Back to .md Directory
-2 is the default value when no feasible has been found
Compares constrained EI with standard EI for Bayesian optimisation under a cost constraint, tracking feasible incumbent values across repeated trials.
ai
View sourceWhat this file does
Compares constrained EI with standard EI for Bayesian optimisation under a cost constraint, tracking feasible incumbent values across repeated trials.
When to use it
- You want to enforce a constraint during Bayesian optimisation
- You need to compare constrained vs unconstrained acquisition functions
- You are working with BoTorch and a custom cost function
- You want to visualise optimisation progress with confidence intervals
Assumes this stack
PythonPyTorchBoTorchGPyTorchMatplotlib
import warnings
import botorch
import gpytorch
import matplotlib.pyplot as plt
import numpy as np
import torch
from tqdm.notebook import tqdm
from bo.objectives import forrester
from bo.train import fit_gp_model
def objective(x):
return forrester(x)
def cost(x):
return -(0.1 * objective(x) + objective(x - 4)) / 3 + x / 3 - 0.5
lb = -5
ub = +5
bounds = torch.tensor([[lb], [ub]], dtype=torch.float)
xs = torch.linspace(lb, ub, 201)
objective_ys = objective(xs)
cost_ys = cost(xs)
plt.plot(xs, objective_ys, label="objective", c="r")
plt.plot(xs, cost_ys, linestyle="--", label="cost", c="k")
plt.legend();
class GPModel(gpytorch.models.ExactGP, botorch.models.gpytorch.GPyTorchModel):
_num_outputs = 1
def __init__(self, train_x, train_y, likelihood):
super().__init__(train_x, train_y, likelihood)
self.mean_module = gpytorch.means.ConstantMean()
self.covar_module = gpytorch.kernels.ScaleKernel(
gpytorch.kernels.MaternKernel(nu=2.5, ard_num_dims=1)
)
def forward(self, x):
mean_x = self.mean_module(x)
covar_x = self.covar_module(x)
return gpytorch.distributions.MultivariateNormal(mean_x, covar_x)
num_queries = 10
num_repeats = 10
strategy = "ei" # "cei" or "ei"
strategy = strategy.upper()
# -2 is the default value when no feasible has been found
default_value = -2
feasible_incumbents = torch.ones((num_repeats, num_queries)) * default_value
for trial in range(num_repeats):
print("trial", trial)
torch.manual_seed(trial)
train_x = bounds[0] + (bounds[1] - bounds[0]) * torch.rand(1, 1)
train_utility = objective(train_x)
train_cost = cost(train_x)
for i in tqdm(range(num_queries)):
feasible_flag = (train_cost <= 0).any()
if feasible_flag:
feasible_incumbents[trial, i] = train_utility[train_cost <= 0].max()
utility_likelihood = gpytorch.likelihoods.GaussianLikelihood()
utility_model = GPModel(train_x, train_utility.squeeze(-1), utility_likelihood)
utility_model.likelihood.noise = 1e-4
fit_gp_model(utility_model, utility_likelihood, train_x, train_utility.squeeze(-1))
cost_likelihood = gpytorch.likelihoods.GaussianLikelihood()
cost_model = GPModel(train_x, train_cost.squeeze(-1), cost_likelihood)
cost_model.likelihood.noise = 1e-4
fit_gp_model(cost_model, cost_likelihood, train_x, train_cost.squeeze(-1))
if feasible_flag:
best_f = train_utility[train_cost <= 0].max()
else:
best_f = torch.tensor(default_value)
if strategy == "EI":
policy = botorch.acquisition.analytic.LogExpectedImprovement(
model=utility_model, best_f=train_utility.max()
)
if strategy == "CEI":
policy = botorch.acquisition.analytic.LogConstrainedExpectedImprovement(
model=botorch.models.model_list_gp_regression.ModelListGP(
utility_model, cost_model
),
best_f=best_f,
objective_index=0,
constraints={1: [None, 0]},
)
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=RuntimeWarning)
next_x, acq_val = botorch.optim.optimize_acqf(
policy, bounds=bounds, q=1, num_restarts=40, raw_samples=100
)
next_utility = objective(next_x)
next_cost = cost(next_x)
