Sequential growing-and-pruning learning for recurrent neural networks using unseented or extended Kalman filter

This paper presents a sequential growing-and-pruning learning algorithm employing an unscented or extended Kalman filter (SGAPL-UKF or SGAPL-EKF) for a recurrent neural network (RNN). The RNN is constructed using a sequential-learning algorithm that employs growing-and-pruning (GAP) criteria based on the concept of the significance of hidden neurons to yield a compact network; and an unscented or extended Kalman filter improves the learning accuracy by providing estimates of the parameters of the RNN from incomplete samples. As an example, this method was used to estimate the output of a Mackey-Glass time series. A comparison of the results obtained with a UKF and an EKF yielded guidelines about which situations each type of filter is suitable for. Verification results show the effectiveness of the learning algorithm.