Study of potential improvements of the CMS H->tau tau¯ analysis using artifical neural networks with multiple layers

Abstract

After the discovery of a boson with a mass of approximately 125 GeV in 2012, the next step is to test whether this boson is the Standard Model Higgs boson, e.g. by measuring its couplings to fermions and gauge bosons. One of the fermion couplings is measured by looking at the Higgs - -- decay channel. One way to reach the traditional discovery significance of 5- for this decay is to increase the amount of data. Another way is to find a good signal-versus-background classifier to extract the desired events from the recorded data more efficiently. To obtain this classifier a standard way is to use machine learning models, for instance neural networks. Recent advances in the field of deep learning have shown that deep neural networks are able to retrieve more information out of a given set of input functions than other machine learning methods. The main goal of this thesis is to use deep-learning techniques on a simulated Higgs --- dataset and to compare the performance with other current Machine Learning (ML) techniques. The training is performed with a GPU-accelerated python library. For the tuning of the hyperparameters, a Bayesian optimization algorithm is used. The obtained result is that deep neural networks trained on this simulated dataset can not compete with the other ML techniques used as a benchmark. A possible explanation is that the training set is by far too small to train the deep neural network at a competitive level.