Skip to contents

Residual Bootstrap LCE Learner Wrapper

Source: R/LearnerLCEBootstrap.R
mlr_learners_lce.bootstrap.Rd

Wraps an arbitrary base LearnerLCE and equips it with an SE prediction via residual bootstrap on the per-batch fit.

Training proceeds as:

  1. Fit the base learner once on the training task; collect the per-batch fitted values and the residuals on the lce_link scale (residual = g(truth) - g(fitted)).

  2. For each of n_bootstrap replicates, resample the link-scale residuals with replacement, add them to the fitted values, back-transform to the natural scale (so the synthetic targets respect the link's support), and refit the base learner on a clone of the training task.

Each replicate is a full curve, so the replicates are draws of the mean curve f(b). The point forecast response is their link-scale mean, back-transformed (the predictive median under the link). The epistemic se_epistemic is the link-scale spread of the replicate mean curves; the total predictive se adds the aleatoric residual variance back. The samples, quantiles, and target_reached predict types describe the realised future performance y_b: each draw is a replicate mean curve plus a resampled link-scale residual, so they reflect both the bootstrap mean-curve uncertainty and the residual noise.

Residual bootstrap is the natural choice here because every batch contributes a single per-batch performance value: row-bootstrap of the archive would simply reweight identical rows, while batch-bootstrap with multiplicity would not be picked up by base learners that aggregate by batch before fitting.

Creates a new bootstrap-wrapped LCE learner.

Arguments

learner

(LearnerLCE)
Base LCE learner to wrap. Its predict type is forced to "response" while the bootstrap replicates are fitted.

Parameters

The bootstrap wrapper's own parameters are exposed alongside the base learner's parameters via paradox::ParamSetCollection; the base learner's parameters carry the base. prefix (e.g. base.rate_lower).

Own parameters:

  • n_bootstrap :: integer(1)
    Number of bootstrap replicates. Initialized to 100.

  • seed :: integer(1) | NULL
    RNG seed for the residual resampling – both the bootstrap refits at train time and the realised-trajectory draws at predict time, so a fixed seed makes the whole prediction reproducible. If NULL (default), the current RNG state is used.