How to run or resume the BO loop

For how the surrogate and acquisition function actually work, see Multi-fidelity Bayesian optimization. For a zero-cost way to try the mechanics first, see Your first optimization loop. This page is just running it for real, against Basilisk.

1. Write a BO config

Copy config/bo_config.yaml and edit:

experiment_dir: experiments/bo_run_001

lf_fidelity: 5   # 32×32, ~5–15 min
hf_fidelity: 7   # 128×128, ~1–2 h

n_lf_init: 8     # LF samples for initial DoE (rule of thumb: ≈10× free params)
n_hf_init: 3     # HF samples for initial DoE (seeds the transfer-learning correlation)
n_iter: 20       # BO iterations after DoE

kla_key: kLa_25  # objective: kLa_10 | kLa_25 | kLa_50
n_candidates: 2000

walltime: "02:00:00"
job_timeout: 7200    # seconds to wait for results.json per job
t_buffer: 150

Parameter bounds come from config/param_space.yaml — see params.json reference.

2. Run it

python scripts/loop.py config/bo_config.yaml

The loop is resumable: it checks how many runs already exist in experiment_dir and skips the DoE phase if it's already complete. Kill and restart freely — nothing needs to be cleaned up first.

Standalone surrogate tools

Useful when you want to inspect or reuse the surrogate without running the whole loop.

Train manually:

python scripts/train_surrogate.py experiments/bo_run_001 surrogate/model.pkl \
    --lf-fidelity 5 --hf-fidelity 7 --kla-key kLa_25

Writes a pickled model.pkl with a .predict(X) -> (mean, var) interface.

Query the next suggested point:

python scripts/suggest.py experiments/bo_run_001 config/param_space.yaml \
    --model-path surrogate/model.pkl --kla-key kLa_25 --n-candidates 2000

Prints a JSON params dict to stdout — the highest-EI candidate.