Each gold dot is a hyperparameter trial (learning rate, regularisation) actually run. A Gaussian-process (GP) regression fits a posterior mean and variance over the whole grid from those trials — the heatmap colour is the predicted validation score, and a dark "fog" overlay thickens wherever the model is still unsure. The Upper Confidence Bound acquisition (mean + κ·std) picks the pulsing blue point as the next trial to run: raise κ to favour exploring foggy regions, lower it to exploit the best-known peak.
- Heatmap colour — GP posterior mean validation score (blue → red, low → high).
- Fog overlay — posterior standard deviation; darker = less certain.
- Gold dots — trials actually run, sized by how recent.
- Pulsing ring — the acquisition function's current top pick.
- Show true accuracy landscape — swaps the heatmap for the hidden ground-truth function the GP is trying to learn.
- Show acquisition surface — swaps the heatmap for mean + κ·std directly, so you can see exactly what the optimiser is maximising.