18:18 < cult-> hi. I have an issue with hmm and the training data. some of my 'normal' training data are fails and throws error: chol(): decomposition failed

18:19 < cult-> sometimes, the beginning of the data works well, and the later part throws this exception. i am not sure what should I do with my data. also, the longer the data, eg. tens of thousands of observations, the probability of getting this error increases.

18:19 < cult-> how should I overcome of this annoying issue?

18:22 < cult-> i observed that if my data has more extreme residuals, this is more likely to happen too, the cholesky decomposition.

19:11 < rcurtin> cult-: what's the observation type, is it gaussian or GMMs?

19:11 < rcurtin> with Gaussians the cholesky decomposition can fail when the covariance matrix gets very small, it becomes non-invertible

19:12 < rcurtin> so one thing you can do is add noise to your observations, or possibly reducing the number of states can help too

19:46 < cult-> gaussian

19:59 < rcurtin> yeah, so try adding some noise to the observations and see if that helps

20:00 < rcurtin> I thought I had written code to help prevent this situation some time back though... are you using a new(ish) version of mlpack? 2.1.0 and newer should incorporate those fixes

20:06 < cult-> i tried both 2.0.x and 2.1.x

20:06 < cult-> 2.1.1 but not the master branch

20:18 < rcurtin> you can try the master branch, but I doubt it's different---I think HISTORY.md has documentation of when the HMM fixes occured; I think it was either 2.0.1 or 2.1.0

20:21 < cult-> i will first try some cleaning on my data, because it has some bad periods

20:21 < cult-> i think it will help but let's see

20:34 < rcurtin> yeah, often this might happen if, e.g., one state has observations that are all identical

20:34 < rcurtin> then the covariance matrix becomes noninvertible for that state

20:34 < rcurtin> anyway I have to go for now, I'll be back later---good luck :)

20:48 < cult-> thats the case, but i haven't yet cleaned it, thanks again!