On-line nonparametric regression to learn state-dependent disturbances

A combination of recursive least squares and weighted least squares is made which can adapt its structure such that a relation between in- and output can be approximated, even when the structure of this relation is unknown beforehand. This method can adapt its structure on-line while it preserves information offered by previous samples, making it applicable in a control setting. This method has been tested with computer-generated data, and it is used in a simulation to learn the non-linear state-dependent effects, both with good success.