A new robust BP algorithm and its application on the identification of dynamic system

In this paper, we propose a new robust BP algorithm that is resistant to the noise effects and is capable of rejecting entrapment into local minima during the identification of dynamic system. The spirit of this algorithm comes from the pioneering work in an old robust BP algorithm by David Chen, but our work improves the old one in two aspects: 1) enhance the stability by employing the relative residual in defining the objective function. 2) prevent entrapping into local minima through letting the learning rate be a specific function of the LS energy function and its gradient. In contrast to the standard BP algorithm, four advantages of the new algorithm are: 1) it need not interpolate all the training data; 2) it is robust against gross errors; 3) its rate of convergence is improved since the influence of incorrect data is suppressed; 4) it prevents entrapping into local minima effectively.