AutoMAX Search

AutoMAX uses SMAC3 — a Bayesian optimization framework — to search over hyperparameter configurations for your chosen loss/optimizer pair.

How It Works

1. AutoMAX reads your config and builds a configuration space from the predefined search space of the selected loss function.

2. It starts with the default configuration (via DefaultInitialDesign) to establish a baseline.

3. SMAC3 proposes new configurations, trains a full model for each trial, and uses the returned target metric to guide the next suggestion.

4. After n_trials evaluations, the best configuration is saved to output_directory.

Resuming a Search

SMAC3 serializes the run history to disk. If a search is interrupted, it can be resumed by re-running the same command — as long as output_directory and name remain unchanged and overwrite: false.

automax:
  name: my_search
  output_directory: ./automax_output
  overwrite: false   # don't wipe previous trials

What Gets Tuned

AutoMAX tunes the loss and optimizer hyperparameters defined in each loss function’s search space. Common parameters across all pairs include:

• lr — learning rate (log-scale range)

• epoch_decay — learning rate decay factor

• weight_decay — L2 regularization

• margin — surrogate loss margin

Loss-specific parameters (e.g. beta, eta, gamma, Lambda, tau) are included automatically when that loss is selected.

See Loss / Optimizer Pairs for the full list per loss function.

Output

After the search completes, results are written to output_directory/name/:

automax_output/
└── my_search/
    ├── runhistory.json  #
    ├── configspace.json  #
    ├── intensifier.json  #
    ├── optimization.json #
    └── scenario.json #