powerlaw

Power law function for modelling learning curve (taken to mean change in expected loss per sample with training set size)
Recommended in <a href="https://arxiv.org/abs/2103.10948">review of learning curve forms</a>
If <code>theta=c(a,b,c)</code> then models as <code>a n^(-b) + c</code>. Note <code>b</code> is negated.
Note that <code>powerlaw(n,c(a,b,c))</code> has limit <code>c</code> as <code>n</code> tends to infinity, if <code>a,b &gt; 0</code>

Predictive scores must be updated with care, because actions taken on the basis of existing risk scores causes bias in risk estimates from the updated score. A holdout set is a straightforward way to manage this problem: a proportion of the population is 'held-out' from computation of the previous risk score. This package provides tools to estimate a size for this holdout set and associated errors. Comprehensive vignettes are included. Please see: Haidar-Wehbe S, Emerson SR, Aslett LJM, Liley J (2022) <doi:10.48550/arXiv.2202.06374> (to appear in Annals of Applied Statistics) for details of methods.

James Liley

OptHoldoutSize

Estimation of Optimal Size for a Holdout Set for Updating a
Predictive Score

Sami Haidar-Wehbe

Sam Emerson

Louis Aslett

powerlaw function

<dl><dt>n</dt>
<dd>Set of training set sizes to evaluate</dd>
<dt>theta</dt>
<dd>Parameter of values</dd></dl>

Arguments

Power law function for modelling learning curve (taken to mean change in expected loss per sample with training set size)
Recommended in <a href='https://arxiv.org/abs/2103.10948'>review of learning curve forms</a>
If <code>theta=c(a,b,c)</code> then models as <code>a n^(-b) + c</code>. Note <code>b</code> is negated.
Note that <code>powerlaw(n,c(a,b,c))</code> has limit <code>c</code> as <code>n</code> tends to infinity, if <code>a,b &gt; 0</code>

Power law function — powerlaw

<dl>

<dt>n</dt>
<dd>Set of training set sizes to evaluate</dd>


<dt>theta</dt>
<dd>Parameter of values</dd>

</dl>

powerlaw: Power law function

Description

Usage

Value

Arguments

Examples