Wing Sheung Chan

48 Object reconstruction and identification The outputs of the low-level algorithms are combined into a final discriminant by a high-level tagging algorithm. Two high-level tagging algorithms have been developed: the MV2c10 algorithm that uses boosted decision trees (BDT) [76] and the DL1 algorithm that uses a deep feed-forward neural network. For a high target b -tagging efficiency, the MV2c10 and DL1 algorithms provide similar rejection against light-quark- or c -quark- initiated jets. Figure 3.1 shows the expected output distributions and the background rejection against the signal efficiency of the MV2c10 algorithm. 1 − 0.8 − 0.6 − 0.4 − 0.2 − 0 0.2 0.4 0.6 0.8 1 MV2c10 BDT output distribution 3 − 10 2 − 10 1 − 10 1 Event fraction |<2.5 η >20 GeV, | T jet p ATLAS Simulation t = 13 TeV, t s b-jets c-jets Light-flavour jets (a) 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 b-jet efficiency 10 2 10 3 10 4 10 Background rejection |<2.5 η >20 GeV, | T jet p Light-flavour jet rejection c-jet rejection ATLAS Simulation (b) Figure 3.1.: The (a) expected output distributions and (b) background rejection against the signal efficiency of the MV2c10 algorithm [75] . 3.2. Hadronic τ decays Decays of a τ lepton can be either leptonic ( B ≈ 35 . 2% ) or hadronic ( B ≈ 64 . 8% ). Either type of decay produces neutrinos that can be partially reconstructed as part of the missing transverse momentum of the event. The remaining products are the visible decay products. The visible decay product of a leptonic τ decay is an electron or muon, which can be reconstructed like any other isolated electrons or muons using the algorithms later described in Sections 3.3 and 3.4. The visible decay products of a hadronic τ decay ( τ had - vis ), on the other hand, requires more specialised algorithms, which are detailed in this section. The performance of these algorithms is a crucial factor to the sensitivity of the search for Z → `τ decays. The reconstruction and identification of τ had - vis are intricate tasks. Due to the complex- ity, these tasks could often benefit from the use of MVA and machine learning techniques