A neural network architecture for the second level trigger in the H1-experiment at the electron proton collider HERA

The H1-experiment is a general purpose particle detector at the electron proton collider HERA. The typical collision rate is about 10 Hz, the background rate 100 kHz. The challenge for the experiment is to suppress the high background rate without losing too many physics reactions. This is performed by a multi-level trigger system. We present the concept, the design and the status for the second level hardware trigger based on a neural network architecture. The design is centered around a VME implementation of the CNAPS parallel computer. We show, with simulated and real data, that we can fulfil the demands with a multi-layer perceptron trained with the backpropagation algorithm. A systematic investigation of the learning parameters is performed on RISC workstations. A new measure for the importance of input variables is presented. A two-dimensional visualization of the network decisions is discussed. We address the problem of building a network for “new” physics reactions which are not present in the training data