Monotone Spline LCE Learner

Fits a shape-constrained smoothing spline (monotone increasing or monotone decreasing) to the per-batch surrogate performance via scam::scam() with a monotone P-spline basis ("mpi" / "mpd"). With direction = "auto" (the default), the direction is inferred from the task's measure: utility-style targets increase and loss-style targets decrease. This is the smoother, lower-bias analogue of LearnerLCEIsotonic.

Extrapolation beyond the training-batch range linearly extends the last fitted spline segment, so very long-horizon forecasts can drift; use LearnerLCEParametricExponential or one of the parametric families when a hard asymptote is needed. The spline is fit on the task's lce_link scale.

At least five distinct batches are required: scam's monotone basis needs a basis dimension k of at least 4, and k is clamped to at most n_batches - 1 to keep the fit identifiable.

When predict_type = "se" the learner returns the spline's pointwise standard errors (as reported by scam::predict.scam()) as the epistemic se_epistemic and adds the spline's residual variance back to form the total predictive se, both on the link scale.

Creates a new instance of this learner.

Parameters

• direction :: character(1)
"auto" (default), "increasing", or "decreasing". Selects the monotone P-spline basis (mpi / mpd) after resolution.

• k :: integer(1) | NULL
  Basis dimension passed to scam::s(). Defaults to NULL, which uses min(10, n_batches - 1). Floored at 4 (the smallest k the monotone basis supports) and clamped to at most n_batches - 1 at training time to keep the fit identifiable.

• bs :: character(1)
  Spline basis. Restricted to "mpi" (monotone increasing) and "mpd" (monotone decreasing) and is set automatically from direction.