Goals Versus Algorithms
Abstract
Computational methods in statistics often are defined through an algorithm. It is argued that a precise finite sample specification of the goals to be achieved by the algorithm is at least equally important. The issues are discussed in the special case of Projection Pursuit Regression. An interesting initial result of this work, which is still in progress, is that the Friedman- Stuetzle algorithm appears to be systematically biased toward overfitting.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 30, 1992
- Accession Number
- ADA256106
Entities
People
- Peter J. Huber
Organizations
- Massachusetts Institute of Technology