The Assessment and Application of Lineage Information in Genetic Programs for Producing Better Models

Abstract

One of the challenges in data mining, and in particular genetic programs, is to provide sufficient coverage of the search space in order to produce an acceptable model. Traditionally, genetic programs generate equations (chromosomes) and consider all chromosomes within a population for breeding purposes. Considering the enormity of the search space for complex problems, it is imperative to examine genetic programs breeding efforts in order to produce better solutions with less training. This research examines chromosome lineage within genetic programs in order to identify breeding patterns. Fitness values for chromosomes are sorted, then partitioned into five classes. Initial experiments reveal a distinct difference between upper, middle, and lower classes. Based upon initial results, a novel genetic programming process is proposed which breeds a new generation exclusively from the top 20 percent of a population. A second set of experiments statistically validate this proposed approach