In this tutorial, we illustrate how to implement a simple multi-objective (MO) Bayesian Optimization (BO) closed loop in BoTorch.
In general, we recommend using Ax for a simple BO setup like this one, since this will simplify your setup (including the amount of code you need to write) considerably. See here for an Ax tutorial on MOBO. If desired, you can use a custom BoTorch model in Ax, following the Using BoTorch with Ax tutorial. Given a MultiObjective, Ax will default to the $q$NEHVI acquisiton function. If desired, this can also be customized by adding "botorch_acqf_class": <desired_botorch_acquisition_function_class>, to the model_kwargs.
We use the parallel ParEGO ($q$ParEGO) [1], parallel Expected Hypervolume Improvement ($q$EHVI) [1], and parallel Noisy Expected Hypervolume Improvement ($q$NEHVI) [2] acquisition functions to optimize a synthetic BraninCurrin problem test function with additive Gaussian observation noise over a 2-parameter search space [0,1]^2. See botorch/test_functions/multi_objective.py for details on BraninCurrin. The noise standard deviations are 15.19 and 0.63 for each objective, respectively.
Since botorch assumes a maximization of all objectives, we seek to find the Pareto frontier, the set of optimal trade-offs where improving one metric means deteriorating another.
For batch optimization (or in noisy settings), we strongly recommend using $q$NEHVI rather than $q$EHVI because it is far more efficient than $q$EHVI and mathematically equivalent in the noiseless setting.
Note: $q$EHVI and $q$NEHVI aggressively exploit parallel hardware and are both much faster when run on a GPU. See [1, 2] for details.
import os
import torch
tkwargs = {
"dtype": torch.double,
"device": torch.device("cuda" if torch.cuda.is_available() else "cpu"),
}
SMOKE_TEST = os.environ.get("SMOKE_TEST")
from botorch.test_functions.multi_objective import BraninCurrin
problem = BraninCurrin(negate=True).to(**tkwargs)
We use a list of SingleTaskGPs to model the two objectives with known noise variances. If no noise variances were provided, SingleTaskGP would infer (homoskedastic) noise levels instead.
The models are initialized with $2(d+1)=6$ points drawn randomly from $[0,1]^2$.
from botorch.models.gp_regression import SingleTaskGP
from botorch.models.model_list_gp_regression import ModelListGP
from botorch.models.transforms.outcome import Standardize
from gpytorch.mlls.sum_marginal_log_likelihood import SumMarginalLogLikelihood
from botorch.utils.transforms import unnormalize, normalize
from botorch.utils.sampling import draw_sobol_samples
NOISE_SE = torch.tensor([15.19, 0.63], **tkwargs)
def generate_initial_data(n=6):
# generate training data
train_x = draw_sobol_samples(bounds=problem.bounds, n=n, q=1).squeeze(1)
train_obj_true = problem(train_x)
train_obj = train_obj_true + torch.randn_like(train_obj_true) * NOISE_SE
return train_x, train_obj, train_obj_true
def initialize_model(train_x, train_obj):
# define models for objective and constraint
train_x = normalize(train_x, problem.bounds)
models = []
for i in range(train_obj.shape[-1]):
train_y = train_obj[..., i : i + 1]
train_yvar = torch.full_like(train_y, NOISE_SE[i] ** 2)
models.append(
SingleTaskGP(
train_x, train_y, train_yvar, outcome_transform=Standardize(m=1)
)
)
model = ModelListGP(*models)
mll = SumMarginalLogLikelihood(model.likelihood, model)
return mll, model
The helper function below initializes the $q$EHVI acquisition function, optimizes it, and returns the batch $\{x_1, x_2, \ldots x_q\}$ along with the observed function values.
For this example, we'll use a relatively small batch of optimization ($q=4$). For batch optimization ($q>1$), passing the keyword argument sequential=True to the function optimize_acqfspecifies that candidates should be optimized in a sequential greedy fashion (see [1] for details why this is important). A simple initialization heuristic is used to select the 10 restart initial locations from a set of 512 random points. Multi-start optimization of the acquisition function is performed using LBFGS-B with exact gradients computed via auto-differentiation.
