Determining Fitness Function Parameters for GA-Boost

• Activity Number: 236
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Dong-Yop Oh*+ and J. Brian Gray
• Companies: University of Alabama and University of Alabama
• Keywords: AdaBoost ; classification ; genetic algorithm ; predictive model ; weak classifier

Abstract

Our recently proposed genetic boosting algorithm, GA-Boost, directly solves for the weak classifiers in an ensemble and their weights using a genetic algorithm. The fitness function consists of three parameters (a, b, and p) that limit the number of weak classifiers (by b) and control the effects of outliers (by a) to maximize an appropriately chosen p-th percentile of margins. We use several artificial data sets to compare GA-Boost performance at 16 different treatment levels, as well as how it compares to AdaBoost, at four different noise levels. Through these simulations, we verify that GA-Boost has better performance with simpler predictive models than AdaBoost when there is a large proportion of outliers in a data set. GA-Boost is applied to real data sets with three different weak classifier options and compared to other robust boosting methods. We also consider graphical methods for selecting the value of p.