5:24 a.m.
Hi Geoffrey,
I'm developing the algorithm for Vehicle Routing Problem with Time Windows (written by
Gendreau Herz and Laporte). To do it I need a new type of scoring and I have developed
DOUBLE_SCORE (double hard and double soft).
If you wont I can send you this code.
I write you because have a problem with solving time....
I'm testing this algorithm on Solomon's data set (100 nodes and 10-20 vehicle) but
the execution time of solver is very hight and often after one day not end....(termination
criteria is steps number).
There are twelve rules and the number of moves is #nodes*#Vehicle^2 but with
relativeSelection may decrease and I'm using Taboo Search.
I wont to ask you if for your experience the drools-solver can solve a problem of this
dimension.
regards
Marco Caminiti
12:56 p.m.
With kind regards,
Geoffrey De Smet
dmzpippo-drools(a)yahoo.it schreef:
Hi Geoffrey,
I'm developing the algorithm for Vehicle Routing Problem with Time Windows (written
by Gendreau Herz and Laporte). To do it I need a new type of scoring and I have developed
DOUBLE_SCORE (double hard and double soft).
If you wont I can send you this code.
Nice :) Did you copy from DefaultHardAndSoftScore? I recommend to do
that so the hashcode, equals etc are correct.
I don't know if we want a build-in HardAndSoftScore based doubles in
drools planner:
- doubles are inherently rounding-prone (try adding 0.09 and 0.01)
- BigDecimal is better, but slower
Don't let it stop you from using it yourself though :)
Feel free to make an issue anyway, so other can vote on it.
If there are enough votes I'll add it anyway.
All in all, the performance of the score doesn't matter that much.
Because even though it does impact performance to not use ints (like
3%), premature optimization is the root of all evil.
So use the score implementation which fits you best.
I write you because have a problem with solving time....
1) Do you have a StartingSolutionInitializer?
Take a look at CurriculumCourseStartingSolutionInitializer.
2) Are you using relativeSelection?
Those 2 are the 2 main ingredients for getting feasible solutions.
I'm testing this algorithm on Solomon's data set (100 nodes and 10-20 vehicle)
but the execution time of solver is very hight and often after one day not
end....(termination criteria is steps number).
There are twelve rules and the number of moves is #nodes*#Vehicle^2 but with
relativeSelection may decrease and I'm using Taboo Search.
I wont to ask you if for your experience the drools-solver can solve a problem of this
dimension.
It should definitly be able to get feasible solutions for this problem.
Avoid a "score trap" (see manual).
I prefer to use a time based termination criteria.
Turn the log level on info, you should see a step occur between 1 or 2
seconds. Lower your relativeSelection if a step takes more then 2 seconds.
I 'll implement absoluteSelection soon.
Try the benchmarker.
Soon the benchmarker will support a time VS bestscore graph
as you can see in my devoxx movie:
http://beta.parleys.com/#st=5&id=1714
regards
Marco Caminiti
_______________________________________________
rules-users mailing list
rules-users(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users
3:55 p.m.
Hi Geoffrey,
1) Do you have a StartingSolutionInitializer?
I don't use StartingSolutionInitializer but I generete a XML file with the initial
solution. Is not the best way but it should not affect the running time of the solver
2) Are you using relativeSelection?
I'm using relativeSelection and after your suggestion I set it to have a step every 1
or 2 seconds
But now I have another problem. If I didn't misunderstand, to select the move at the
step i the solver makes a move ->fires the rules to evaluate the score -> makes the
undo move. This happens with all moves and then the decider selectes the best move it
makes it and gets the associated score.
The problem is that I have an object that modifies an attribute in the mothod doMove (and
with the undoMove I reset the attribute) but the same object has other attributes those
are modified in the rules...
Not to change these attributes for any attempted move I think to modify the execution of
the solver thus
for every movemake a move (modify the first attribute)
fire the rules (modify other attributes)
make a undoMove (reset all attributes)the decider selects a move, it makes it and re-fire
the rules to calculate the score and to modify consistently attributesI do not know an
alternative way to achieve this goal because if I move the amendment of all attributes in
the method doMove I not have rules any more in the DRL file.
regards
Marco Caminiti
6:12 a.m.
With kind regards,
Geoffrey De Smet
dmzpippo-drools(a)yahoo.it schreef:
Hi Geoffrey,
> 1) Do you have a StartingSolutionInitializer?
I don't use StartingSolutionInitializer but I generete a XML file with
the initial solution. Is not the best way but it should not affect the
running time of the solver
> 2) Are you using relativeSelection?
I'm using relativeSelection and after your suggestion I set it to have a
step every 1 or 2 seconds
But now I have another problem. If I didn't misunderstand, to select the
move at the step i the solver makes a move ->fires the rules to evaluate
the score -> makes the undo move. This happens with all moves and then
the decider selectes the best move it makes it and gets the associated
score.
Turn on verifyUndoMoveIsUncorrupted, see
https://jira.jboss.org/jira/browse/JBRULES-2391
The problem is that I have an object that modifies an attribute in the
mothod doMove (and with the undoMove I reset the attribute) but the same
object has other attributes those are modified in the rules...
Score rules shouldn't modify the facts.
They should only change the ConstraintOccurrence's and the scoreCalculator.
The doMove method can change the facts, but the undoMove method should
undo all those changes.
Change = add, modify and/or delete
Not to change these attributes for any attempted move I think to
modify
the execution of the solver thus
* for every move
o make a move (modify the first attribute)
o fire the rules (modify other attributes)
o make a undoMove (reset all attributes)
* the decider selects a move, it makes it and re-fire the rules to
calculate the score and to modify consistently attributes
I do not know an alternative way to achieve this goal because if I move
the amendment of all attributes in the method doMove I not have rules
any more in the DRL file.
regards
Marco Caminiti
------------------------------------------------------------------------
_______________________________________________
rules-users mailing list
rules-users(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users
5447
days inactive
5450
days old
3 comments
2 participants
participants (2)
-
dmzpippo-drools＠yahoo.it -
Geoffrey De Smet