train_x = torch.cat([train_x, next_x])
train_utility = torch.cat([train_utility, next_utility])
train_cost = torch.cat([train_cost, next_cost])
trial 0
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 1
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 2
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 3
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 4
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 5
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 6
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 7
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 8
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
trial 9
0%| | 0/10 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
0%| | 0/500 [00:00<?, ?it/s]
!mkdir forrester
torch.save(feasible_incumbents, f"./forrester/{strategy}.pth")
mkdir: forrester: File exists
def show_agg_progress(path, name, linestyle="solid"):
def ci(y):
return 2 * y.std(axis=0) / np.sqrt(num_repeats)
incumbents = torch.load(path)
avg_incumbent = incumbents.mean(axis=0)
ci_incumbent = ci(incumbents)
plt.plot(avg_incumbent, linestyle=linestyle, label=name)
plt.fill_between(
np.arange(num_queries),
avg_incumbent + ci_incumbent,
avg_incumbent - ci_incumbent,
alpha=0.1,
)
plt.figure(figsize=(8, 6))
show_agg_progress(
"./forrester/CEI.pth",
"constrained EI",
)
show_agg_progress("./forrester/EI.pth", "EI", linestyle="--")
plt.axhline(
objective_ys[cost_ys <= 0].max(),
c="k",
linestyle=":",
label="feasible best",
)
plt.xlabel("# queries", fontsize=15)
plt.ylabel("accuracy", fontsize=15)
plt.legend(fontsize=15);
ei_incumbents = torch.load("./forrester/EI.pth")
cei_incumbents = torch.load("./forrester/CEI.pth")
torch.set_printoptions(precision=1)
print(ei_incumbents)
tensor([[ 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8],
[-2.0, 2.7, 2.7, 3.3, 4.1, 4.2, 4.2, 4.2, 4.2, 4.2],
[ 1.8, 1.8, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2],
[ 4.1, 4.1, 4.1, 4.2, 4.2, 4.2, 4.2, 4.2, 4.2, 4.2],
[ 1.0, 1.6, 4.0, 4.2, 4.2, 4.2, 4.2, 4.2, 4.2, 4.2],
[-2.0, 1.9, 3.5, 4.1, 4.2, 4.2, 4.2, 4.2, 4.2, 4.2],
[ 1.2, 1.2, 1.9, 1.9, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2],
[ 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8],
[ 1.6, 1.6, 2.1, 2.1, 2.1, 2.1, 2.2, 2.2, 4.2, 4.2],
[ 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2, 2.2]])
What's inside
3 code cells for setup and plotting, 1 class definition, 1 main loop over 10 trials and 10 queries, 2 saving/loading cells
Change this for your project
- Replace
forresterobjective andcostfunction with your own - Replace
./forrester/save path with your project directory - Replace
num_queries = 10andnum_repeats = 10with your desired counts
Where it goes
Keep it in your repository where the agent or team that needs it will read it.
Worth borrowing
- Using
default_value = -2as placeholder when no feasible point exists - Tracking feasible incumbents in a tensor for later analysis
Related Documents
CONSTRAINTS.md
Xilinx Constraints
Explains Xilinx XDC constraint ordering, compilation precedence, and key references for FPGA timing closure.
aieval
0
6
lastweekCONSTRAINTS.md
Technical Constraints & Requirements
Defines platform, performance, security, and deployment constraints for a Scottish mountain weather app with native iOS/Android and PWA fallback.
ai
0
5
m-deaneCONSTRAINTS.md
Token Constraints Implementation
Defines a token constraint system that counts tokens, truncates content, and streams responses to stay within an LLM's token limit.
aillmrag
0
4
markoveranoCONSTRAINTS.md
Constraints
Explains how to model constraints as Ceres Solver cost functions in the fuse optimization framework, with a complete code example.
aisafety
0
4
locusrobotics