Algorithms

When you create an experiment, you pick a search algorithm. The platform suggests parameter sets, runs trials, observes metrics, and suggests again until limits or goals are reached.

Recommended default

Bayesian optimization (TPE) is the default for most experiments. It balances exploration early on with focusing on promising regions as results arrive. Leave it selected unless you have a specific reason to change it.

Algorithms in the dashboard

Algorithm	Best when	Notes
Bayesian (TPE)	Continuous or mixed spaces, typical research sweeps	Default. Learns from completed trials.
Grid search	Small discrete spaces you want to cover exhaustively	Builds a grid from int/double ranges and categorical choices. Stops when the grid is exhausted.
CMA-ES	Continuous spaces where covariance-aware search helps	Uses CMA-ES sampling; configure restart and population options in the wizard when available.
Sobol sequence	Space-filling initial coverage	Quasi-Monte Carlo (Sobol) sampling for structured exploration.
Random search	Baseline comparison	Available in the wizard as a search mode.
Hyperband	Listed in the wizard	Available as a dashboard option; prefer Bayesian, Grid, CMA-ES, or Sobol for well-defined search modes today.

Start with Bayesian

For a first working experiment, keep Bayesian selected, use a narrow search space, and raise trial budget only after metrics collect correctly.

How sampling interacts with your experiment

Completed trials teach

Successful trials with objective metrics inform the next suggestions. Failed or metric-less trials are treated as failures for the study.

Limits still apply

Max trials, failure budget, parallelism, spend limit, and credit balance stop new suggestions regardless of algorithm.

Multi-objective

When you define multiple maximize/minimize objectives, the study optimizes those directions together.

Related

• Search spaces — what the algorithm is allowed to try
• Create an experiment — where you pick the algorithm
• Control an experiment — pause, resume, and stop