Regression Tree Analysis of Battle Simulation Data
Wei-Yin Loh, (University of Wisconsin, Madison), firstname.lastname@example.org
The sparse and high-dimensional nature of data from battle simulations creates two difficult problems. First, many regression methods cannot deal with these data without some sort of variable selection. If the wrong variables are deleted, much information can be lost. Second, the models are usually difficult to interpret. Even when interpretation is possible, the conclusions can be wrong.
We introduce some new ways of thinking about regression trees that can overcome the above problems. The new methods produce models that are instantly recognizable, are not biased by variable selection, and have high prediction accuracy relative to the best tree and non-tree methods.