Adaptive neural nets filter using a recursive Levenberg-Marquardt search direction

This paper proposes a recursive Levenberg-Marquardt (LM) search direction as the training algorithm for non-linear adaptive filters, which use multi-layer feed forward neural nets as the filter structures. The neural nets can be considered as a class of non-linear adaptive filters with transversal or recursive filter structures. In the off-line training, the LM method is regarded as an intermediate method between the steepest descent (SD) and Gauss-Newton (GN) methods, and it has better convergence properties than the other two methods. In the echo cancellation experiments, the recursive LM algorithm converges faster and gives higher echo return loss enhancement (ERLE) than the recursive SD and GN algorithms.