EDF
Overview
The EDF (Empirical Distribution Function) chart shows what fraction of completed
trials achieved an objective value at or below each point on the X axis. It is
Optuna's plot_edf equivalent: instead of looking at objective value versus
trial order (as Optimization History does), it looks at the overall shape of the
distribution of results, independent of when each trial ran.
Controls
| Item | Description |
|---|---|
| Objective (multi-objective studies) | Use the dropdown to select which objective's distribution to display. |
| Log Scale | Toggle log-scale on the value (X) axis. Trials with a non-positive objective value are dropped from the curve while log scale is active, because they cannot be represented on a log axis. |
Operations
Pan, zoom, reset, and legend toggling (to show/hide a study's curve when comparison studies are loaded) follow the standard 2D chart controls.
| Item | Description |
|---|---|
| Hover | Hover near the curve to see a tooltip with the objective value and the cumulative fraction of trials at that point. |
How to Read
- X axis: the objective value. Y axis: the fraction (0 to 1) of completed trials whose objective value is at or below that X value — this is the empirical CDF.
- Steep, narrow curve: most trials converge to a similar objective value — the search is well-behaved and consistent, or the search space is not very sensitive to the sampled hyperparameters.
- Shallow, wide curve: objective values are spread out over a broad range — trials vary widely in quality, suggesting high sensitivity to hyperparameters or an unstable optimization process.
- Left-shifted curve (for a minimization objective): the study as a whole found better values more often — comparing two studies' EDFs this way tells you which sampler/configuration produced better results overall, not just in the single best trial.
- Right-shifted curve: worse values are more common in that study.
- Comparing studies: overlaying the base study and one or more comparison studies makes it easy to see which one has a systematically better (further left, for minimization) distribution of outcomes across the whole run, not just at the single best value.
- Flat sections: no trials landed in that value range — often visible as gaps or plateaus in the staircase.
- Vertical jumps: many trials share (or are very close to) the same objective value at that point — common with discrete/categorical search spaces or with a sampler that repeatedly proposes similar configurations.