Numerical experiments

We proceed to demonstrate the performance of the KRLS-T algorithm, first on a stationary benchmark and then on a tracking problem with non-stationary data. We include several other relevant algorithms in the comparison, in particular Approximate Linear Dependency KRLS (ALD-KRLS) and Sliding-Window KRLS (SW-KRLS). The ALD-KRLS algorithm iteratively constructs the solution to a stationary batch regression problem, and hence it is not suitable for tracking. To slow down dictionary growth it uses an approximate linear dependency criterion (see [4] for details). A notable characteristic of ALD-KRLS is that it does not intrinsically handle forgetting or regularization, but rather achieves its regularization by constructing a sparse basis. The SW-KRLS algorithm [7,13] is, to the best of our knowledge, the only relevant KRLS algorithm capable of tracking, apart from the proposed algorithm. Some additional algorithms will be mentioned briefly throughout the experiments.