HPO step

Step five is HPO. Turn hyperparameter optimization on or off, then set population size, evolution cadence, tournament selection, and mutation.

For how the population, tournament selection, and mutations fit together, see How evolutionary hyperparameter optimization works.

In the wizard

The main control is Hyperparameter optimization (toggle). When enabled, Arena trains a population of agents in parallel. At intervals, agents are evaluated, the fittest are kept (tournament selection and elitism), survivors are cloned, and mutations explore new network shapes and learning hyperparameters. That happens inside one training run instead of many sequential trials.

Accordions open for:

• Evolution

• Tournament Selection

• Mutation

Field labels include Population size, Evolution frequency (steps), Evolution frequency (batches), Evolution frequency (episodes), Tournament size, and mutation-related names.

Next and Back submit the form so validation runs before you leave. On a Draft, any remaining form error blocks Next.

While a run is already going

HPO is not only a wizard concern. On the Experiments tab, the row menu (⋮) offers Update mutation parameters when status is Running. That opens the Update mutation parameters dialog. After you save, the success toast reads Mutation parameters updated for experiment ‘{name}’. Failures show Update failed with Please try again later.

That dialog is separate from Stop training, which halts the job. See Train, halt, and resume.

To pause evolution without stopping the job, set mutation probabilities to 0 for every type except None.

Next step

Finish step five, then use Train on Summary. To change mutation on a live run, use the table menu when status is Running.