Wing Sheung Chan

Event selection and classification 79 Table 4.6.: Input to the NN classifiers. The list is common between the classifiers for the eτ and µτ channels. The top six variables are the low-level variables and are measured in the transformed frame of reference as described in text. The bottom four variables are the high-level variables and are measured in the laboratory frame of reference. Variable 1P region 3P region NN Z ττ NN Wjets NN Z `` NN Z ττ NN Wjets ˆ p z ( ` ) • • • • • ˆ E ( ` ) • • • • • ˆ p x ( τ had - vis ) • • • • • ˆ p z ( τ had - vis ) • • • • • ˆ E ( τ had - vis ) • • • • • ˆ E miss T • • • • • ∆ α ( `, τ ) • • • • • m vis ( `, τ ) • • • • • m coll ( `, τ ) • • • • • m ( `, τ track ) • have a tendency to increase after a certain epoch (see Appendix B) . In view of that, no regularisation or dropout is added. Other unspecified configurations or hyperparameters have not been optimised explicitly and the default values from Keras 1.1.0 are used. The configurations and hyperparameters of the NN architecture and optimiser are summarised in Table 4.7. Figure 4.2 shows the expected output distributions of the targeted background and signal of the different NN classifiers. 4.2.4. Combined output For every event, each NN classifier would output a score ranging from zero to one. The score reflects the likelihood of an event to be a signal event. The most background-like events would be given a score close to zero, while the most signal-like events would receive a score close to one. For each channel ( eτ or µτ , and 1P or 3P) separately, the output scores from the different classifiers are combined into one final discriminant, called the “combined NN output” or “NN comb output”, using the formula combined NN output = 1 − s P b w b (1 − NN b output ) 2 P b w b , (4.8)