Reference Point
$q$EHVI requires specifying a reference point, which is the lower bound on the objectives used for computing hypervolume. In this tutorial, we assume the reference point is known. In practice the reference point can be set 1) using domain knowledge to be slightly worse than the lower bound of objective values, where the lower bound is the minimum acceptable value of interest for each objective, or 2) using a dynamic reference point selection strategy.
Partitioning the Non-dominated Space into disjoint rectangles
$q$EHVI requires partitioning the non-dominated space into disjoint rectangles (see [1] for details).
Note: FastNondominatedPartitioning will be very slow when 1) there are a lot of points on the pareto frontier and 2) there are >5 objectives.
from botorch.optim.optimize import optimize_acqf, optimize_acqf_list
from botorch.acquisition.objective import GenericMCObjective
from botorch.utils.multi_objective.scalarization import get_chebyshev_scalarization
from botorch.utils.multi_objective.box_decompositions.non_dominated import (
FastNondominatedPartitioning,
)
from botorch.acquisition.multi_objective.monte_carlo import (
qExpectedHypervolumeImprovement,
qNoisyExpectedHypervolumeImprovement,
)
from botorch.utils.sampling import sample_simplex
BATCH_SIZE = 4
NUM_RESTARTS = 10 if not SMOKE_TEST else 2
RAW_SAMPLES = 512 if not SMOKE_TEST else 4
standard_bounds = torch.zeros(2, problem.dim, **tkwargs)
standard_bounds[1] = 1
def optimize_qehvi_and_get_observation(model, train_x, train_obj, sampler):
"""Optimizes the qEHVI acquisition function, and returns a new candidate and observation."""
# partition non-dominated space into disjoint rectangles
with torch.no_grad():
pred = model.posterior(normalize(train_x, problem.bounds)).mean
partitioning = FastNondominatedPartitioning(
ref_point=problem.ref_point,
Y=pred,
)
acq_func = qExpectedHypervolumeImprovement(
model=model,
ref_point=problem.ref_point,
partitioning=partitioning,
sampler=sampler,
)
# optimize
candidates, _ = optimize_acqf(
acq_function=acq_func,
bounds=standard_bounds,
q=BATCH_SIZE,
num_restarts=NUM_RESTARTS,
raw_samples=RAW_SAMPLES, # used for intialization heuristic
options={"batch_limit": 5, "maxiter": 200},
sequential=True,
)
# observe new values
new_x = unnormalize(candidates.detach(), bounds=problem.bounds)
new_obj_true = problem(new_x)
new_obj = new_obj_true + torch.randn_like(new_obj_true) * NOISE_SE
return new_x, new_obj, new_obj_true
Integrating over function values at in-sample designs
$q$NEHVI integrates over the unknown function values at the previously evaluated designs (see [2] for details). Therefore, we need to provide the previously evaluated designs (train_x, normalized to be within $[0,1]^d$) to the acquisition function.
Efficient batch generation with Cached Box Decomposition (CBD)
$q$NEHVI leveraged CBD to efficiently generate large batches of candidates. CBD scales polynomially with respect to the batch size where as the inclusion-exclusion principle used by qEHVI scales exponentially with the batch size.
Pruning baseline designs
To speed up integration over the function values at the previously evaluated designs, we prune the set of previously evaluated designs (by setting prune_baseline=True) to only include those which have positive probability of being on the current in-sample Pareto frontier.
def optimize_qnehvi_and_get_observation(model, train_x, train_obj, sampler):
"""Optimizes the qEHVI acquisition function, and returns a new candidate and observation."""
# partition non-dominated space into disjoint rectangles
acq_func = qNoisyExpectedHypervolumeImprovement(
model=model,
ref_point=problem.ref_point.tolist(), # use known reference point
X_baseline=normalize(train_x, problem.bounds),
prune_baseline=True, # prune baseline points that have estimated zero probability of being Pareto optimal
sampler=sampler,
)
# optimize
candidates, _ = optimize_acqf(
acq_function=acq_func,
bounds=standard_bounds,
q=BATCH_SIZE,
num_restarts=NUM_RESTARTS,
raw_samples=RAW_SAMPLES, # used for intialization heuristic
options={"batch_limit": 5, "maxiter": 200},
sequential=True,
)
# observe new values
new_x = unnormalize(candidates.detach(), bounds=problem.bounds)
new_obj_true = problem(new_x)
new_obj = new_obj_true + torch.randn_like(new_obj_true) * NOISE_SE
return new_x, new_obj, new_obj_true
The helper function below similarly initializes $q$NParEGO, optimizes it, and returns the batch $\{x_1, x_2, \ldots x_q\}$ along with the observed function values.
