Resources

A Python library for research and teaching in Bayesian optimization, built so you can see how each method works, compare methods fairly, and use them in user studies. It goes from a Gaussian process written in plain NumPy to transfer, continual, multi-objective and meta-learned BO.

• Readable from scratch: a NumPy Gaussian process with RBF and Matérn-5/2 kernels and expected improvement, benchmarked against a BoTorch version of the same loop.
• Transfer, continual and meta-learned BO: reuse what earlier tasks or users taught you, with TAF, TAF+, continual BO (ConBO) with a population model, and neural acquisition functions meta-trained with reinforcement learning (NAF).
• Several objectives: correlated multi-output GPs and expected hypervolume improvement, plus TAF-EHVI, OpenBO’s own transfer method for multi-objective BO.
• Humans in the loop: almost every optimizer has a WebSocket server, so a user study, game or device evaluates the designs and OpenBO decides what to try next.
• Reproducible benchmarks: test functions, task families and archived benchmark campaigns, each with the script that reproduces it. Every optimizer shares one ask/tell interface.

git clone https://github.com/yichiliao/openbo
cd openbo
uv sync
uv run python scripts/run_benchmark.py --help