Training one machine learning model requires choosing dozens of settings — learning rate, tree depth, regularisation, number of layers — and the "best" combination is different for every dataset. AutoML automates that search. This lab renders the space of possible hyperparameter combinations as a 3D terrain, where height is validation accuracy: tall gold peaks are great configurations, low blue-indigo valleys are poor ones. Glowing trial agents are AutoML jobs testing configurations, hunting for the highest peak without ever seeing the whole map in advance.
On rugged, multimodal landscapes, purely random search often beats a naive grid of the same size — a result formalised by Bergstra & Bengio (2012) — which is part of why modern AutoML tools favour randomised and Bayesian strategies over exhaustive grids.
Glowing search agents explore a 3D accuracy landscape shaped by two hyperparameters, hunting for the highest peak the way AutoML tools like random search, grid search and Bayesian optimisation hunt for the best model configuration.
The terrain's height is validation accuracy across a learning-rate and model-depth grid, with gold marking the global optimum and lower peaks acting as misleading local optima. Each trial agent trains and evaluates a configuration, then reports its score before choosing the next one.
Switch between random, grid and Bayesian-style search to compare coverage patterns, add more parallel trials, speed up the search, and raise landscape ruggedness to add more deceptive local optima. Watch the gold beacon jump each time a trial beats the current best.
Bergstra & Bengio's influential 2012 paper showed that plain random search often outperforms grid search of the same budget on high-dimensional problems — a key reason modern AutoML frameworks default to randomised or Bayesian strategies instead.