Efficiency Improvements in Monte Carlo Algorithms for High-Multiplicity Processes

Monte Carlo event generators are essential tools bridging the gap between theoretical predictions and measurements from contemporary collider experiments. Due to their importance and intensive use, it is crucial that event generators provide precise predictions while preserving reasonable computational costs. This thesis presents efficiency improvements for the Sherpa event generator, in particular for high-multiplicity processes. A short introduction to Monte Carlo techniques as well as concepts of perturbative QCD is given to serve as a basis for the following discussions. Subsequently, the idea behind the Machine Learning based unweighting is presented and the optimisation of the utilised Neural Network architecture described. Resulting performance improvements in comparison to the "classic" unweighting prescription are investigated. Significant time improvements are achieved for all examined processes. Finally, the problem of negative weighted events, as well as mechanisms to reduce the negative weight fraction, are discussed with focus on Z and tt production in association with jets. In both cases, the amount of negative weighted event can roughly be halved for the investigated process setups.