Robust sequential learning algorithm for function approximation base on strong tracking filter

This paper addresses the problem that network whose parameters are updated using EKF can not obtain robust performance if the system state saltates when EKF reach stable state. Strong tracking filter which introduces suboptimal fading factor matrix to overcome the problem is utilized to adjust the network parameters to obtain robust performance. The winner neuron updating strategy is also employed to reduce the computation load for online application. Experimental results show the proposed algorithm can achieve smaller approximation error and more compact network structure than several other typical sequential learning algorithms.