As promised, I have done some research into Dempster Shafer Theory. As discussed in class, it is a generalization of Bayes rule that can take ignorance into account. The question raised in class was -- why don't we always use it? Here are some thoughts on that:
- First, it is important to realize that DST is really a starting place within the study of uncertainty. It may not be the final say on how to do this, but to me what is great about it is the fact that it introduces a number of concepts that go beyond probability in modeling uncertainty. To understand some of its limitations, see the references below:
- Here is a nice report that discusses some of the advances in using it and lays out information relating to the practical side of it: http://www.sandia.gov/epistemic/Reports/SAND2002-0835.pdf
- Here is a theoretical take on an inconsistency in the theory itself: http://www.cis.temple.edu/~wangp/Publication/D-S.pdf
- Second, it comes with a cost in efficiency, and may be overkill (e.g. in situations where ignorance is not at issue or we have little data about it).
However, the question of why it has not been adopted more widely may also be one of human issues (we do what we know, we propagate what we ourselves learned about)....
To investigate more deeply, you may want to read A mathematical theory of hints by Kohlas and Monney or A survey on Probabilistic Argumentation (available freely online)