Re: BNs from Decision Trees

Lars M. Nielsen (ln@hugin.dk)
Wed, 23 Sep 1998 08:22:58 +0200

Michael Mayo wrote:
>
> Here's a question for you all...
>
> If I have a decision tree with probabilities attached to each and
> every outcome arc and leaf nodes representing conclusions, are there
> any algorithms for inducing a bayesian network?
>
> Has anyone done any work on this? Would it be possible to induce a
> bayesian network with a single conclusion node where each value
> corresponds to a leaf node in the DT?

I don't know any particular algorithms. However, we had some discussions
about this in the decision support group at Aalborg University some time
ago. We concluded that in some cases - if the same decision sub tree is
repeated on the different branches - it would be favourable to convert
it to a Bayesian network or influence diagram. Otherwise, if the
branches of the decision tree has very different structure, the best
representation would probably be the decision tree.

If there are only a few irregularities in the decision tree, you can do
some tricks to get around it when converting to a Bayesian network
(well, you also can if there are many, but the Bayesian network would be
ugly).

In these in-between cases, it might be interesting to be able to combine
the two modelling techniques. I don't know if anybody has tried to do
that!?

Lars

-- 
 Lars Moltsen Nielsen           HUGIN Expert A/S
 Email: ln@hugin.dk             Niels Jernes Vej 10
 Phone (o): +45 96 35 45 48     DK-9220 Aalborg East
 Phone (h): +45 98 13 13 56     Denmark