Applying Rule Ensembles to the Search for Super-Symmetry at the Large Hadron Collider

Abstract

In this note we give an example application of a recently presented predictive learning method called Rule Ensembles. The application we present is the search for super-symmetric particles at the Large Hadron Collider. In particular, we consider the problem of separating the background coming from top quark production from the signal of super-symmetric particles. The method is based on an expansion of base learners, each learner being a rule, i.e. a combination of cuts in the variable space describing signal and background. These rules are generated from an ensemble of decision trees. One of the results of the method is a set of rules (cuts) ordered according to their importance, which gives useful tools for diagnosis of the model. We also compare the method to a number of other multivariate methods, in particular Artificial Neural Networks, the likelihood method and the recently presented boosted decision tree method. We find better performance of Rule Ensembles in all cases. For example for a given significance the amount of data needed to claim SUSY discovery could be reduced by 15 % using Rule Ensembles as compared to using a likelihood method.