Slice Chart
The Slice Chart is a scatter plot of one parameter (X axis) against one objective (Y axis), using the actual observed trials. It is a quick way to see how an objective relates to a single parameter across everything you have run. It uses no model — every dot is a real trial.
What it tells you
- The shape of the relationship between a parameter and an objective: a downward trend, a band that narrows toward good values, clusters, or no clear relationship at all.
- Where good values concentrate along a parameter's range, which hints at promising regions.
- The spread at a given parameter value: wide vertical scatter means other parameters also strongly influence the objective there.
How to read it
- Dots: each observed trial. The X position is the chosen parameter, the Y position is the chosen objective.
- Highlighted dots:
- Single-objective study: Best trials are highlighted against the rest (Trials).
- Multi-objective study: Pareto trials are highlighted against Non-Pareto ones.
- A clear downward (or upward) trend suggests the parameter matters for the objective; a flat, evenly spread cloud suggests it matters little on its own.
Controls
| Item | Description |
|---|---|
| Parameter | Choose the parameter on the X axis. |
| Objective | Choose the objective on the Y axis. |
| Log Scale | Plot the Y axis, the objective, on a logarithmic scale for positive values only. |
| Click a point | Inspect that trial's details, including parameters, objective values, and Pareto rank. |
How to read it carefully
- This is a marginal view: the other parameters are not held fixed — they vary freely from dot to dot. So a flat band does not prove the parameter is unimportant; its effect may be masked by interactions with other parameters. Cross-check with the Importance Chart.
- Because it plots only observed trials, sparsely sampled parameter ranges will simply have fewer dots — there is no interpolation or extrapolation.
Difference from PDP
- PDP Chart: fits a surrogate model and shows one parameter's averaged effect with the other variables marginalized out by the model — including extrapolation into unsampled regions.
- Slice Chart: plots the raw observed trials with no model. The other parameters are not controlled, so the scatter reflects what actually happened, not an isolated model effect.