T03 — Bayesian Experimental Design (Single Objective)

Goal: Use a trained surrogate GP to recommend the next experiment via LogEI acquisition, maximising titer.

Script: examples/03_bayesian_experimental_design.py

Prerequisite: model.pt from T02 (or the script’s built-in SingleTaskGP placeholder).

What the script does

1. Builds (or loads) a GP surrogate over the 6-dimensional control space.

2. Calls build_acquisition("LogEI", gp, best_f=best_f) to construct a numerically-stable Expected Improvement acquisition.

3. Optimises the acquisition via optimise_acquisition(acqf, bounds, q=1) (20 restarts, 512 raw samples).

4. Prints the recommended next operating point.

Running

python examples/03_bayesian_experimental_design.py

Key code

from perfusio.bed.acquisitions import build_acquisition
from perfusio.bed.search import optimise_acquisition
from perfusio.config import DEFAULT_AMBR250_DESIGN_SPACE

DS = DEFAULT_AMBR250_DESIGN_SPACE
bounds = DS.bounds_tensor()  # shape (2, n_controls)

acqf = build_acquisition("LogEI", surrogate_gp, best_f=current_best)
candidate, value = optimise_acquisition(acqf, bounds=bounds, q=1)

Available acquisition functions

See Choosing an Acquisition Function for a full guide. The most common choices:

Use case	Recommended acquisition
Single run, noisy	LogEI
Batch of 4	qLogEI
Exploration emphasis	UCB with high β

Acquisition parameters

build_acquisition accepts keyword arguments forwarded to BoTorch:

build_acquisition("UCB", model, beta=2.0)
build_acquisition("qLogEI", model, best_f=0.42, num_samples=128)
build_acquisition("qNEHVI", model, ref_point=ref_pt, partitioning=part)

Next step

Proceed to T04 — Multi-Objective BED.