We want to estimate parameters based on a number
of observations.
With Recursive Least Squares (RLS) we do it
in an online fashion,
updating the prediction whenever we get a new observation.
The old observations can be discarded.
The basic concept is as follows:
Initial values:
The initial value for beta is often set
to 0, with delta we can adjust our trust in that value.
A high delta (typically around 100) implies low trust.
The first example compares batch least squares regression (bls)
with RLS.
In the second example we estimate a sin curve with RLS using 300 samples.