$q$NParEGO uses random augmented chebyshev scalarization with the qNoisyExpectedImprovement acquisition function. In the parallel setting ($q>1$), each candidate is optimized in sequential greedy fashion using a different random scalarization (see [1] for details).
To do this, we create a list of qNoisyExpectedImprovement acquisition functions, each with different random scalarization weights. The optimize_acqf_list method sequentially generates one candidate per acquisition function and conditions the next candidate (and acquisition function) on the previously selected pending candidates.
from botorch.acquisition.monte_carlo import qNoisyExpectedImprovement
def optimize_qnparego_and_get_observation(model, train_x, train_obj, sampler):
"""Samples a set of random weights for each candidate in the batch, performs sequential greedy optimization
of the qNParEGO acquisition function, and returns a new candidate and observation."""
train_x = normalize(train_x, problem.bounds)
with torch.no_grad():
pred = model.posterior(train_x).mean
acq_func_list = []
for _ in range(BATCH_SIZE):
weights = sample_simplex(problem.num_objectives, **tkwargs).squeeze()
objective = GenericMCObjective(
get_chebyshev_scalarization(weights=weights, Y=pred)
)
acq_func = qNoisyExpectedImprovement( # pyre-ignore: [28]
model=model,
objective=objective,
X_baseline=train_x,
sampler=sampler,
prune_baseline=True,
)
acq_func_list.append(acq_func)
# optimize
candidates, _ = optimize_acqf_list(
acq_function_list=acq_func_list,
bounds=standard_bounds,
num_restarts=NUM_RESTARTS,
raw_samples=RAW_SAMPLES, # used for intialization heuristic
options={"batch_limit": 5, "maxiter": 200},
)
# observe new values
new_x = unnormalize(candidates.detach(), bounds=problem.bounds)
new_obj_true = problem(new_x)
new_obj = new_obj_true + torch.randn_like(new_obj_true) * NOISE_SE
return new_x, new_obj, new_obj_true
The Bayesian optimization "loop" for a batch size of $q$ simply iterates the following steps:
Just for illustration purposes, we run one trial with N_BATCH=20 rounds of optimization. The acquisition function is approximated using MC_SAMPLES=128 samples.
Note: Running this may take a little while.
import time
import warnings
from botorch import fit_gpytorch_mll
from botorch.exceptions import BadInitialCandidatesWarning
from botorch.sampling.normal import SobolQMCNormalSampler
from botorch.utils.multi_objective.box_decompositions.dominated import (
DominatedPartitioning,
)
from botorch.utils.multi_objective.pareto import is_non_dominated
warnings.filterwarnings("ignore", category=BadInitialCandidatesWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)
N_BATCH = 20 if not SMOKE_TEST else 5
MC_SAMPLES = 128 if not SMOKE_TEST else 16
verbose = True
hvs_qparego, hvs_qehvi, hvs_qnehvi, hvs_random = [], [], [], []
# call helper functions to generate initial training data and initialize model
train_x_qparego, train_obj_qparego, train_obj_true_qparego = generate_initial_data(
n=2 * (problem.dim + 1)
)
mll_qparego, model_qparego = initialize_model(train_x_qparego, train_obj_qparego)
train_x_qehvi, train_obj_qehvi, train_obj_true_qehvi = (
train_x_qparego,
train_obj_qparego,
train_obj_true_qparego,
)
train_x_qnehvi, train_obj_qnehvi, train_obj_true_qnehvi = (
train_x_qparego,
train_obj_qparego,
train_obj_true_qparego,
)
train_x_random, train_obj_random, train_obj_true_random = (
train_x_qparego,
train_obj_qparego,
train_obj_true_qparego,
)
mll_qehvi, model_qehvi = initialize_model(train_x_qehvi, train_obj_qehvi)
mll_qnehvi, model_qnehvi = initialize_model(train_x_qnehvi, train_obj_qnehvi)
