Solving a Higgs optimization problem with quantum annealing for machine learning

Alex Mott, Joshua Job, Jean-Roch Vlimant, Daniel A. Lidar, Maria Spiropulu

Abstract: A machine learning algorithm implemented on a quantum annealer—a D-Wave machine with 1,098 superconducting qubits—is used to identify Higgs-boson decays from background standard-model processes. Quantum computers show promise for various applications, yet there are few examples of them performing practically useful tasks. Here, Alex Mott and colleagues implement a 'signal versus background' machine learning optimization problem that can be used in the search for Higgs bosons at the Large Hadron Collider at CERN and at future colliders. The researchers use a programmable quantum annealer—a D-Wave machine with 1,098 superconducting qubits—to perform the quantum optimization. Compared to classical machine learning methods, the quantum machine learning approach was robust against overtraining. The discovery of Higgs-boson decays in a background of standard-model processes was assisted by machine learning methods1,2. The classifiers used to separate signals such as these from background are trained using highly unerring but not completely perfect simulations of the physical processes involved, often resulting in incorrect labelling of background processes or signals (label noise) and systematic errors. Here we use quantum3,4,5,6 and classical7,8 annealing (probabilistic techniques for approximating the global maximum or minimum of a given function) to solve a Higgs-signal-versus-background machine learning optimization problem, mapped to a problem of finding the ground state of a corresponding Ising spin model. We build a set of weak classifiers based on the kinematic observables of the Higgs decay photons, which we then use to construct a strong classifier. This strong classifier is highly resilient against overtraining and against errors in the correlations of the physical observables in the training data. We show that the resulting quantum and classical annealing-based classifier systems perform comparably to the state-of-the-art machine learning methods that are currently used in particle physics9,10. However, in contrast to these methods, the annealing-based classifiers are simple functions of directly interpretable experimental parameters with clear physical meaning. The annealer-trained classifiers use the excited states in the vicinity of the ground state and demonstrate some advantage over traditional machine learning methods for small training datasets. Given the relative simplicity of the algorithm and its robustness to error, this technique may find application in other areas of experimental particle physics, such as real-time decision making in event-selection problems and classification in neutrino physics.