mdraguet (Oxford) will be responsible for the maintenance and the development of the framework to train the high level taggers. The primary goal is to prepare the training of the DL1r (or equivalent)
tagger for the pFlow jets in the upcoming release 22. The task will initially familiarise with the current training framework and review/optimise the list of inputs information as the low level algorithms are evolving in the new release. The training procedure will particularly focus on first matching, then improving the performance of the current tagger with additional attention to pileup-robustness,
hight pt behaviour and charm-tagging performance. Improvements to be investigated are related to the addition of more input information, output node categorisation (i. e. inclusive VS exclusive b-decays), training strategy and general network architecture.
All the procedure and corresponding software will be documented and made available in GitLab.
Start date: 15 November 2020
Local supervisor: dbortole
Technical supervisor: vdao