Create an experiment

An experiment connects an image from your org Images registry to a search space, objectives, and execution limits. Once launched, HyperOptimizer schedules trials and ranks results in the dashboard.

Wizard flow

Open Dashboard → Experiments → New.

1. Pick an image from your org registry (repository + tag). Publish first if nothing is listed — see Publish an image.
2. Edit the command in the workbench. The platform seeds from the image CMD when available. Managed parameters are appended as --hpo-<slug> by default, or plain --<slug> if you switch flag mode. You can use {{slug}} placeholders in a custom template.
3. Define parameters — names (slugs), types, and ranges or choices. See Search spaces.
4. Set objectives — one or more metrics to maximize or minimize. Names must match hpo.metrics.<key> lines. Optional goals and conditions are available in the UI.
5. Configure run settings — parallelism, max trials, max failed trials (failure budget), max runtime per trial, machine size, and an optional experiment spend limit in €.
6. Add environment variables (optional) — up to 32 custom vars. Names starting with HYPEROPTIMIZER_ are reserved for the platform.
7. Choose an algorithm — Bayesian (TPE) is the recommended default. See Algorithms.
8. Launch with desired status running (or create paused if you prefer to start later).

Fields summary

Field	Required	Description
Image	Required	Org Images repository and tag (Enterprise may allow custom Docker outside the org registry).
Command	Optional	Override or extend the image command; managed params appended or substituted via {{slug}}.
Parameters	Required	Search space passed as CLI flags.
Objectives	Required	Metric keys to maximize or minimize.
Parallelism	Optional	Concurrent trial containers (capped by plan).
Max trials / failed / runtime	Optional	Execution limits; failed-trial cap is your failure budget.
Machine size	Optional	Compute SKU — see Machines and limits.
Spend limit	Optional	Soft € cap for this experiment.
Env vars	Optional	Custom environment; platform injects trial/experiment/org IDs.
Algorithm	Optional	How the next parameter sets are chosen.

Platform-injected environment

Every trial receives:

HYPEROPTIMIZER_TRIAL_ID
HYPEROPTIMIZER_EXPERIMENT_ID
HYPEROPTIMIZER_ORG_ID

Do not define custom variables with the HYPEROPTIMIZER_ prefix.

Recommended first run

Start with a narrow search space and a small trial budget. The first goal is to validate the contract: arguments arrive, the workload runs, and metrics are collected.

1. Narrow the space: Use two or three parameters, not your entire strategy surface.
2. Pick a reliable metric: Choose a metric that always prints for successful runs.
3. Keep parallelism low: Start with low parallelism until data and image access are confirmed.
4. Use Small machines: Confirm cost and runtime before scaling up.

Tip

Presets such as Free Evaluation, Starter Optimization, and Pro Optimization prefill capacity within your plan. You can still edit every field.

After launch

Watch the first few trials on the experiment page. Pause, resume, or stop from there — see Control an experiment. If trials fail, read logs and check Troubleshooting.