regularization: Implements the regularization functions used in the...

In blockcpd: Change Point Detection for Multiple Aligned Independent Time Series

Implements the regularization functions used in the estimation

Description

The estimator in this package computes the optimum of -l(C, p) + λ*R(leftIndex, rightIndex, nrow, ncol), where l is the log likelihood of the family, lambda is the penalization constant and R is the regularization function. The user can create his own regularization function and pass as an argument to fit_blockcpd. It should have four arguments, in the following order: left_index, right_index, nrow and ncol. Each argument is explained in the parameter section. If the function depends on leftIndex and rightIndex, it will be non-homogeneous, which might be interesting in some applications. The package implements some functions as an example, but uses only bic_loss as the default. The algorithm is consistent as long as the the regularization is bounded by a constant.

Usage

toy_regularization(left_index, right_index, nrow, ncol)

Arguments

left_index	First index of the interval
right_index	Last index of the interval
nrow	Number of rows/signals/series
ncol	Number of columns/variables

Examples

my_reg <- function(leftIndex, rightIndex, nrow, ncol){
    block_size = (rightIndex - leftIndex + 1)
    return(log(nrow*ncol)*(1/block_size))
}

blockcpd documentation built on Aug. 12, 2022, 5:07 p.m.