Boosted Decision Trees for the ATLAS Level-1 Calorimeter Tau Trigger

The Large Hadron Collider at CERN generates 10$^{34}$ proton-proton collisions per square centimeter per second in the ATLAS detector. Unfortunately, it is not feasible to save to disk such enormously large volumes of data - necessitating the concept of the trigger. The ATLAS Trigger System decides which collision events contain interesting physics and which do not. The interesting events are saved to disk and the rest are discarded. This thesis specifically discusses the Tau Trigger System which attempts to detect in real-time if a given collision contains a signal compatible with a tau lepton. Until the writing of this thesis, a simple energy cutoff algorithm was used which was moderately successful at classifying true tau events. Since the Trigger System suffers from harsh latency constraints as well as being based on hardware with limited resources, it was challenging to improve upon these results. This thesis proposes a novel, hardware specific, machine learning based approach which improves the performance of the ATLAS Tau Trigger. The use of this new method will result in the recording of more tau lepton collision data which is critical to many ongoing analyses.