4:24 p.m.
Hi,
I've been using Drools Planner with success for a while for our Conversation
planning tool. It's used by schools to plan conversations between teachers
and parents of schoolgoing children.
Recently I checked on the new developments on Drools (the current version
we're using in production is 5.4.0.FINAL) and started to upgrade to
OptaPlanner 6.0.0.Beta5.
Everything looked fine, and seeing how it's now easier to check which
constraints are broken in a given solution, I decided to have a look at that
as well.
And that's where I noticed that we actually had an issue with score
corruption in one of our rules...
The goal of the rule in case is to minimize the amount of timeslots between
the first and last Conversation of a family (they can have conversations for
multiple children.)
The rule uses custom helper objects, like in the NurseRostering example,
which are added with insertLogical.
Conversation is the planning entity, which has the planning variable:
timeslot. FamilyStart and FamilyEnd are custom helper objects to register
the timeslot of the first and last conversation for a family.
The relevant rules are as follows:
rule "familyStart"
salience 1
when
Conversation(
family != null, $family: family,
timeslot != null, $timeslot: timeslot);
not Conversation(
family == $family,
timeslot < $timeslot);
then
insertLogical(new FamilyStart($family, $timeslot));
end
rule "familyEnd"
salience 1
when
Conversation(
family != null, $family: family,
timeslot != null, $timeslot: timeslot);
not Conversation(
family == $family,
timeslot > $timeslot);
then
insertLogical(new FamilyEnd($family, $timeslot));
end
rule "familyCompact"
when
$start: FamilyStart(
family != null, $family: family,
timeslot != null, $timeslotBegin: timeslot);
$end: FamilyEnd(
family == $family,
(timeslot.id - $timeslotBegin.id > family.minimumTimeslotsRequired),
timeslot != null, $timeslotEnd: timeslot);
then
scoreHolder.addSoftConstraintMatch(kcontext, -(($timeslotEnd.getId() -
$timeslotBegin.getId()) - $family.getMinimumTimeslotsRequired()));
end
When running one of my example schedules in FULL_ASSERT mode, the following
error appears during the LocalSearch phase:
Exception in thread "main" java.lang.IllegalStateException: Score
corruption: the workingScore (-2hard/-119soft) is not the uncorruptedScore
(-2hard/-118soft) after completedAction (Teacher 3 family 11 timeslot 2 =>
12):
The corrupted scoreDirector has 1 ConstraintMatch(s) which are in excess
(and should not be there):
defaultpkg/familyCompact/level1/[FamilyEnd Family 11, Timeslot 12,
FamilyStart Family 11, Timeslot 7]=-1
The corrupted scoreDirector has no ConstraintMatch(s) which are missing.
Check your score constraints.
I use a custom ChangeMove and SwapMove implementation. I've also tried using
the new generic Moves, these have the same problem.
It seems like when a move happens that should update an inserted FamilyStart
or FamilyEnd because of a change in timeslot, this change isn't properly
reflected in the scoreDirector.
Do you have any idea what could cause this corruption? The moves do call
beforeVariableChanged and afterVariableChanged on the scoreDirector, and the
FamilyStart and FamilyEnd classes implement an equals and hashcode based on
the Family and Timeslot fields.
(Actually, I've also tried using the default equals and hashcode of Object
for these 2 classes, which causes corruption much faster and with more
ConstraintOccurrences.)
Thanks for any insights,
Pieter
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
4:54 p.m.
On 16-07-13 16:24, pvandenbrink wrote:
Great to hear :)
Sounds like an interesting use case story.
Do you have this issue in 5.4 too? E.g.: the 5.4
equivalent of full
assert (TRACE) does not detect score corruption?
Maybe that "[FamilyEnd Family 11,
Timeslot 12, FamilyStart Family 11,
Timeslot 7]" exists twice?
Once way to find out is to put a debug breakpoint at the throw of the
exception
and compare the ConstraintMatchTotal's ConstraintMatch's justificationList.
Interesting: that proves
it's not a custom move problem.
Your test
above proves that's ok. Use generic moves for further tests
first pls.
and the
FamilyStart and FamilyEnd classes implement an equals and hashcode based on
the Family and Timeslot fields.
1) But how do Family and Timeslot implement their
equals/hashcode?
My first guess would be to start looking at this.
2) There's also the remote probability there's a bug in drools - in
which case we'd want a reproducer attached on a jira so we can fix it.
HTH :)
11:59 a.m.
ge0ffrey wrote
Do you have this issue in 5.4 too? E.g.: the 5.4 equivalent of full
assert (TRACE) does not detect score corruption?
I've tried this and actually, yes, it happens with TRACE in 5.4 as well. So
it's not an issue introduced by the upgrade to Beta5, then.
ge0ffrey wrote
Ok, using the generic changeMoveSelector and swapMoveSelector the error is
almost the same, just a slight difference in timeslot values:
Exception in thread "main" java.lang.IllegalStateException: Score
corruption: the workingScore (0hard/-123soft) is not the uncorruptedScore
(0hard/-120soft) after completedAction (Conversation 24: Family 11, Student
24, Teacher 3, Timeslot 14(Day 3) => Timeslot 14):
The corrupted scoreDirector has 1 ConstraintMatch(s) which are in excess
(and should not be there):
defaultpkg/familyCompact/level1/[FamilyEnd Family 11, Timeslot 14,
FamilyStart Family 11, Timeslot 7]=-3
The corrupted scoreDirector has no ConstraintMatch(s) which are missing.
