Recursive least squares approach to learning in recurrent neural networks

This paper presents a new approach to learning in recurrent neural networks, based on the descent of the error functional in the space of the linear outputs of the neurons (neuron space approach). At each step of the learning process a linear system is solved for the weights using a recursive least squares technique. This approach, with respect to traditional gradient-based algorithms, guarantees better performances from the point of view of both the speed of convergence and the numerical robustness