The High Luminosity upgrade of the Large Hadron Collider (LHC) will increase the instan-
taneous luminosity from 2 ×1034cm−2 s−1 to 5 − 7.5 × 1034 cm−2 s−1. This will increase the
average number of simultaneous proton-proton collisions from the current value of 60 to 140
and eventually 200. This poses a large challenge for the experiments, particularly for the trigger
systems which process data quickly to determine which events to store. The CMS High Level
Trigger (HLT) will receive data at a rate 5-7.5 times higher than at nominal operation. It will
need to accommodate an acceptance rate up to 7.5 kHz, which will require around 20 times the
computing power that is required now.
This thesis presents detailed timing studies of the HLT algorithms, identifying charged par-
ticle tracking as the main cause of the increased computing requirement. The computational
performance of the current Kalman filter based tracking algorithm is evaluated along with a
parallelised algorithm called mkFit. The potential for these algorithms to be accelerated with
co-processors is discussed. Two machine learning algorithms are then explored. Graph neural
nets (GNNs) have shown promising performance for tracking. This thesis applies them to CMS
data and removes commonly used simplifications. Building graphs is shown to be the most
time consuming element of the GNN pipeline, making low latency applications challenging.
Two novel reinforcement learning methods for tracking are introduced, exploring both a con-
tinuous and a discrete environment. They show a hit classification score up to 90%, depending
implementation options. The classification score is not yet sufficient for future LHC require-
ments, but with a small and simple neural net, it shows potential for speeding up tracking.