Slope and learning rate adaptation scheme for neural networks and its application to blind equalization

Error back propagation (EBP) is the most used training algorithm for feedforwrd artiJicia1 neural networks (FFANNs). Howevei; it is generally believed that it is vely slow if it does converge, especially the network size is not too large compared to the problem at hand. The speed of the learning phase depends both on learning rate (LR) and on the choice of activation functions (AFs). In this papei; a non gradient scheme is proposed to enhance the convergence of EBP algorithm; this scheme is based on the observation that keeping high LR and linear AF during startup enhances the learning capability. But as the network output comes close to the target value, a gradual decrement in LR and increment in the slope of AF ensure a better steady state mapping. The proposed scheme is applied on a blind neural equalizer and it performed better than the standard EBP