A NEAT Solution: Feature Selection through Neuroevolution in Deep Learning-Based \($H \rightarrow \tau \tau$\) Classification

Abstract

The presence of background noise from other Standard Model processes makes identifying (H \rightarrow \tau \tau) particularly difficult. Contemporary methods involve building deep neural networks to classify the Higgs signal from the background and initial success has been seen. Difficulty arises however in determining, both from a physics and statistical perspective, the optimal selection of features for these networks. We present an implementation of the Neuroevolution of Augmenting Topologies (NEAT) algorithm in feature selecting for a bag of multilayer dropout neural networks. Our method called ``Deep NEAT'' achieves an Approximate Median Significance (AMS) of (3.595 \pm 0.027), outperforming benchmarks set by TMVA (AMS: 3.120) and Multiboost (AMS: 3.405). Deep NEAT demonstrates a method for automatic feature selection and confirms the benefit of machine learning techniques in high energy particle physics.