Fits a monotone, piecewise-constant learning curve via isotonic regression
(stats::isoreg()) to the per-batch surrogate performance. With
direction = "auto" (the default), the direction is inferred from the task's
measure: utility-style targets (R², accuracy, ...) increase and loss-style
targets (MAE, RMSE, ...) decrease. Set direction = "increasing" or
"decreasing" to override this; the implementation negates the target before
fitting for the decreasing case.
Predictions outside the observed batch range are extrapolated as
constants of the closest endpoint fit, matching the monotone-shape
constraint. Within the observed range either piecewise-constant
(step function) or piecewise-linear interpolation is used, controlled
by interpolation. The fit is performed on the task's lce_link scale.
When predict_type = "se" the epistemic se_epistemic is a constant
link-scale residual standard deviation divided by sqrt(n_batches), and the
total predictive se adds that residual standard deviation back as the
aleatoric spread. The residual standard deviation uses n_batches - 1 degrees
of freedom (it does not discount the degrees of freedom the monotone fit
itself consumes) and is constant across batches, so the uncertainty is a
coarse estimate that ignores how the fit degrades away from the observed
batches.
Creates a new instance of this learner.