On the basis of applying neural networks to the track finding problem,
investigations are made according to the specific properties of such
discrete detectors as multiwire proportional chambers. These investiga
tions result in the modification of the so-called rotor model of a neu
ral network. The energy function of a network in this modification con
tains only one ''cost'' term. This speeds up calculations considerably
. The reduction of the energy function is done by the neuron selection
with the help of simple geometrical and energetical criteria. Besides
, cellular automaton was applied to preliminary selection of data that
made it possible to create an initial network configuration with the
energy closer to its global minimum. We study also the following probl
ems of track information extraction by our ANN model: providing an ini
tial ANN configuration by an algorithm general enough to be applicable
for any discrete detector in- or out of a magnetic field; robustness
to heavy contaminated raw data; stability to the growing event multipl
icity. The algorithm was tested on real and simulated events obtained
from the ARES spectrometer with satisfactory results.