A TRAINING METHOD FOR DISCRETE
George D. Magoulas, M.N. Vrahatis, T. N. Grapsa, George Androulakis · 1997
In this contribution a new training method is proposed for neural networks that are based on neurons whose output can be in a particular state. This method minimises the well known least square criterion by using information concerning only the signs of the error function and inaccurate gradient values. The algorithm is based on a modified one-dimensional bisection method and it treats supervised training in networks of neurons with discrete output states as a problem of minimisation based on imprecise values. Subject classification: AMS(MOS) 65KlO, 49DlO, 68T05, 68G05.