Statistical Methods for Learning Curves and Cost Analysis
Published Date: March 1, 2003
Abstract:D6870 This monograph examines the statistical methods that have been used to estimate the learning-rate parameter in learning curves, as well as the methods used to calibrate cost-estimating relationships (CERs). We argue that some widely-used methods, such as lot-midpoint iteration, do not have either a strong mathematical or statistical justification. The properties of other, general-purpose methods, such as non-linear least squares and iteratively reweighted least squares, are well established in the statistics literature. These latter methods can be applied to learning curve and CER estimation, and possess stronger properties than the heuristic methods traditionally applied by cost analysts.
