Neural network trigger algorithms for heavy quark event selection in a fixed target high energy physics experiment

The study of particles containing heavy quarks is currently a major topic in High Energy Physics. In this paper. neural net trigger algorithms are developed to distinguish heavy quark (signal) events from light quark (background) events in a fixed target experiment. The event tracks which are parametrized by the impact parameter D and the angle @ of the track with respect to the beam line, vary in number and in position in the Q-D plane. An invariant second order moment feature set and an invariant D-sequence representation are derived to characterize the signal and background event track patterns in the Q-D plane. A 3-layer perceptron is trained to classify events as signal/background via the moments and D-sequences. A nearest neighbor classifier is also developed to serve as a benchmark for comparing the performance of the neural net triggers. Results indicate that the selected moment feature set and the D-sequence representation contain essential signal/background discriminatory information. The results also show that the neural network trigger algorithms are superior to the nearest neighbor trigger algorithms, A very high discrimination against background events and a very high efficiency for selecting signal events is obtained with the D-sequence neural net trigger algorithm. SUMMARY The study of particles containing heavy quarks is currently a major topic in High Energy Physics and, in this paper, neural net trigger algorithms are developed to distinguish heavy quark (signal) events from light quark (background) events in a fixed target experiment. The event tracks which are parameterized by the impact parameter D and the angle 4, of a track with respect to the beam line vary in number and in position in the 0-D plane and cannot, therefore, be used as inputs to a neural network directly. This problem is overcome by deriving an invariant second order moment feature set and a Dsequence representation to characterize the signal and background tracks in the Q-D plane. The moments feature set characterizes the dispersion of the tracks and the orientation of the tracks in the (P-D plane. The D-sequence which is obtained through a simple set of transformations captures the track variations along the D-axis. A 3-layer perceptron is trained to classify events as signal/background via the normalized moments feature set and the Dsequences. The key to a successful study of heavy quark physics is a very high discrimination against background events and a high efficiency for selecting signal events. A training strategy is developed to keep the background misclassifications at a minimum. A nearest neighbor classifier is also developed to serve as a benchmark for comparing the performance of the neural net trigger algorithms. The very high efficiency obtained for rejecting background events and for selecting signal events clearly indicate that the selected moment feature set and the D-sequence representation contain essential signal/background discriminatory information. The results obtained also show that the neural net triggers are superior to the nearest neighbor triggers and the D-sequence neural net trigger is superior to the moments neural net trigger. It is important to note that the results obtained are very impressive as tests on randomly selected events indicate that, in many cases, it is impossible to visually distinguish between signal and background events from the track patterns in the OD plane.