Bruce D'Ambrosio wrote:
>
> As described in my AIMag article last year, SPI (symbolic
> probabilistic inference) treats individual marginals as just a special
> case of the general optimization problem of finding efficient ways to
> compute arbitrary subjoints, and does a surprisingly good job at
> arbitrary subnjoints. Variable elimination should work also.
>
> The open research problem that no-one is working on (to my knowledge)
> is to find efficient ways of computing an arbitrary collection of
> arbitrary
> subjoints. No existing method seems to provide an interesting
> starting point for that problem.
>
> cheers - Bruce
I would do as follows: Start off with your best junction tree and
performing a propagation of variables ( a collect to some root - let it
be lazy or no-lazy). Before I do so, I calculate the cost for each
clique of using that as the root. This is done by sending messages of
costs around in the junction tree ( easy arithmetic, one propagation is
sufficient), and then you choose the cheapest root.
Doesn't this work?
/Finn
This archive was generated by hypermail 2b29 : Fri Mar 31 2000 - 12:41:49 PST