Approximation Via Value Unification TITLE2
Paul E. Utgoff, David John Stracuzzi · 1999
Numerical function approximation over a Boolean domain is a classical problem with wide application to data modeling tasks and various forms of learning. A great many function approximation algorithms have been devised over the years. Because the goal is to produce an approximating function that has low expected error, algorithms are typically guided by error reduction. This guiding force, to reduce error, can bias the algorithm in a detrimental manner. We illustrate this bias, and then propose an alternative approach based on a notion of value unification. We compare these two biases in the realm of regression tree induction, and then present a new function approximation algorithm AVVU based on value unification. EMail: [email protected]