Robust BP theory and algorithms based on several kinds of error estimators

The mean-squared error estimator was used by standard BP (backpropagation) algorithms. Therefore these algorithms might get trapped in some local minimum, have slow convergence and be sensitive to initial weight values etc. In this paper, a kind of new robust BP mathematical theory which is based on the Lagrangian multiplier method and several kinds of robust error estimators is investigated in detail. Robust BP algorithms are obtained. Experiments illustrate: our algorithms not only converge fast and are less sensitive to initial weight values, but also can overcome the influence of “outliers”. The algorithms are robust for little noise perturbation and gross error