# compute hypervolume
bd = DominatedPartitioning(ref_point=problem.ref_point, Y=train_obj_true_qparego)
volume = bd.compute_hypervolume().item()
hvs_qparego.append(volume)
hvs_qehvi.append(volume)
hvs_qnehvi.append(volume)
hvs_random.append(volume)
# run N_BATCH rounds of BayesOpt after the initial random batch
for iteration in range(1, N_BATCH + 1):
t0 = time.monotonic()
# fit the models
fit_gpytorch_mll(mll_qparego)
fit_gpytorch_mll(mll_qehvi)
fit_gpytorch_mll(mll_qnehvi)
# define the qEI and qNEI acquisition modules using a QMC sampler
qparego_sampler = SobolQMCNormalSampler(sample_shape=torch.Size([MC_SAMPLES]))
qehvi_sampler = SobolQMCNormalSampler(sample_shape=torch.Size([MC_SAMPLES]))
qnehvi_sampler = SobolQMCNormalSampler(sample_shape=torch.Size([MC_SAMPLES]))
# optimize acquisition functions and get new observations
(
new_x_qparego,
new_obj_qparego,
new_obj_true_qparego,
) = optimize_qnparego_and_get_observation(
model_qparego, train_x_qparego, train_obj_qparego, qparego_sampler
)
new_x_qehvi, new_obj_qehvi, new_obj_true_qehvi = optimize_qehvi_and_get_observation(
model_qehvi, train_x_qehvi, train_obj_qehvi, qehvi_sampler
)
(
new_x_qnehvi,
new_obj_qnehvi,
new_obj_true_qnehvi,
) = optimize_qnehvi_and_get_observation(
model_qnehvi, train_x_qnehvi, train_obj_qnehvi, qnehvi_sampler
)
new_x_random, new_obj_random, new_obj_true_random = generate_initial_data(
n=BATCH_SIZE
)
# update training points
train_x_qparego = torch.cat([train_x_qparego, new_x_qparego])
train_obj_qparego = torch.cat([train_obj_qparego, new_obj_qparego])
train_obj_true_qparego = torch.cat([train_obj_true_qparego, new_obj_true_qparego])
train_x_qehvi = torch.cat([train_x_qehvi, new_x_qehvi])
train_obj_qehvi = torch.cat([train_obj_qehvi, new_obj_qehvi])
train_obj_true_qehvi = torch.cat([train_obj_true_qehvi, new_obj_true_qehvi])
train_x_qnehvi = torch.cat([train_x_qnehvi, new_x_qnehvi])
train_obj_qnehvi = torch.cat([train_obj_qnehvi, new_obj_qnehvi])
train_obj_true_qnehvi = torch.cat([train_obj_true_qnehvi, new_obj_true_qnehvi])
train_x_random = torch.cat([train_x_random, new_x_random])
train_obj_random = torch.cat([train_obj_random, new_obj_random])
train_obj_true_random = torch.cat([train_obj_true_random, new_obj_true_random])
# update progress
for hvs_list, train_obj in zip(
(hvs_random, hvs_qparego, hvs_qehvi, hvs_qnehvi),
(
train_obj_true_random,
train_obj_true_qparego,
train_obj_true_qehvi,
train_obj_true_qnehvi,
),
):
# compute hypervolume
bd = DominatedPartitioning(ref_point=problem.ref_point, Y=train_obj)
volume = bd.compute_hypervolume().item()
hvs_list.append(volume)
# reinitialize the models so they are ready for fitting on next iteration
# Note: we find improved performance from not warm starting the model hyperparameters
# using the hyperparameters from the previous iteration
mll_qparego, model_qparego = initialize_model(train_x_qparego, train_obj_qparego)
mll_qehvi, model_qehvi = initialize_model(train_x_qehvi, train_obj_qehvi)
mll_qnehvi, model_qnehvi = initialize_model(train_x_qnehvi, train_obj_qnehvi)
t1 = time.monotonic()
if verbose:
print(
f"\nBatch {iteration:>2}: Hypervolume (random, qNParEGO, qEHVI, qNEHVI) = "
f"({hvs_random[-1]:>4.2f}, {hvs_qparego[-1]:>4.2f}, {hvs_qehvi[-1]:>4.2f}, {hvs_qnehvi[-1]:>4.2f}), "
f"time = {t1-t0:>4.2f}.",
end="",
)
else:
print(".", end="")
The plot below shows the a common metric of multi-objective optimization performance, the log hypervolume difference: the log difference between the hypervolume of the true pareto front and the hypervolume of the approximate pareto front identified by each algorithm. The log hypervolume difference is plotted at each step of the optimization for each of the algorithms.
