hmmFit and hmmLogprob for classification

[ido]

From [email protected] on January 16, 2011 04:00:38

What steps will reproduce the problem? 1. train separate models using positive discrete samples from different classes 2. cross-test for the highest likelihood using hmmLogprob What is the expected output? What do you see instead? hmmLogprob sometimes crashes when a model is tested on data that do not belong to the same class

What version / revision of the product are you using? On what operating system? pmtk3 nov 2010 Please provide any additional information below. The problem is that the discretized positive samples in the training set might not include all the possible observations for a class, but hmmFit estimates the number of observations from the positive samples. So, if there are 8 observation symbols, but a class makes use of only the first 6, the trained model also has 6 observation states, and cannot evaluate samples belonging to other classes, a crucial step for classification.

Solution: There should be an option to set the number of observations directly.

Original issue: http://code.google.com/p/pmtk3/issues/detail?id=12