DYNAMIC BIAS IS NECESSARY IN REAL WORLD
Date
1991-04-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This paper discusses the bias present in machine learning
systems, emphasizing its effect on learnability and complexity. A good bias
must allow more concepts to be learned and/or decrease the complexity
associated with learning. The paper develops an exhaustive framework
for bias, with two important distinctions: \fIstatic\fR versus \fIdynamic\fR
and \fIfocus\fR versus \fImagnify\fR. The well-known candidate elimination
algorithm (Mitchell) is used to illustrate the framework. Real world
learners need dynamic bias. The paper examines two representative systems.
\s+2S\s-2TABB (Utgoff) dynamically magnifies the description space where
learning would otherwise be impossible. \s+2E\s-2TAR is a prototype for
learning robot assembly tasks from examples--a dynamic focusing mechanism
reduces both the real world description space and the task construction
complexity. Inductive learning must be viewed as a problem of dynamic
search control.
Description
Keywords
Computer Science