Interactive demo · Probabilistic Machine Learning: Advanced Topics · Chapter 6

Bayesian optimisation: surrogate and acquisition

An expensive function is evaluated one point at a time. After each evaluation a Gaussian process gives a mean and an uncertainty band for the function everywhere; an acquisition function turns that band into the next point to try (Algorithm 6.5). The first objective is real: the 5-fold cross-validated log-loss of logistic regression on the 569 breast-cancer cases as a function of the regularisation strength \(\log_{10}C\), tabulated on 61 values. Change the acquisition rule, the kernel length scale and the number of evaluations, and watch where the next query lands. The table compares all four rules from the same two starting points.

Objective

Acquisition function

Evaluations n

Kernel length scale ℓ (fraction of the range)

UCB multiplier β

Try these

The surrogate is a zero-mean Gaussian process with the squared-exponential kernel \(k(x,x')=\exp(-(x-x')^2/2\ell^2)\) on the range rescaled to [0, 1], observation noise \(10^{-6}\), and the observed values standardised. We maximise the negative log-loss. Each rule picks the best not-yet-queried grid point; both objectives start from the same two evaluations. Thompson sampling draws one function from the posterior with a fixed seed.

In the lecture notes