Every machine learning model has knobs that aren't learned from data — a learning rate, a regularization strength, the depth of a tree, the size of a hidden layer. Choosing good values means evaluating a model's validation loss across a "hyperparameter surface" that is expensive to sample and rarely smooth. This simulation renders that surface as a 3D landscape over two hyperparameters and lets you compare how three classic search strategies explore it.
Bergstra & Bengio's 2012 paper "Random Search for Hyper-Parameter Optimization" showed that random search often outperforms grid search with the same budget, because real loss surfaces are usually sensitive to only a few of the tuned dimensions.
A 3D validation-loss landscape over two hyperparameters — learning rate and regularization strength — that grid search, random search and Bayesian optimization each explore in real time so you can watch how their sampling patterns differ.
Grid search sweeps a fixed lattice, random search scatters trials uniformly, and Bayesian optimization uses past trials to pick each next point, trading exploration of uncertain regions against exploiting the current best basin.
Pick a search algorithm and evaluation budget, then watch trial markers land on the loss surface. Adjust ruggedness to add local minima, and tune exploration vs exploitation to see Bayesian search behave more greedily or more broadly.
Bergstra & Bengio (2012) showed random search often beats grid search at equal budget, because most hyperparameter surfaces are dominated by only a few truly important dimensions.