In this paper we will describe a real-time handwritten characters recognizer based on Multilayer Perceprons with back error propagation learning algorithm. The system has been implemented on a Personal Computer. The necessity of evaluating the performance of neural networks for handwritten characters recognition in real environment brought us to the implementation of a complete system available for end users. This approach foorces to consider the variables and uncertainties not foreseen in simulation phase: the final performance will take into account the recognition efficiency, the adaptation capability to different applications and the facilities provided for user friendly operations. The nentwork design has been made with the aid with a recently published theorem that estabilishes a relation between the number of regions separable by the network and the number of hidden units. The recognition accuracy, obtained by testing the system with a data-base of digits handwritten by six people,was 80%. Experimental resulta are reported in terms of recognition accuraracy and speed of convergence of the learning phase in the diffe3rent examinated conditions. Finally, the effiecient implementation of the system (computational time very close to real-time on a PC AT compatible) opens to practical applications in the area of improved man-machine communication; this aspect will be briefly explored in the paper.
