Bayesian Optimization of Configurations
Bayesian optimization finds good designs with few expensive evaluations by modeling the objective and choosing each next point to learn the most.
When to use it
Some objectives are expensive to evaluate and have no gradient, such as the performance of a full machine configuration computed by a heavy simulation. Bayesian optimization is built for exactly this: it finds a good input with as few evaluations as possible, which matters when each evaluation is costly.
How it works
- Fit a probabilistic model, often a Gaussian process, to the results seen so far.
- Use an acquisition function to pick the next point, balancing exploring uncertain regions against exploiting promising ones.
- Evaluate the true objective there, update the model, and repeat.
Explore versus exploit
The core idea is the acquisition function, which values a candidate both for its expected performance and for how much it would reduce uncertainty. This keeps the search from prematurely converging on a local optimum while still driving toward good designs.
Applied to Kronos
For the Hyperion breeder and the burner, Bayesian optimization tunes configuration parameters where each candidate is expensive to simulate. It complements broad space-filling sweeps: the sweep maps the landscape, Bayesian optimization refines the best regions.
With surrogates
When many cheap evaluations are available through a surrogate, other optimizers may be faster; Bayesian optimization earns its place when evaluations stay genuinely expensive. The methods are chosen per problem, not by habit.
Uncertainty native
Because it is built on a probabilistic model, Bayesian optimization naturally reports uncertainty about where the optimum is, feeding uncertainty-driven decisions.