors on the inverse lengthscales to avoid overfitting in high-dimensional spaces. Specifically, SAASBO uses a hierarchical sparsity prior consisting of a global shrinkage parameter $\tau \sim \mathcal{HC}(\beta)$ and inverse lengthscales $\rho_d \sim \mathcal{HC}(\tau)$ for $d=1, \ldots, D$, where $\mathcal{HC}$ is the half-Cauchy distribution. While half-Cauchy priors favor values near zero they also have heavy tails, which allows the inverse lengthscales of the most important parameters to escape zero. To perform inference in the SAAS model we use Hamiltonian Monte Carlo (HMC) as we found that to outperform MAP inference.
We find that SAASBO performs well on problems with hundreds of dimensions. As we rely on HMC and in particular the No-U-Turn-Sampler (NUTS) for inference, the overhead of SAASBO scales cubically with the number of datapoints. Depending on the problem, using more than a few hundred evaluations may not be feasible as SAASBO is designed for problems with a limited evaluation budget.
In general, we recommend using Ax for a simple BO setup like this one. See here for a SAASBO tutorial in Ax, which uses the Log Noisy Expected Improvement acquisition function. Therefore, this tutorial shows a minimal illustrative example of how to use SAASBO with only BoTorch. To customize the acquisition function used with SAASBO in Ax, see the custom acquisition tutorial, where adding \"surrogate\": Surrogate(SaasFullyBayesianSingleTaskGP), to the model_kwargs of BOTORCH_MODULAR step is sufficient to enable the SAAS model.
import os
import torch
from torch.quasirandom import SobolEngine
from botorch import fit_fully_bayesian_model_nuts
from botorch.acquisition.logei import qLogExpectedImprovement
from botorch.models.fully_bayesian import SaasFullyBayesianSingleTaskGP
from botorch.models.transforms import Standardize
from botorch.optim import optimize_acqf
from botorch.test_functions import Branin
SMOKE_TEST = os.environ.get("SMOKE_TEST")
tkwargs = {
"device": torch.device("cuda" if torch.cuda.is_available() else "cpu"),
"dtype": torch.double,
}
The time to fit the SAAS model can be decreased by lowering
WARMUP_STEPS and NUM_SAMPLES.
We recommend using 512 warmup steps and 256 samples when possible and to not use fewer than 256 warmup steps and 128 samples. By default, we only keep each 16th sample which with 256 samples results in 32 hyperparameter samples.
To make this tutorial run faster we use 256 warmup steps and 128 samples.
WARMUP_STEPS = 256 if not SMOKE_TEST else 32
NUM_SAMPLES = 128 if not SMOKE_TEST else 16
THINNING = 16
We generate a simple function that only depends on the first parameter and show that the SAAS model sets all other lengthscales to large values.
train_X = torch.rand(10, 4, **tkwargs)
test_X = torch.rand(5, 4, **tkwargs)
train_Y = torch.sin(train_X[:, :1])
test_Y = torch.sin(test_X[:, :1])
By default, we infer the unknown noise variance in the data. You can also pass in a known
noise variance (train_Yvar) for each observation, which may be useful in cases where you for example
know that the problem is noise-free and can then set the noise variance to a small value such as 1e-6.
In this case you can construct a model as follows:
gp = SaasFullyBayesianSingleTaskGP(train_X=train_X, train_Y=train_Y, train_Yvar=torch.full_like(train_Y, 1e-6))
gp = SaasFullyBayesianSingleTaskGP(
train_X=train_X,
train_Y=train_Y,
outcome_transform=Standardize(m=1)
)
fit_fully_bayesian_model_nuts(
gp,
warmup_steps=WARMUP_STEPS,
num_samples=NUM_SAMPLES,
thinning=THINNING,
disable_progbar=True,
)
with torch.no_grad():
posterior = gp.posterior(test_X)
Computing the median lengthscales over the MCMC dimensions makes it clear that the first feature has the smallest lengthscale
print(gp.median_lengthscale.detach())
In the next cell we show how to make predictions with the SAAS model. You compute the mean and variance for test points just like for any other BoTorch posteriors. Note that the mean and posterior will have an extra batch dimension at -3 that corresponds to the number of MCMC samples (which is 8 in this tutorial).
print(posterior.mean.shape)
print(posterior.variance.shape)
We also provide several convenience methods for computing different statistics over the MCMC samples:
mixture_mean = posterior.mixture_mean
mixture_variance = posterior.mixture_variance
mixture_quantile = posterior.quantile(q=0.95)
print(f"Ground truth: {test_Y.squeeze(-1)}")
print(f"Mixture mean: {posterior.mixture_mean.squeeze(-1)}")
We take the standard 2D Branin problem and embed it in a 30D space. In particular, we let dimensions 0 and 1 correspond to the true dimensions. We will show that SAASBO is able to identify the important dimensions and efficiently optimize this function. We work with the domain $[0, 1]^d$ and unnormalize the inputs to the true domain of Branin before evaluating the function.
branin = Branin().to(**tkwargs)
def branin_emb(x):
"""x is assumed to be in [0, 1]^d"""
lb, ub = branin.bounds
return branin(lb + (ub - lb) * x[..., :2])
DIM = 30 if not SMOKE_TEST else 2
# Evaluation budget
N_INIT = 10
N_ITERATIONS = 8 if not SMOKE_TEST else 1
BATCH_SIZE = 5 if not SMOKE_TEST else 1
print(f"Using a total of {N_INIT + BATCH_SIZE * N_ITERATIONS} function evaluations")
We use 10 initial Sobol points followed by 8 iterations of BO using a batch size of 5, which results in a total of 50 function evaluations. As our goal is to minimize Branin, we flip the sign of the function values before fitting the SAAS model as the BoTorch acquisition functions assume maximization.
