Adaptation of Boolean networks using back-error propagation

Summary form only given, as follows. The back-error-propagation method has been adapted to give a training procedure for networks of Boolean units. The method has been applied to pyramidal networks of units, and simulation studies of the method have been undertaken. Solutions have been found for hard problems, including the 8-bit parity problem. The results compare favorably with those obtained for networks of semilinear units, but also show the difficulties of applying optimization methods where large plateaus are present in the error function. The method is compared with adaptive procedures for training networks of probabilistic logic neurons.<>