Platform For AI:How AutoML works

AutoML automates hyperparameter tuning so you can find optimal model configurations without manually running and comparing training jobs.

AutoML searches for the best hyperparameter values by running structured experiments. Each experiment spawns multiple trials, and each trial executes one or more training tasks — using the hyperparameter combination assigned to it — until a stop condition is met.

Key concepts

How it works

Configure the hyperparameter search space, search algorithm, and experiment stop conditions. AutoML creates an experiment from this configuration, then proceeds as follows:

1. The experiment generates hyperparameter combinations using the configured search algorithm and assigns one combination to each trial.

   Note

   Configure multiple trials to run concurrently to accelerate model training. Concurrency increases resource costs.

2. Each trial runs one or more training tasks based on its assigned hyperparameter combination. Tasks execute as DLC jobs on general computing resources or intelligent computing LINGJUN resources, or as MaxCompute tasks on MaxCompute computing resources. Billing, configuration, and resource usage differ between DLC jobs and MaxCompute tasks.

3. AutoML continuously monitors task metrics during the experiment.

4. The experiment stops when any of the following conditions is met: the maximum number of search combinations is reached, the algorithm's own stop condition is triggered, or all combinations have been evaluated.

5. AutoML returns results containing hyperparameter combinations or trained models for each trial. Specify a model storage path to view models. Results are also available in logs.

What to configure

Before starting an experiment, configure the following parameter groups: