On the Forgetting Factors of the RLS Algorithm Based on Third-Order Tensor Decomposition

The recently developed recursive least-squares algorithm based on third-order tensor decomposition, namely RLSTOT, proves to be efficient for the identification of long length impulse responses. Such systems can be encountered in many real-world applications, like network/acoustic echo cancellation. The RLS-TOT combines the estimates of three shorter adaptive filters, which have different convergence parameters. In this paper, we investigate some strategies for setting the forgetting factors of the RLS-TOT. These are the main parameters that control the memory of the algorithm and, consequently, its main performance criteria. Simulations performed in the framework of echo cancellation support the theoretical findings.