Inlägg "Hey all, I am having some issues with NaNs in my ..." | Artificiell intelligens (DD2380)

Hey all,

I am having some issues with NaNs in my HMM implementation and wonder if anybody has come across this or even better knows a solution. I pretty much follow Stamp's tutorial to the letter but sometimes I get a situation where during the last iteration of the alpha pass, either all alphas or values of B for this step are zero, resulting in a zero sum of all alphas, which in turn results in an infinite scaling factor.

Now in the alpha pass I don't really notice this since 0 * infinity = 0.

This problem only shows up further down the line where I use the scaling factor in the beta pass, multiplying with it and getting an all-NaN beta array.

I found this post from 2004 with someone having the very same problem while using the stamp pseudocode. This suggests to me that ...

The pseudocode still has a problem/bug regarding numerical stability OR
This is to be expected OR
I made the same mistake as other people ( I know of at least two).

Using this implementation I am still able to get a 200points score though since it doesn't happen every time.

Does anybody have pointers to where the problem might lie?

I had a similar problem and also used the stamp tutorial. The problem seemed to be and it sounds as if you are having the same problem that you try to determine the probability of an impossible sequence according to the hmm you have. There is two quite easy ways around this. You could either adjust your forward algorithm to handle this case or you could make sure there is never zeros in your matrices in which case you will only get very low probabilities instead.