No subject

[Bill Macready]

We would like to announce a paper entitled:

    An Efficient Method To Estimate Bagging's Generalization Error

                   D.H. Wolpert, W.G. Macready


In bagging one uses bootstrap replicates of the training set to try to
improve a learning algorithm's performance. The computational
requirements for estimating the resultant generalization error on a
test set by means of cross-validation are often prohibitive; for
leave-one-out cross-validation one needs to train the underlying
algorithm on the order of $m^2$ times, where $m$ is the size of the
training set. This paper presents several ways to exploit the
bias-variance decomposition to estimate the generalization error of a
bagged learning algorithm without invoking yet more training of the
underlying learning algorithm. In a set of experiments, the accuracy
of this estimator was compared to both the accuracy of using
cross-validation to estimate the generalization error of the
underlying learning algorithm, and the accuracy of using
cross-validation to estimate the generalization error of the bagged
algorithm. The estimator presented here was comparable in its accuracy
to, and sometimes even more accurate than, the alternative
cross-validation-based estimators. 


This paper is available from the web site:
"http://www.santafe.edu/~wgm/papers.html" 
or by ftp from 
""ftp://ftp.santafe.edu/pub/wgm/error.ps.gz"