A practical variable forgetting factor recursive least-squares algorithm

In the context of adaptive filtering, the recursive least-squares (RLS) is a very popular algorithm, especially for its fast convergence rate. The most important parameter of this algorithm is the forgetting factor. It is well-known that a constant value of this parameter leads to a compromise between misadjustment and tracking. In this paper, we present a variable forgetting factor approach, aiming to better compromise between the performance criteria of the RLS algorithm. Also, we propose a practical solution to estimate the power of the system noise (in a system identification scenario), which is required within the algorithm. Experiments performed in the context of network echo cancellation support the advantages of the proposed approach.