Mark,
The approach I was looking into that time was not feasible because we
used to keep order of asserted fact/tuples on a node basis. With the
core changes you made in 3.1, we can implement range ordering by
replacing the current hashmap index for a tree index. No need for
training data, in my understanding.
Maybe we will need a composed index approach to work some cases, but
the general solution idea is simple.
[]s
Edson
Mark Proctor wrote:
Actually I was just thinking about some stuff Edson has done. With
solvers we know the available data and ranges, right? We can use this
to order indexes, I know this was something Edson looked into - but
without training data, we couldn't make it worth while - same for
custom indexing. So we can start to incorporate those to get faster
joins for known data sets.
Mark
Geoffrey De Smet wrote:
The more I learn from JCHS (or prolog for that matter),
the more I am starting to think that this is a different way of solving.
1) JCHS/prolog looks like (or is) declarative solving.
2) Taseree is actually more hybrid, the general idea behind it is:
- Drools (declarative programming) is very easy for evaluation
but very difficult for solving.
- Local/tabu search (procedural programming) is easy for solving
but difficult for evaluation.
Both have it's disadvantages and advantages, for example:
Local search is generally faster but doesn't recognize the optimal
solution.
To me it seems they are both interesting to implement,
there must be some common ground too.
We should hold a conference call about it this weekend?
It would be a good idea to compare JCHS and Taseree on a couple of
problems, like the tt problem:
http://mat.gsia.cmu.edu/TOURN/
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
--
Edson Tirelli
Software Engineer - JBoss Rules Core Developer
Office: +55 11 3124-6000
Mobile: +55 11 9218-4151
JBoss, a division of Red Hat @ www.jboss.com
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
