First reinforcement learning application in high energy physics
The N3PDF project (http://n3pdf.mi.infn.it/) aims at applying modern Machine Learning and Deep Learning techniques to theoretical physics. In this context, a first novel application of Reinforcement Learning to jet grooming was presented on arxiv.
Jet grooming is a key analysis tools used at hadron colliders such as the LHC. In this paper, we introduce the first application in high energy physics of reinforcement learning - a powerful machine learning paradigm. We train a deep Q Network agent to automatically define a jet grooming strategy according to a carefully constructed reward function, and show that the algorithm found by the machine outperforms heuristic state-of-the-art methods used by experimental collaborations. This jet grooming procedure also shows a strong robustness to poorly modeled non-perturbative contributions, and can be applied successfully to data outside of its training range.