The tuning of error signal for back-propagation algorithms

Abstract

Despite of Back-propagation (BP) algorithm existence for almost four decades, it is still widely used in many fields to solve range of real world problems. However, it suffers from slow convergence and tends to trap in local minima. So, an improved two-term error function is proposed to overcome the existing problems. This new algorithm is proven to be a better algorithm. The main purpose of this study is to evaluate the efficiency of improved two-term error function by applying three different values of ß parameter in the activation function. The improved twoterm error with different error signal (d ) will replace the conventional error signal in standard BP. These both algorithms will be tested on three universal datasets; Iris, Balloon and Cancer by comparing the accuracy and the convergence speed. The ultimate outcome of the study would be handful information to get a better justification on these both Back-propagation algorithms usages in real application.