Response Surface 3D

The Response Surface 3D widget renders the raw predicted surface of a trained surrogate over two chosen parameters, sliced through an anchor design point. It is the direct visualization counterpart of what the Surrogate Optimizer and Robustness widgets output numerically.

Slice, not marginal

The 2D PDP widget shows a marginal effect: at each grid point of the two selected parameters, the remaining parameters are averaged out. This widget shows a conditional slice: the remaining parameters are fixed at the anchor point's values, so the surface is exactly the model's prediction along a 2-parameter plane through that design:

$$z(x_1, x_2) = \hat{f}(x_1, x_2, \mathbf{x}_{\text{anchor}, \setminus {1,2}})$$

The two answer different questions. The PDP is the better tool for "how does this parameter act in general"; the slice is the better tool for "what is the local shape of the objective around my candidate" — for example, whether there is still a direction for improvement.

Caveats

• Surface validity is limited by surrogate quality. Check fit quality before reading local geometry (for example, with the Surrogate Optimizer validation plot).
• Different anchors produce different slices. When comparing neighborhoods of two candidates, keep axis parameters identical and switch only the anchor.
• Axis parameters must be numeric. Categorical parameters cannot form the slice plane and are frozen at the anchor values, like other non-axis parameters.