The plot shows that $q$NEHVI outperforms $q$EHVI, $q$ParEGO, and Sobol.
import numpy as np
from matplotlib import pyplot as plt
%matplotlib inline
iters = np.arange(N_BATCH + 1) * BATCH_SIZE
log_hv_difference_qparego = np.log10(problem.max_hv - np.asarray(hvs_qparego))
log_hv_difference_qehvi = np.log10(problem.max_hv - np.asarray(hvs_qehvi))
log_hv_difference_qnehvi = np.log10(problem.max_hv - np.asarray(hvs_qnehvi))
log_hv_difference_rnd = np.log10(problem.max_hv - np.asarray(hvs_random))
fig, ax = plt.subplots(1, 1, figsize=(8, 6))
ax.errorbar(
iters,
log_hv_difference_rnd,
label="Sobol",
linewidth=1.5,
)
ax.errorbar(
iters,
log_hv_difference_qparego,
label="qNParEGO",
linewidth=1.5,
)
ax.errorbar(
iters,
log_hv_difference_qehvi,
label="qEHVI",
linewidth=1.5,
)
ax.errorbar(
iters,
log_hv_difference_qnehvi,
label="qNEHVI",
linewidth=1.5,
)
ax.set(
xlabel="number of observations (beyond initial points)",
ylabel="Log Hypervolume Difference",
)
ax.legend(loc="lower left")
To examine optimization process from another perspective, we plot the true function values at the designs selected under each algorithm where the color corresponds to the BO iteration at which the point was collected. The plot on the right for $q$NEHVI shows that the $q$NEHVI quickly identifies the pareto front and most of its evaluations are very close to the pareto front. $q$NParEGO also identifies has many observations close to the pareto front, but relies on optimizing random scalarizations, which is a less principled way of optimizing the pareto front compared to $q$NEHVI, which explicitly attempts focuses on improving the pareto front. $q$EHVI uses the posterior mean as a plug-in estimator for the true function values at the in-sample points, whereas $q$NEHVI than integrating over the uncertainty at the in-sample designs Sobol generates random points and has few points close to the Pareto front.
from matplotlib.cm import ScalarMappable
fig, axes = plt.subplots(1, 4, figsize=(23, 7), sharex=True, sharey=True)
algos = ["Sobol", "qNParEGO", "qEHVI", "qNEHVI"]
cm = plt.get_cmap("viridis")
batch_number = torch.cat(
[
torch.zeros(2 * (problem.dim + 1)),
torch.arange(1, N_BATCH + 1).repeat(BATCH_SIZE, 1).t().reshape(-1),
]
).numpy()
for i, train_obj in enumerate(
(
train_obj_true_random,
train_obj_true_qparego,
train_obj_true_qehvi,
train_obj_true_qnehvi,
)
):
sc = axes[i].scatter(
train_obj[:, 0].cpu().numpy(),
train_obj[:, 1].cpu().numpy(),
c=batch_number,
alpha=0.8,
)
axes[i].set_title(algos[i])
axes[i].set_xlabel("Objective 1")
axes[0].set_ylabel("Objective 2")
norm = plt.Normalize(batch_number.min(), batch_number.max())
sm = ScalarMappable(norm=norm, cmap=cm)
sm.set_array([])
fig.subplots_adjust(right=0.9)
cbar_ax = fig.add_axes([0.93, 0.15, 0.01, 0.7])
cbar = fig.colorbar(sm, cax=cbar_ax)
cbar.ax.set_title("Iteration")
