Intermediate Values
Overview
The Intermediate Values chart overlays the learning curve of every trial that reported intermediate values during optimization — the same values samplers and pruners see as a trial progresses (e.g. validation loss per epoch, per boosting round, per environment step). It lets you compare how trials evolve over their reporting steps and see, at a glance, which ones were cut short by pruning.
Controls
| Item | Description |
|---|---|
| Log Scale | Toggle log-scale on the value axis. Points with a non-positive value are dropped from a curve while log scale is active, because they cannot be represented on a log axis. |
Operations
Pan, zoom, and reset follow the standard 2D chart controls.
| Item | Description |
|---|---|
| Hover | Hover near a curve to highlight it and see a tooltip with the trial number, state, step, and value at that point. Other curves are dimmed while one is highlighted. |
How to Read
- Each line is one trial: the X axis is the reporting step (e.g. epoch or boosting round) and the Y axis is the intermediate value at that step.
- Color by trial state: complete trials, pruned trials, running trials, and failed trials are colored differently (see the legend). Pruned curves typically stop earlier than complete ones — that is the pruner cutting off an unpromising trial.
- Tight bundle of curves: most trials follow a similar trajectory, suggesting the search space is not very sensitive to the sampled hyperparameters, or the model/data dominates the outcome.
- Wide spread, especially early: hyperparameters strongly affect the learning trajectory; pruning based on early steps is likely to be effective.
- Many short pruned curves: the pruner is aggressively cutting unpromising trials early, which speeds up the search — as long as it is not also cutting trials that would have improved later.
Performance note: if the study has more than 2000 trials with intermediate values, the chart evenly subsamples down to 2000 curves and shows a note indicating how many trials are displayed.