Wing Sheung Chan

52 Object reconstruction and identification Table 3.1.: Input variables to the BDT of the MVA TES algorithm. The cluster variables are the energy-weighted averages of all the clusters in the core region. More detailed descriptions of the cluster moments can be found in Reference [73] . Variable Description Event variables h µ i Average number of interactions per bunch crossing N PV Number of pile-up vertices τ had - vis variables p comb T Transverse momentum at the combined TES p LC T /p comb T Transverse momentum at LC scale divided by p comb T p pantau T /p comb T Transverse momentum at the pantau TES divided by p comb T η pantau Pseudorapidity from substructure reconstruction N tracks Number of associated tracks Υ Relative energy difference between h ± ’s and π 0 ’s: ( E ( h ± ) − E ( π 0 )) / ( E ( h ± ) + E ( π 0 )) BDT(1p0n vs 1p1n) BDT classification score for differentiating 1p0n and 1p1n BDT(1p1n vs 1pXn) BDT classification score for differentiating 1p1n and 1pXn BDT(3p0n vs 3pXn) BDT classification score for differentiating 3p0n and 1pXn Cluster variables (energy-weighted average) λ centre Distance of the shower centre from the calorimeter front face measured along the shower axis h λ 2 i Second moment in λ , the distance of cells from the shower centre along the shower axis h ρ i First moment in ρ , the energy density f PS Energy fraction in the presamplers P EM Classification probability that the shower is electromagnetic The residues and the correlations are believed to be mainly caused by effects of pile-ups and underlying events. Instead of multiple BDTs for τ had - vis ’s in different | η | regions or decay modes, only one BDT for all τ had - vis ’s is trained. Such an inclusive approach allows the BDT to make use of continuous variables such as | η | and the BDT outputs of the decay mode classifier to dynamically classify the τ had - vis ’s while calibrating the energy. A full list of input variables and their descriptions are shown in Table 3.1. The BDT is trained to minimise the mean squared error between its output and the training target, which is the ratio p truth T /p comb T . In other words, the output of the BDT