Author(s): Domingos, Pedro | Abstract: Current rule induction systems (e.g. CN2) typically rely on a "separate and conquer" strategy: they induce one rule at a time, removing the newly covered examples from the training set after each step. This results in a dwindling number of examples being available for learning successive rules, which in turn causes several problems that adversely affect the accuracy of the resulting rules. The research reported here investigates the alternative: learning all rules simultaneously using the entire training set for each. A viable approach using this strategy is proposed and implemented in the RISE 1 system. Empirical comparison of the new system with CN2 suggests that "conquering without separating" performs similarly to its counterpart in simple domains, but achieves increasingly substantial gains in accuracy as the domain difficulty grows, without sacrificing speed.
[1]
J. Ross Quinlan,et al.
C4.5: Programs for Machine Learning
,
1992
.
[2]
Ryszard S. Michalski,et al.
A Theory and Methodology of Inductive Learning
,
1983,
Artificial Intelligence.
[3]
Larry A. Rendell,et al.
Lookahead Feature Construction for Learning Hard Concepts
,
1993,
International Conference on Machine Learning.
[4]
Robert C. Holte,et al.
Concept Learning and the Problem of Small Disjuncts
,
1989,
IJCAI.
[5]
Peter Clark,et al.
Rule Induction with CN2: Some Recent Improvements
,
1991,
EWSL.