There is a great deal of information to be gained from studying the details within a GA run. This paper investigates the roles of crossover and mutation by observing the actions and eeects of individual occurrences of each genetic operation. The observations are compared with some of the common expectations of these operators.
A critical part of the design of eeective EAs is to obtain a proper balance between exploration and exploitation. A key element of this balance is the selection algorithm used. Although we have for some time now understood the relative diierences between various selection methods for standard centralized EAs, our understanding of decentralized EAs has been less complete. In this paper we present analysis tools and results which provide considerable insight into the properties of various decentralized selection methods, including the selection pressures they induce. Understanding this permits a more informed choice of selection methods when designing and implementing spatially structured EAs.
A learning classifier system, EpiCS, was used to de rive a continuous measure of disease risk in a series of 250 individuals. Using the area under the receiver-ope rating characteristic curve, this measure was compared wit h the risk estimate derived for the same individuals by logistic regression. Over 20 training-testing trial s, risk estimates derived by EpiCS were consistently more accurate (mean area=0.97, SD=0.01) than that derive d by logistic regression (mean area=0.89, SD=0.02). The areas for the trials with minimum and maximum classification performance on testing were signific antly greater (p=0.019 and p<0.001, respectively) than th e area for the logistic regression curve. This inves tigation demonstrated the ability of a learning classifier s ystem to produce output that is clinically meaningful in diagnostic classification.
