Fig. 3: Bayesian optimisation applied to a 1D test function.

a Conceptual flow diagram of Bayesian optimisation. In the deep kernel variant, the update to the surrogate model involves the joint training of the neural network and kernel parameters. b-d Optimisation iterations i = [1, 8, 11], respectively. The dashed red line represents f_true. The surrogate model (mean) is represented by the dark solid blue line and the blue shaded areas represent the model’s confidence (standard deviation). The thin solid blue lines show functions sampled from the model’s posterior distribution. The purple dots indicate where measurements of f_true have been made and the dash-dotted green line represents the acquisition function, the maximum of which (vertical dotted purple line) dictates the next sample point at iteration i + 1.