PATTERN MATCHING IN HIGH ENERGY PHISICS BY USING NEURAL NETWORK AND GENETIC ALGORITHM MARCELLO CASTELLANO, GIUSEPPE MASTRONARDI, VITOANTONIO BEVILACQUA, E. NAPPI* Dipartimento di Elettrotecnica ed Elettronica Politecnico di Bari, Via Orabona 4, 70126, Bari, ltaly Phone:+39805460252;Fax:+39805460410 E-mail: mastrona @poliba.it INFN sezione di Bari Abstract In this paper two different approaches to provide information from events by high energy phisics experiments are shown. Usually the representations produced in such experiments are spot-composed and the classical algorithms to be needed for data analysis are time consuming. For this reason the possibility to speed up pattern recognition tasks by soft computing approach with parallel algorithms has been investigated. The first scheme shown in the following is a two layer neural network with forward connections, the second one consists of an evolutionary algorithm with elitistic strategy and mutation and cross-over adaptive probability. Test results of these approaches have been carried out analysing a set of images produced by an optical Ring imaging Cherenkov (RICH) detector at CERN. Keywords: neural network, genetic algorithms, pattern recognition, high energy physics. 1 Introduction The main effort in the analysis of data obtained in high energy physics depends on the statistical nature of investigated phenomena. The use of optical imaging detectors provides an advanced solutions of the problem but the presence of noise in detected images makes pattern recognition more complicated. The output produced by optical imaging detectors consists of grey-level images wich record the light stored in each pixel. Several pattern recognition algorithms have been proposed up to now to solve automatically many decision making problem [2], they involved criteria as template matching, detection of similarity and clustering and heuristic, mathematic or syntactic implementative methodologies. Experimental observation of phenomena at subnuclear level requires quanta of high energy. Particle accelerators today achieve energies of the order of the TeV allowing the investigation at distances down cm. The physics production system can be realized either firing a beam of highly energetic particles toward a fixed target of material or by the collision between circulating beams of highly energetic particles. The resulting secondary particles are observed by a detectors system. Each detector is devoted to explore a feature of the incident particle by an interaction with it to performe the measure process. Pattern recognition techniques can be used to define the feature value as the first step toward the particle identification task. ALICE is an HEP experiment at the Large Hadron Collider at CERN optimized for the study of high energy heavy-ion collisions [I]. A Ring Imaging Cherenkov detector RICH is adopted in ALICE with the aim to identify high-momentum particles in the range from 1 to SGeV/c. To perform this task, the RICH is able to produce images in which to search for circular regions to charged particles recognition. 2 THE NEURAL NETWORK MODEL The proposed neural network [6,10] consists of an input and an output layer with feedforward interconnection between them and is based on adaptive linear threshold neurons. The input of the first layer consists of incoming (n x n) binarized images and the output layer designed in order to guarantee the shift invariance of the position of the stimulus pattern consists of a set of m planes where each plane provides (h x h) neurons . In each plane the neurons have the same synaptic spatial distribution and are connected with presynaptic neurons belonging to the