This master thesis is a report of a survey aimed at understanding and reducing background sources in central exclusive production events measured at the ALICE experiment, located at CERNLHC. The ALICE experiment consists of a central barrel and a forward muon spectrometer. Additional smaller detectors for global event characterization and triggering are located at small angles outside of the central barrel. Such a geometry allows the investigation of many properties of diffractive reactions at hadron colliders, for example the measurement of single and double diffractive dissociation cross-sections and the study of central exclusive production (CEP). Central diffractive events are defined experimentally by hits in the central barrel and no activity outside of it, creating an activity gap in the observed rapidity of measured particles. The study of Pythia-8 simulations of these processes show a drastic reduction of non-diffractive events (background) by enforcing the rapidity gap condition. The remaining background is largely composed of partially reconstructed CEP events, so called feed-down events. Often feed-down events are accompanied by neutral particles, which are not detected. This missing mass and momentum leads to a shift of the invariant mass spectrum to lower masses. This thesis aims at understanding and suppressing background sources in the two pion invariant mass spectrum in X to pi+ pidecays of the centrally produced system X. This is done in two ways: First, a feed-down template is constructed by using background events marked by a detected gamma in the main calorimeter of ALICE, and by using events with more than two detected charged tracks. Despite facing possibly tedious efficiency corrections for the sake of complete feed-down descriptions, this method yields promising results. Second, machine learning methods for background suppression of CEP events are employed. The measured variables e.g. the four-momentum of particles, energy loss in the detectors, deduced kinematic quantities, and global event characteristics are generally correlated. To obtain a maximal separation of signal and background it is necessary to treat these bservables in a fully multivariate way. Although achieving good results, i. e. the signal purity can be increased by 30% while maintaining a nearly constant signal efficiency, the trained classifiers tend to obtain a strong mass bias which results in a cut-like behavior of the trained model. It can be concluded that multivariate techniques trained on Pythia-8 generated CEP simulations generally suffer from incomplete Monte Carlo descriptions, including only high mass continuum production. However, promising new packages are currently being developed which provide interesting prospects for further studies.