AutoML Lab 2D: Grid vs Random vs Real Gaussian-Process Search

This 2D companion pits the same three hyperparameter-search strategies against each other on a rippled synthetic loss landscape — a bowl toward a hidden global minimum overlaid with sinusoidal traps — but runs them step-by-step in real time on one shared view instead of three separate static panels. Grid search lays a fixed lattice before seeing any result, random search samples uniformly each trial, and the Bayesian optimizer fits an exact Gaussian-process posterior (RBF kernel, Cholesky-solved, not a heuristic shortcut) after every trial and chooses its next point by minimizing a lower-confidence-bound acquisition function that explicitly balances predicted loss against remaining uncertainty. A live convergence chart tracks each strategy's best loss found against evaluation count as the run plays out, so the exploration-exploitation trade-off is visible while it happens rather than only in the final tally.