Particle-level event classification for LHC trigger, with e-to-end sorting RNNs
| Main Authors: | Weitekamp, Daniel, Pierini, Maurizio, Vlimant, Jean-Roch |
|---|---|
| Format: | Report Journal |
| Bahasa: | eng |
| Terbitan: |
, 2017
|
| Subjects: | |
| Online Access: |
https://zenodo.org/record/1034814 |
Daftar Isi:
- The LHC experiment produces an overwhelming amount of data, the vast majority of which consists of Quantum Chromodynamic multi-jet events (QCD). However, physicists are generally concerned with rarer physics processes buried in the QCD background. Culling out this background while maintaining signals of interest is one of the greatest challenges at the LHC. Algorithms for distilling pure signals of interest are called triggers, and generally employ simple rule based filters based on physics based features. However, these features are abstractions of hundreds of stable byproducts of a proton-proton collisions, and therefore constitute a significant reduction in information, much of which could have potentially been used to infer the presence of a signal process. We demonstrate that utilizing long-term memory capable recurrent neural networks (RNNs) we can utilize the full range of information available in a given collision event for the purpose of event classification. Furthermore, we demonstrate the effectiveness of our classifier in both trigger and offline settings. Finally we discuss the potential for further improvements, chiefly related to the ordering of the RNN input, and further extensions of the network backpropagation for this purpose.