Experiment wizard overview

New and draft experiments use a six-step flow. The header shows New experiment and a horizontal stepper; the body swaps content as you move.

Steps (in order)

1. Resources: nodes and resource class.

2. Environment: gym (classic) or gym / dataset (Advanced Training).

3. Agent: algorithm and model or network.

4. Training: rollout length, batch sizes, and related fields.

5. HPO: hyperparameter optimization.

6. Summary: read-only review; Train or Resume here.

Classic vs Advanced Training

When you created the project, the Training switcher set the type:

• Classic RL: gym environment grid, standard Agent form, classic Training accordions.

• Advanced Training: combined environment grid (datasets and RL Environment rows), Pipelines tab on the project, and LLM-oriented agent and training UI where the flow needs it.

See Projects.

Navigation

Back and Next appear on every step except Summary, which has Back plus Train or Resume.

On a Draft, Next does not advance if something is missing. You get a toast with the reason (no resource class, no environment, no algorithm, form errors, reward validation not finished, bad prompt template, and similar).

On a non-draft experiment you can click step names to review configuration.

Save appears only on Summary (with Train or Resume) for Draft experiments that already have a resource class. It keeps the draft without scheduling training. Mid-steps use Next and idle autosave instead.

After Resources, you can change compute later with Edit resources (opens Select training cluster) from later wizard steps.

A draft also saves on its own after about five minutes of idle activity (mouse, keyboard, scroll, or click). Success toasts look like Updated draft experiment: {name} or Created draft experiment: {name}.

Summary step

On Summary for a draft, Train schedules the run. Tooltip: training begins shortly when resources are ready, usually within about ten minutes. Status becomes Pending first; allow up to roughly ten minutes for the job to move to Running when the cluster is busy. Success toast: Experiment {name} scheduled for training…

If the experiment is no longer a Draft, Resume replaces Train (unless status is Running or Pending). Resume opens a dialog to pick a checkpoint and name a new experiment.

After a successful Train from the wizard, Arena returns you to the project Experiments tab and the stepper resets to Resources.

Step guides

Wizard step	Guide
Resources	Choosing compute
Environment	Environment step
Agent	Agent step
Training	Training step
HPO	Hyperparameter search

After training, see Experiment statuses.