A ≺ B ⇔ ∀i fᵢ(A) ≤ fᵢ(B) ∧ ∃j fⱼ(A) < fⱼ(B)
Every glowing sphere is one automated hyperparameter trial, plotted in 3D by three competing objectives: model size (left → right), accuracy (bottom → top) and training time (back → front). Because shrinking the model tends to hurt accuracy and speed things up at the same time, there is no single "best" trial — only a frontier of trade-offs.
A trial A is Pareto-optimal when no other trial beats it on every objective at once. Formally, A dominates B when:
∀i: fᵢ(A) ≤ fᵢ(B) and ∃j: fⱼ(A) < fⱼ(B)
Trials that are dominated by at least one other trial (dimmer blue) are discarded in favor of the non-dominated set (bright gold), which is connected into the visible frontier. The Hypervolume stat is the normalized area of the accuracy/size plane swept out below that frontier — it only grows when a trial pushes the trade-off outward, so it is a single number that tracks overall search progress.
Search strategy select (random / Bayesian-style / evolutionary), trial-rate slider, show-dominated toggle, show-frontier toggle, play/pause, reset
Real AutoML systems such as Optuna and Google Vizier use exactly this idea — instead of chasing one metric, they maintain a Pareto front of configurations (e.g. accuracy vs. inference latency vs. model size) so an engineer can pick the trade-off that fits their deployment target after the search finishes, rather than before it starts.
An automated tuner searches a live hyperparameter space in 3D — every trial is plotted by accuracy, model size and training time, and the Pareto-optimal frontier of best trade-offs emerges as random search, Bayesian-style exploitation or evolutionary refinement run.
Real hyperparameter tuning rarely optimizes just one number. This simulation plots every automated trial in a 3D objective space — model size, accuracy and training time — and highlights the Pareto-optimal frontier: the trials no other trial beats on every objective at once. Watch how the frontier expands and its shape changes as trials accumulate.
Pick a search strategy (random search, Bayesian-style exploitation, or evolutionary mutation of the current frontier), adjust the trial rate, and toggle whether dominated trials and the frontier line are shown. Drag to orbit the objective-space cube and scroll to zoom.
Real AutoML systems such as Optuna and Google Vizier use exactly this idea — instead of chasing one metric, they maintain a Pareto front of configurations (e.g. accuracy vs. inference latency vs. model size) so an engineer can pick the trade-off that fits their deployment target after the search finishes, rather than before it starts.