Search Space Analysis: Hyperparameter Sampling Strategies (2D)

Search Setup

Search strategy

Sample budget 120

Landscape ruggedness 55%

Sampling speed 12/s

Show contour lines

Live Readout

Samples evaluated

Best loss found

—

Best (x1, x2)

—

Eval rate

0/s

How it works

The heatmap is a 2D loss surface loss(x1, x2) — a convex bowl overlaid with tunable sine-wave ruggedness, exactly the same closed-form function driving the 3D version's mesh. Teal is low loss (good), magenta is high loss (bad). Each strategy proposes the next (x1, x2) to evaluate differently:

Grid — a uniform √budget × √budget lattice; exhaustive, but cost explodes with more dimensions.
Random — uniform random draws; often beats grid in higher dimensions.
Latin hypercube — each axis stratified into equal bins then shuffled, so samples spread evenly with less clustering than pure random.
Bayesian-guided — starts broad, then narrows a shrinking search radius around the current best point as the budget is spent (a simplified stand-in for a surrogate-model acquisition function).

The convergence chart on the right plots the running best loss against samples evaluated — the standard way to compare sampler efficiency: a strategy that reaches a low value with fewer samples is more sample-efficient.

Low loss High loss Best found