This projects explores several classifiers to discriminate between signal particle processes (involving 4 top quarks in our case) and background processes (involving a top and antitop quark). The same analysis can also be applied to distinguish between different particle processes.
The most important findings of this project are presented in a report: Distinguishing_4_top_events_from_background.pdf.
-
This project start with exploratory data analysis in:
Data Exporation and Preprocessing.ipynb
-
Several conventional machine learning classifiers are tested in:
Algorithm Selection.ipynb
Tested are: single Decision Trees, boosted ensembles of Decision Trees, bagged ensembles of Decision Trees and Random Forest classifiers. -
Performance of Deep Neural Networks is optimised in:
Train Deep Neural Networks.ipynb
Also, rotational symmetry and inversion symmetry of particle processes is incorporated here. -
Convolutional Neural Networks are employed to better extract spatial relations in:
Train Convolutional Neural Networks.ipynb
Here rotational symmetry and inversion symmetry of particle processes is translated to their effect in the generated images. Additionally, ensembles of Convolutional Neural Networks are used to boost performance.
The used dataset is a subset of a generated LHC-like dataset as part of [1], and is available at https://www.phenomldata.org/. The subset consists of 83.300 ttbar events and 16.700 4top events. The used subset is available in TrainingValidationData.csv.
[1] G. Brooijmans, A. Buckley, S. Caron, et. al. Les Houches 2019 Physics at TeV Colliders: New Physics Working Group Report. arXiv:2002.12220, Feb 2020.