Identification of Hadronic Tau Lepton Decays at the ATLAS Detector Using Artificial Neural Networks

Tau leptons play an important role in a wide range of physics analyses at the LHC, such as the verification of the Standard Model at the TeV scale or the determination of Higgs boson properties. For the identification of hadronically decaying tau leptons with the ATLAS detector, a sophisticated, multi-variate algorithm is required. This is due to the high production cross section for QCD jets, the dominant background. Artificial neural networks (ANNs) have gained much attention in recent years by winning several pattern recognition contests. In this thesis, a survey of ANNs is given with a focus on developments of the past 20 years. Based on this work, a novel, ANN-based tau identification is presented which is competitive to the current BDT-based approach. The influence of various hyperparameters on the identification is studied and optimized. Both stability and performance are enhanced through formation of ANN ensembles. Additionally, a score-flattening algorithm is presented that is beneficial to physics analyses with no defined working point in terms of signal identification efficiency.