Rank Plot
Overview
The Rank Plot is Optuna's plot_rank equivalent. It is a scatter plot of two
selected parameters where each completed trial is colored not by its raw
objective value, but by its rank — the percentile position of its objective
value among all completed trials (0 = best, 1 = worst). Coloring by rank instead
of raw value makes the plot robust to outliers and highlights how the
combination of two parameters relates to overall trial quality.
Controls
| Item | Description |
|---|---|
| X / Y parameter selectors | Choose which two parameters to plot against each other. Defaults to the first two parameters in the study. |
| Objective selector (multi-objective studies) | Choose which objective's rank to color by. |
| Fewer than 2 parameters | The chart shows a message instead of a plot, because a rank plot needs two axes. |
Operations
Pan, zoom, reset, and click-for-detail follow the standard 2D chart controls.
| Item | Description |
|---|---|
| Hover | Hover over a point to see the trial number, the X/Y parameter values, the objective value, and the rank percentile. |
How to Read
- Color scale: points colored toward the "Best" end of the legend have objective values near the top of the study (best observed outcomes); points toward "Worst" have objective values near the bottom. The direction (minimize/maximize) of the selected objective is already accounted for — "best" always means best, regardless of direction.
- Why rank instead of raw value: a handful of extreme outliers can wash out a raw-value color scale, making most of the plot look like a single flat color. Ranking removes that problem — colors are always spread evenly across the full range of trials, so subtle structure is visible even when a few trials have unusually extreme results.
- Clusters of "Best"-colored points in a region: that combination of the two parameters tends to produce good outcomes — a promising region of the search space.
- Clusters of "Worst"-colored points in a region: that combination tends to produce poor outcomes — worth avoiding or investigating why.
- No visible pattern / colors evenly mixed everywhere: the two selected parameters (as a pair) don't strongly determine the outcome on their own — other parameters, or interactions with a third parameter, may dominate. Try other parameter pairs, or check the Importance Chart to see which parameters matter most.
- Diagonal or curved gradients: suggests an interaction effect between the two parameters — the optimal value of one depends on the value of the other.
- Categorical parameters: shown using their encoded category index on the axis, same as other scatter-based widgets (Slice Chart, Scatter Matrix).