Featureless LCE Learner

Baseline LearnerLCE which ignores the batch_nr value and predicts a constant performance for every future batch. The constant is one of three summaries of the per-batch training performances, selected via type:

• "average": the mean (or median, when robust = TRUE) of the considered training batches. The "no learning-curve information used at all" sanity-check baseline.

• "best": the best performance observed among the considered training batches. The "no progress beyond the best we have already seen" baseline. The optimization direction (whether higher or lower is better) is taken from the task's measure, so the task must carry one for this type.

• "last": the performance of the most recent batch. The "no progress from here" baseline.

The window parameter restricts "the considered batches" to the most recent window of them (all of them by default). Location and dispersion are computed on the task's lce_link scale. When predict_type = "se", se_epistemic is the (robust) link-scale dispersion of the considered per-batch performances divided by the square root of their count (the standard error of the constant), and the total predictive se adds that dispersion back as the aleatoric spread (se = sqrt(dispersion^2 + se_epistemic^2)), both replicated for every prediction. For type = "average" this is the textbook predictive standard deviation of a new observation; for type = "best" / "last" the constant is not the window mean, so the symmetric interval it implies is only a coarse heuristic. The uncertainty reflects nothing beyond the spread of the observed curve.

Creates a new instance of this learner.

Parameters

• type :: character(1)
"average", "best", or "last". Initialized to "best".

• robust :: logical(1)
  When TRUE, use median + MAD instead of mean + SD. Affects the "average" location and the dispersion of all types. Initialized to FALSE.

• window :: integer(1)
  Number of most recent batches considered, clamped to the number of training batches. Inf (the default) uses all batches. For type = "last" only the se is affected, since the location is always the most recent batch.