New improvements on the real-time recurrent learning algorithm

This paper presents some improvements on the real-time recurrent learning (RTRL) algorithm based on second derivatives and on Catfolis (1993) version. The algorithm use estimates to the Hessian matrix that is computed recursively on line with elements based on the sensitivity parameter as defined by Williams and Zipper (1989). Experiments were done to compare the proposed learning algorithm with existing ones in the presence of noise. The new algorithm had shorter learning periods and kept the basic properties of the original RTRL. The proposed algorithm can still be an attractive alternative because its high computing demands can be compensated by the use of very small fully connected neural networks.