BO_init_tests

Experiments comparing Bayesian Optimisation libraries (and, later, initialisation strategies) on test and engineering design problems.

• BO_optimiser_test.ipynb: the original exploratory notebook.
• bo_bench/: a small benchmark harness that runs every method with the same seeds and evaluation budget.

Setup

pip install -r requirements.txt

Running the benchmark

python -m bo_bench.run --objective rastrigin2d --budget 50 --n-init 8 --seeds 0 1 2 3 4

Flag	Default	Meaning
--objective	rastrigin2d	Objective to optimise (rastrigin2d, rastrigin2d_narrow)
--methods	all	Any of bayes_opt, skopt
--seeds	0 1 2 3 4	One run per method per seed
--budget	50	Total objective evaluations per run, initial points included
--n-init	8	Random initial points (part of the budget)
--out	results	Output directory

Outputs, in --out:

• <objective>_results.csv: one row per evaluation with columns objective, method, seed, iteration, value, best_so_far.
• <objective>_convergence.png: median best-so-far per method across seeds, with an interquartile band.

All values use the maximisation convention (higher is better; Rastrigin's optimum is 0). skopt minimises, so it is given the negated objective (positive Rastrigin) and its results are flipped back before recording.

Layout

• bo_bench/objectives.py: objectives with bounds and known optimum. Add a new one to OBJECTIVES.
• bo_bench/optimizers.py: one adapter per library. Each takes (objective, budget, n_init, seed) and returns the evaluated values in order. Add a new one to OPTIMIZERS.
• bo_bench/run.py: CLI runner that writes the CSV and plot.
• bo_bench/plotting.py: convergence plot.