Gain Variation in Recurrent Error Propagation Networks