Incremental learning algorithm for function approximation and its realization using neural networks

In this paper an incremental learning algorithm for function approximation is presented. The algorithm utilizes the current training pattern to generate an approximately learned function with minimum change in the previously learned one. The algorithm does not store previous learned data for retraining, thus it emulates the biological incremental learning process. The new learned function coefficients are computed using the least squares and the goal attainment optimization methods. The proposed algorithm can be applied to learn systems that operate dynamically in changed environments, such as learning inverse dynamics of robots.