Real machine-learning models are rarely tuned for a single number. A model can almost always be made more accurate at the cost of extra latency or a larger footprint, and vice versa. Each glowing point in this scene is a candidate model configuration plotted by three competing objectives — validation error, inference latency, and model size — all of which we want to minimize simultaneously.
score = w₁·error + w₂·latency + w₃·size. The point on the Pareto front with the lowest score is highlighted gold and linked to the origin — this is the "best" model for the priorities you set.The concept is named after economist Vilfredo Pareto, who used it to describe allocations where no one can be made better off without making someone else worse off. In multi-objective hyperparameter tuning, algorithms like NSGA-II and Bayesian multi-objective search try to approximate this entire front in one run, rather than collapsing it to a single scalarized number in advance.
A 3D Pareto front of candidate model configurations trading off validation error, inference latency, and model size — the core geometry behind multi-objective hyperparameter tuning.
Bright indigo points are Pareto-optimal: no other candidate beats them on every objective at once. Dim grey points are dominated — some other configuration is strictly better in every dimension.
Adjust the weight sliders to change how much error, latency, and size matter; the gold marker jumps to the front point that minimizes your weighted score. Regenerate the population to sample a new trade-off surface.
Algorithms like NSGA-II approximate the whole Pareto front in one optimization run, letting engineers pick a trade-off after seeing all the options — rather than committing to fixed weights in advance.