Among the main goals of the Large Hadron Collider (LHC) physics program is the search for physics beyond the Standard Model, also known as New Physics. This implies the analysis of a vast amount -petabytes- of data from the collisions in the search for faint discrepancies between the collected data, from ATLAS in our case, and the expected physics from the Standard Model. Most of the New Physics searches at the LHC consist on assuming a (e.g. SUSY) model that would become manifest as a specific set of the aforementioned discrepancies; however, in order to have a more complete coverage of New Physics searches in our data, we need to probe more generic scenarios that don’t depend on the specifics of a model, in as many experimental signatures as possible.
In recent years, there have been several efforts to port techniques from the booming field of Machine Learning into LHC analyses. We present preliminary studies of two methods used in the context of New Physics searches, more specifically in the dijet final state scenario. The first of them is an anomaly detection method that uses mixture models in a semi-supervised approach, performing regularization by means of a penalized likelihood. The second method uses Gaussian Processes to model smooth background and generic signals in one-dimensional distributions. Additionally, we present a set of tasks performed within TADA: a fast monitoring system in ATLAS.