earlystop: Early stopping

In egenn/rtemis: Machine Learning and Visualization

Early stopping

Description

Check loss vector for early stopping criteria: - either total percent decrease from starting error (e.g. if predictions started at expectation) - or minimum percent decrease (relative to the first value of the vector) over a window of last n steps

Usage

earlystop(
  x,
  window = 10,
  window_decrease_pct_min = 0.01,
  total_decrease_pct_max = NULL,
  verbose = TRUE
)

Arguments

x	Numeric vector: loss at each iteration
window	Integer: Number of steps to consider
window_decrease_pct_min	Float: Stop if improvement is less than this percent over last window steps
total_decrease_pct_max	Float: Stop if improvement from first to last step exceeds this percent. If defined, overrides window_decrease_pct_min
verbose	Logical: If TRUE, print messages to console. Default = TRUE

Details

If the first loss value was set to be the loss when yhat = mean(y) (e.g. in boosting), then total_decrease_pct_max corresponds to R-squared and window_decrease_pct_min to percent R-squared improvement over window last steps.

Author(s)

E.D. Gennatas

egenn/rtemis documentation built on Nov. 22, 2024, 4:12 a.m.