Abstract

The classical genetic algorithm provides a powerful yet domain-independent tool for concept learning. However, in general, learning systems based on the genetic algorithm generally do not perform as well as symbolic learning algorithms. Robert Holte's symbolic learning algorithm 1R demonstrated that simple rules can perform well in non-trivial learning problems, and inspired an approach to machine learning which Holte termed "simplicity first research methodology". A system called ELGAR is proposed, constructed and evaluated in order to investigate the properties of concept learning using the genetic algorithm. A hybrid algorithm is then developed and implemented which integrates the genetic algorithm in ELGAR with the "simplicity first" approach, resulting in a concept learning system that outperforms both 1R and the purely genetic version of ELGAR.