How to obtain the high robustness in radius basis function(RBF) learning is a trouble. Therefore, with the
supposition that the perturbation of training dataset is bounded, a RBF network learning algorithm is proposed in terms of the robust least-square principle. Moreover, a strategy of estimating perturbation bound is proposed. Experimental analysis shows that the proposed method has effective robust learning performance. Similar to standard least square algorithm, no additional parameters are needed for the proposed algorithm, which is benefit to more practical applications.