A SECOND LEVEL NEURAL NETWORK TRIGGER IN THE H1 EXPERIMENT AT HERA

Abstract

At the HERA e-p collider the expected machine background rates are typically 105 times higher than the rates from physics. The greatest challenge in the trigger is finding methods which suppress the machine background without using lengthy pattern recognition algorithms with prohibitive computing times. This task is optimally suited to a neural network solution. Our present investigations show that feed-forward networks, trained on the topological energy sums from the H1 calorimeter and on first level tracking trigger information, provide an additional background suppression factor compared to the traditional method which operates with fixed thresholds. The neural network algorithm — implemented by special purpose fast matrix-vector multiplier chips at the second level of the trigger — allows the lowering of the first level thresholds which is shown to be important to get good efficiencies for specific physics event classes.