Parameter Sweep¶

CORNET's built-in sweep engine expands a cartesian product of parameter axes across one or more runs, writing each variant to its own output directory and leaderboard entry.

1. Add a `sweep` block¶

In your task's config.yaml, add a sweep key under experiment:

experiment:
  name: my_sweep
  duration: 60.0
  output_dir: results
  primary_metric: mean_aoi_ms
  higher_is_better: false

  sweep:
    axes:
      network.scenario.numerology: [1, 2, 3]
      network.scenario.bandwidth_mhz: [20, 40]
    repeats: 2

This expands to 3 × 2 × 2 = 12 runs.

2. Axis key format¶

Axis keys are dot-path references to fields in UnifiedConfig. Examples:

Key	Field
`network.scenario.numerology`	`config.network.scenario.numerology`
`network.scenario.bandwidth_mhz`	`config.network.scenario.bandwidth_mhz`
`experiment.duration`	`config.experiment.duration`
`middleware.rtf`	`config.middleware.rtf`

The key must resolve to a scalar field (int, float, str, bool). Nested objects and lists are not supported as sweep values.

3. Variant ID naming¶

CORNET generates a variant_id string for each expanded variant:

Axes	Repeats	Example variant ID
1 axis	1	`"2"` (the value itself)
2+ axes	1	`"numerology=2_bandwidth_mhz=40"`
any	> 1	`"numerology=2_bandwidth_mhz=40_run1"`

Variant IDs are used as: - output_dir suffix: results/numerology=2_bandwidth_mhz=40_run1/ - variant_id field in leaderboard.json

4. Output directory layout¶

results/
  numerology=1_bandwidth_mhz=20_run1/
    analysis/
      aoi_statistics.json
  numerology=1_bandwidth_mhz=20_run2/
    analysis/
      aoi_statistics.json
  numerology=1_bandwidth_mhz=40_run1/
  ...
tasks/my_task/leaderboard.json    ← one entry per variant

5. Running a sweep¶

python -m cornet tasks/my_task

The sweep runner calls expand_sweep(config) then iterates through each variant sequentially. Failed variants are logged and skipped; the sweep continues.

6. Viewing sweep results¶

python -m cornet view tasks/my_task

The leaderboard table shows one row per variant, sorted by primary metric. Use the web UI for charts across sweep dimensions:

python -m cornet ui tasks/my_task

7. Programmatic sweep expansion¶

from cornet.config.loader import load_config
from cornet.sweep.expander import expand_sweep

config = load_config("tasks/my_task/config.yaml")
variants = expand_sweep(config)
for v in variants:
    print(v.experiment.name, v.network.scenario.numerology)

expand_sweep returns a single-element list (original config, name = "default") when no sweep is defined.

8. Reference: SweepConfig schema¶

Field	Type	Default	Description
`axes`	`dict[str, list]`	`{}`	Mapping from dot-path key to list of values.
`repeats`	`int`	`1`	Number of repeated runs per parameter combination.

See config-schema.md for the full schema.