Quickstart¶
This walkthrough follows the in-app new experiment tutorial: a Classic RL project, one experiment, default settings, train, then view results. Allow about 15–20 minutes if compute queues quickly.
1. Sign in and open Experiments¶
After login you are on Home. In the left sidebar, click Experiments.
2. Create a project¶
On the Experiments page, create a new project:
Name and Description: any values you like.
On the Training switcher, leave Classic RL selected (the default).
Confirm creation. The app opens the new project page.
3. Create an experiment¶
Click New experiment, name it, save. The experiment setup wizard opens with a horizontal stepper in the header: Resources → Environment → Agent → Training → HPO → Summary.
4. Resources¶
Pick a compute resource class from the radio list. The tutorial suggests Arena Medium when your org offers it. Each option shows specs and an hourly credit rate. Select one, then Next. Change it later with Edit resources (dropdown in the Select training cluster modal).
GPU-backed classes cost more credits per hour than CPU-only ones. Your remaining balance appears in the profile menu (progress bars) and on Usage for Managers.
5. Environment¶
Arena ships with many built-in Gym environments. Pick any environment for this walkthrough. Validation summary, rollout reward charts, and an optional render appear on the same step when data exists. Click Next to go to Agent.
For your own code, use Environments in the sidebar and the custom-environment tutorial instead.
6. Agent (algorithm)¶
Choose an algorithm (the tutorial uses PPO). You can edit network and algorithm hyperparameters; the defaults are enough here. Next.
7. Training and HPO¶
On Training, leave step counts and related settings as they are. Next.
On HPO (hyperparameter optimization), evolutionary search is optional. Leave the defaults and Next. See How evolutionary hyperparameter optimization works if you want the full picture before tuning mutation settings.
8. Train¶
On Summary, review the configuration and click Train. The experiment appears in the project’s Experiments table with a live status.
Training debits credits while the job runs. If the balance hits zero, new runs may be blocked until you top up or upgrade (see Credits and plans).
9. Results¶
Open the Results tab for the project (the tutorial highlights this tab at the end). Select your experiment to inspect metrics, logs, and checkpoints when the run finishes.
Statuses such as Running, Succeeded, and Failed are defined in Experiment statuses.
Optional: custom environment tutorial¶
From the sidebar, open Environments and follow the Upload a custom environment tutorial: upload code, pick an entrypoint, validate, save, then launch an experiment from the environment page.
CLI¶
To run jobs from a terminal instead of the UI, create a CLI API key on Profile management and configure the Arena Python CLI. See Arena CLI.