search

psycharo-zz / factor-graph

matlab/c++ factor graph framework
32 stars 8 forks source link

AR parameter estimation #31

Open psycharo-zz opened 12 years ago

psycharo-zz commented 12 years ago

Make a working example of parameter estimation for the autoregressive model. For now do not estimate variances.

psycharo-zz commented 12 years ago

@bertdv , I got some results for this one, however, they are not yet stable (sometimes I am getting NaN's as answers, probably it is something about matrix inversion). But even when the computation is successful, the parameters that are estimated are really different from the ones I specified in the beginning (while the values vector is OK). But from my understanding this is fine, since there can be actually arbitrarily many combinations that lead to the same result. Is this correct?

ghost commented 12 years ago

Good to hear you're getting results. Yes, it is possible that matrix inversions are ill-conditioned and as a result lead to divergences. You can repair this by adding Table-6 in Loe-2007. While this would be a good thing to do anyway, my suspicion is that there is still a bug somewhere though. This is not like PCA where any rotation of the subspace leads to the same errors. I really would expect the system the converge to the correct coefficients, aside from misadjustment as a result of the noise injections. I suggest that you start with a simpler AR problem: learn 1 coefficient with no obervation noise, then add more coefficients (this will be RLS) and once that works add observation noise (now its a Kalman filter).