Experimenting genetic algorithms for training a neural network prototype for photon event identification

A computational system based on a synchronous feedback neural network for the on-line event processing of a photon counting intensified CCD has been implemented. Event identification plays a key role as it affects the whole detector efficiency. Identification quality depends on the goodness of event model. The main difficulty in real photon counting applications is to define a precise event model due to the high number of noise sources that make event shape far from the expected ideal model. This results in an intrinsic difficulty in development of efficient neural network training based on conventional gradient search techniques. In this paper we approach the learning problem with real data by using genetic algorithms. Genetic algorithms seem to provide a rapid convergence to good solutions even using limited computational resources. A GENITOR-like algorithm has been developed and implemented in C++, and some results are shown