Re: Sliding window Viterbi

john <johns@xxxxxxxxxx> wrote:

>I am still maintaining a running cumulative sum of the suriviving path
>error over the entire message. This grows without bound, so I need to
>make this a sliding sum. But the bookkeeping to do that is complicated.
>What I did instead is continuously compute the mean of all of the
>cumulative error sums, and subtrac that off at each step. Then it
>doesn't grow without bound. This gives good BER too.

>Is this a crazy thing to do? I have a bad feeling that it is a dumb
>idea and I should go with the sliding sum instead, however tricky to
>implement.

This is called subtractive normalization and it's not crazy.
(Most people would find the smallest path metric at a given
trellis stage, and subtract it from all of the metrics.
You're subtracting the mean instead, which I suspect works
equally well.)

The other approach is called modulo normaliztion.

Steve
.