Interactive explainer
Interactive: Bayesian optimisation: which experiment comes next?
Bayesian optimisation chooses new experiments using a probabilistic surrogate model. An acquisition function combines potential improvement with uncertainty so that expensive evaluations can be used selectively.
Interactive: Bayesian optimisation: which experiment comes next?
Change smoothness and the exploration incentive, then evaluate the next suggested setting. The Gaussian process is recalculated after every attempt.
Reading the display
Three starting points have already been measured. The main line shows the GP’s expected function value; dashed lines show pointwise model uncertainty. Observation points and the table contain only settings actually evaluated on the synthetic function. The thin connection between points is an orientation aid, not another predictive model.
The text reports the best measured value, the next setting and its expected improvement. Not every evaluation improves the record. A new point can be worthwhile because it reduces uncertainty in a promising region.
Three experiments
- Search step by step: evaluate five additional suggestions. Watch where uncertainty decreases and how subsequent proposals move.
- Change smoothness: reset and compare a short with a long length scale. Observations influence untested regions over different distances.
- Encourage exploration: from the same starting points, compare with . The acquisition function changes the priority of candidate evaluations.
Reset restores the three initial points, length scale 0.18 and . The button is disabled after 20 evaluations. Everything runs locally; no external experiment is triggered. The full explanation describes the kernel, expected improvement and the limits of this one-dimensional search.