- author = "Robert Evan Marmelstein",
- title = "Evolving Compact Decision Rule Sets",
- school = "Faculty of the Graduate School of Engineering of the Air Force Institute of Technology Air University",
- year = "1999",
- address = "USA",
- month = jun,
- keywords = "genetic algorithms, genetic programming, GRaCCE, Matlab",
-
URL = "
ftp://math.chtf.stuba.sk/pub/vlado/thesis_Marmelstein/thesis_Marmelstein.pdf",
-
URL = "
ftp://math.chtf.stuba.sk/pub/vlado/thesis_Marmelstein/thesis_Marmelstein.ps.gz",
- size = "271 pages",
-
abstract = "With the increased proliferation of computing
equipment, there has been a corresponding explosion in
the number and size of databases. Although a great deal
of time and effort is spent building and maintaining
these databases, it is nonetheless rare that this
valuable resource is exploited to its fullest. The
principle reason for this paradox is that many
organizations lack the insight and/or expertise to
effectively translate this information into usable
knowledge. While data mining technology holds the
promise of automatically extracting useful patterns
(such as decision rules) from data, this potential has
yet to be realized. One of the major technical
impediments is that the current generation of data
mining tools produce decision rule sets that are very
accurate, but extremely complex and difficult to
interpret. As a result, there is a clear need for
methods that yield decision rule sets that are both
accurate and compact.
The development of the Genetic Rule and Classifier Construction Environment (GRaCCE) is proposed as an alternative to existing decision rule induction (DRI) algorithms. GRaCCE is a multi-phase algorithm which harnesses the power of evolutionary search to mine classification rules from data. These rules are based on piece-wise linear estimates of the Bayes decision boundary within a winnowed subset of the data. Once a sufficient set of these hyper-planes are generated, a genetic algorithm (GA) based {"}0/1{"} search is performed to locate combinations of them that enclose class homogeneous regions of the data. It is shown that this approach enables GRaCCE to produce rule sets significantly more compact than those of other DRI methods while achieving a comparable level of accuracy. Since the principle of Occam's razor tells us to always prefer the simplest model that its the data, the rules found by GRaCCE are of greater utility than those identified by existing methods.",
- notes = "AFIT/DS/ENG/99-05 Approved for public release; distribution unlimited Appendix B. GRaCCE User's Guide",
