Experimental results with NGE

Abstract

The Each program has been tested in several domains, using two different configurations of the program. The results presented below used four different problems: (1) predicting the recurrence of breast cancer, (2) classifying iris flowers, (3) predicting survival for heart attack victims, and (4) predicting the dependent variable in a discrete event simulation. The simulation problem was based on a very simplified stock market model. In the process of testing Each on these problems, the algorithm has been tuned in various ways, and much has been learned about the kinds of problems on which Each performs well, and those on which it performs poorly. In particular, in the test runs for the simulation, I was able to vary such things as the number of variables and the complexity of the underlying relationships quite freely. For the real data used in the other sets of test runs, the variables were fixed. Thus the simulation allowed me to test the system in a more controlled way, while the real data offered a chance to test the system with data which was in some cases very noisy and incomplete, and, most importantly, to compare Each with other systems.