Inductive learning of decision rules from attribute-based examples: a knowledge-intensive genetic algorithm approach
Cezary Z. Janikow · 1992
Genetic algorithms are stochastic adaptive systems whose search method models natural genetic inheritance and the Darwinian struggle for survival. Their importance results from the robustness and domain independence of such a search. Robustness is a desirable quality of any search method. In particular, this property has led to many successful genetic algorithm applications involving parameter optimization of unknown, possibly non-smooth and discontinuous functions. Domain independence of the search is also a praised characteristic since it allows for easy applications in different domains. However, it is a potential source of limitations of the method as well. In this dissertation, we present a modified genetic algorithm designed for the problem of supervised inductive learning in feature-based spaces which utilizes domain dependent task-specific knowledge. Supervised learning is one of the most popular problems studied in machine learning and, consequently, has attracted considerable attention of the genetic algorithm community. Thus far, these efforts have lacked the level of success achieved in parameter optimization. The approach developed here uses the same high level descriptive language that is used in rule-based supervised learning methods. This allows for an easy utilization of inference rules of the well known inductive learning methodology, which replace the traditional domain independent operators. Moreover, a closer relationship between the underlying task and the processing mechanisms provides a setting for an application of more powerful task-specific heuristics. Initial results indicate that genetic algorithms can be effectively used to process high level concepts and incorporate task-specific knowledge. In this particular case of supervised learning, this new method proves to be competitive to other symbolic systems. Moreover, it is potentially more robust as it provides a powerful framework that uses cooperation among competing solutions and does not assume any prior relationships among attributes.