System identification using the fast LMS-sine algorithm

The authors extend the desirable features inherent in the gradient LMS (least-mean-square) algorithm and explore an approach aimed at improving its convergence rate. They modify the LMS algorithm by adding a nonlinear term in the update recursion. The resulting algorithm emulates the dynamics of a planar pendulum, and in the steady-state it reduces to an LMS algorithm with much smoother learning curves. Additionally, the new algorithm has a much faster convergence rate than existing gradient algorithms.<>