X = SobolEngine(dimension=DIM, scramble=True, seed=0).draw(N_INIT).to(**tkwargs)
Y = branin_emb(X).unsqueeze(-1)
print(f"Best initial point: {Y.min().item():.3f}")
for i in range(N_ITERATIONS):
train_Y = -1 * Y # Flip the sign since we want to minimize f(x)
gp = SaasFullyBayesianSingleTaskGP(
train_X=X,
train_Y=train_Y,
train_Yvar=torch.full_like(train_Y, 1e-6),
outcome_transform=Standardize(m=1),
)
fit_fully_bayesian_model_nuts(
gp,
warmup_steps=WARMUP_STEPS,
num_samples=NUM_SAMPLES,
thinning=THINNING,
disable_progbar=True,
)
EI = qLogExpectedImprovement(model=gp, best_f=train_Y.max())
candidates, acq_values = optimize_acqf(
EI,
bounds=torch.cat((torch.zeros(1, DIM), torch.ones(1, DIM))).to(**tkwargs),
q=BATCH_SIZE,
num_restarts=10,
raw_samples=1024,
)
Y_next = torch.cat([branin_emb(x).unsqueeze(-1) for x in candidates]).unsqueeze(-1)
if Y_next.min() < Y.min():
ind_best = Y_next.argmin()
x0, x1 = candidates[ind_best, :2].tolist()
print(
f"{i + 1}) New best: {Y_next[ind_best].item():.3f} @ "
f"[{x0:.3f}, {x1:.3f}]"
)
X = torch.cat((X, candidates))
Y = torch.cat((Y, Y_next))
We can see that we were able to get close to the global optimium of $\approx 0.398$ after 50 function evaluations.
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline
Y_np = Y.cpu().numpy()
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(np.minimum.accumulate(Y_np), color="b", label="SAASBO")
ax.plot([0, len(Y_np)], [0.398, 0.398], "--", c="g", lw=3, label="Optimal value")
ax.grid(True)
ax.set_title(f"Branin, D = {DIM}", fontsize=20)
ax.set_xlabel("Number of evaluations", fontsize=20)
ax.set_xlim([0, len(Y_np)])
ax.set_ylabel("Best value found", fontsize=20)
ax.set_ylim([0, 8])
ax.legend(fontsize=18)
plt.show()
We fit a model using the 50 datapoints collected by SAASBO and predict on 50 test points in order to see how well the SAAS model predicts out-of-sample. The plot shows the mean and a 95% confidence interval for each test point.
train_X = SobolEngine(dimension=DIM, scramble=True, seed=0).draw(50).to(**tkwargs)
test_X = SobolEngine(dimension=DIM, scramble=True, seed=1).draw(50).to(**tkwargs)
train_Y = branin_emb(train_X).unsqueeze(-1)
test_Y = branin_emb(test_X).unsqueeze(-1)
gp = SaasFullyBayesianSingleTaskGP(
train_X=train_X,
train_Y=train_Y,
train_Yvar=torch.full_like(train_Y, 1e-6),
outcome_transform=Standardize(m=1),
)
fit_fully_bayesian_model_nuts(
gp,
warmup_steps=WARMUP_STEPS,
num_samples=NUM_SAMPLES,
thinning=THINNING,
disable_progbar=True,
)
with torch.no_grad():
posterior = gp.posterior(test_X)
median = posterior.quantile(value=torch.tensor([0.5], **tkwargs))
q1 = posterior.quantile(value=torch.tensor([0.025], **tkwargs))
q2 = posterior.quantile(value=torch.tensor([0.975], **tkwargs))
fig, ax = plt.subplots(1, 1, figsize=(8, 6))
ax.plot([0, 80], [0, 80], "b--", lw=2)
yerr1, yerr2 = median - q1, q2 - median
yerr = torch.cat((yerr1.unsqueeze(0), yerr2.unsqueeze(0)), dim=0).squeeze(-1)
markers, caps, bars = ax.errorbar(
test_Y.squeeze(-1).cpu().numpy(),
median.squeeze(-1).cpu().numpy(),
yerr=yerr.cpu().numpy(),
fmt=".",
capsize=4,
elinewidth=2.0,
ms=14,
c="k",
ecolor="gray",
)
ax.set_xlim([0, 80])
ax.set_ylim([0, 80])
[bar.set_alpha(0.8) for bar in bars]
[cap.set_alpha(0.8) for cap in caps]
ax.set_xlabel("True value", fontsize=20)
ax.set_ylabel("Predicted value", fontsize=20)
ax.set_aspect("equal")
ax.grid(True)
As SAASBO places strong priors on the inverse lengthscales, we only expect parameters 0 and 1 to be identified as important by the model since the other parameters have no effect. We can confirm that this is the case below as the lengthscales of parameters 0 and 1 are small with all other lengthscales being large.
median_lengthscales = gp.median_lengthscale
for i in median_lengthscales.argsort()[:10]:
print(f"Parameter {i:2}) Median lengthscale = {median_lengthscales[i].item():.2e}")