Galkin

pp. 1-8

Algorithm of computer-aided generation of teaching selection for neural network with error backpropagation was considered. Teaching and hand-written character recognition occurred through the generation of finite set of different images of the character. The deviation of trace parameters of the character didn’t exceed 15% (values of angles of significant segments with coordinate axis) from the reference hand-written character. It was shown that neural network learned faster from auto-generated selection than from selection generated by a man. Algorithm was implemented in Microsoft C# language.

The authors studied dependence of speed of on-line handwriting parser created on the basis of artificial neuron networks, on the architecture of these networks and program solution as a whole for manual and automated methods of symbols (letters) selection from a connected sequence (words). For each method the authors developed s simulation model which is realized in GPSS language. The simulation results showed that personal pocket computers (PDA) make it possible to carry out handwriting parsing software capable of granting the user useful functionality at an acceptable level.