Can someone please comment on pull request #181? https://github.com/apache/mahout/pull/181
I have implemented a more numerically stable scaling method for Baum Welch training of HMM. This new scaling method partially addresses Mahout-627. The existing scaling method based on log scaling is not numerically stable when tried with the Mapreduce version of Baum Welch proposed for Mahout-627. With the rescaling method, I am able to successfully run the Mapreduce version of Baum Welch on long training sequences. Thanks Manogna