Check your score constraints.
I noticed two things:
1) The toString of the move claims it moved the conversation from Timeslot
14 to 14, I think it actually prints the updated timeslot twice, rather than
the original one and the updated one. My own change move had the same issue,
which I fixed by storing the original timeslot of the Conversation in a
separate field of the move and reading from that field in the toString.
2) These two constraint matches exist in the corrupted score map:
defaultpkg/familyCompact/level1/[FamilyEnd Family 11, Timeslot 14,
FamilyStart Family 11, Timeslot 7]=-3
defaultpkg/familyCompact/level1/[FamilyEnd Family 11, Timeslot 14,
FamilyStart Family 11, Timeslot 5]=-5
So there are two different FamilyStarts: (Family 11, Timeslot 7) and (Family
11, Timeslot 5), where there should be only one (the one for timeslot 5).
This causes familyCompact to fire twice for the same family, which explains
the score corruption... But what could cause two FamilyStarts to be
inserted?
ge0ffrey wrote
Family, Timeslot and Conversation implement an equals and hashcode based on
their numerical id, like so:
@Override
public boolean equals(Object other)
{
if ((this == other))
return true;
if (!(other instanceof Timeslot))
return false;
Timeslot castOther = (Timeslot) other;
return new EqualsBuilder().append(this.getId(),
castOther.getId()).isEquals();
}
@Override
public int hashCode()
{
return new HashCodeBuilder().append(getId()).toHashCode();
}
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
12:12 p.m.
On 17-07-13 11:59, pvandenbrink wrote:
Yes, this is a known issue that their
toString() after doing the move
doesn't show what it did
(but not easily fixable because ChangeMove's can be cached and reused).
The 7 one probably got inserted with it didn't reach to 5 yet. That
was
a good logicalInsertion.
The issue is, why didn't it get retracted when a move made changes so it
did reach to 5?
This one looks good, but I am mostly interested in those of the
logically inserted objects (FamilyStart, etc).
1:02 p.m.
ge0ffrey wrote
I've added a printline in the rules xml file before the insertLogical. At
one point shortly before the corruption occurs the following two lines
appear immediately one after another:
inserting FamilyStart for Family 11 at tijdslot Timeslot 5
inserting FamilyStart for Family 11 at tijdslot Timeslot 7
This seems strange, it's as if they're inserted almost at the same time.
ge0ffrey wrote
This one looks good, but I am mostly interested in those of the
logically inserted objects (FamilyStart, etc).
FamilyStart implements it as follows:
@Override
public boolean equals(Object other)
{
if ((this == other))
return true;
if (!(other instanceof FamilyStart))
return false;
FamilyStart castOther = (FamilyStart) other;
return new EqualsBuilder().append(this.getFamily(), castOther.getFamily())
.append(this.getTimeslot(), castOther.getTimeslot()).isEquals();
}
@Override
public int hashCode()
{
return new
HashCodeBuilder().append(getFamily()).append(getTimeslot()).toHashCode();
}
Similarly, FamilyEnd:
@Override
public boolean equals(Object other)
{
if ((this == other))
return true;
if (!(other instanceof FamilyEnd))
return false;
FamilyEnd castOther = (FamilyEnd) other;
return new EqualsBuilder().append(this.getFamily(), castOther.getFamily())
.append(this.getTimeslot(), castOther.getTimeslot()).isEquals();
}
@Override
public int hashCode()
{
return new
HashCodeBuilder().append(getFamily()).append(getTimeslot()).toHashCode();
}
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
2:25 p.m.
On 17-07-13 13:02, pvandenbrink wrote:
Did you enable logging in OptaPlanner? See docs on logging. Use trace
logging and it will output a line for every move it tries. That will
prove or disprove that those 2 println's are happening as a result of
the same move. (I presume they won't).
Looks good.
This means there might be a bug in drools expert after all.
Can you make a reproducer and file a jira? For our convenience, try to
isolate as much as possible:
- remove all other rules
- try to reduce the dataset as much as possible (as long as it reproduces)
- using TRACE logging, try to bring down the number for steps to 1 (by
saving the dataset just before the step which goes wrong and then load
that one).
- Don't worry about reducing the number of moves to as little as
possible for that 1 step - I can hack that in DefaultDecider.
5:26 p.m.
ge0ffrey wrote
I was using debug level tracing, changed it to trace.
The above two printlines appear after the last accepted move, before the
exception is thrown:
17-07-13 17:13:51 TRACE - DefaultDecider - Move
index (118), score (0hard/-121soft), accepted (true) for move (Conversation
4: Family 4, Student 4, Teacher 1, Timeslot 2(Day 1) => Timeslot 14).
I'll look into creating a reproducer for a jira issue.
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
11:25 a.m.
Ok, I've created a small reproducer and attached it to this issue:
https://issues.jboss.org/browse/PLANNER-182
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
4:47 p.m.
Turns out this is a user problem in the equals and hashcode of
FamilyStart, not a bug in Drools or OptaPlanner :)
@Override
public boolean equals(Object other)
{
if ((this == other))
return true;
if (!(other instanceof FamilyStart))
return false;
FamilyStart castOther = (FamilyStart) other;
return new EqualsBuilder()
.append(this.getTimeslot(), castOther.getTimeslot()) // <== must be Family
.append(this.getTimeslot(), castOther.getTimeslot())
.isEquals();
}
@Override
public int hashCode()
{
return new HashCodeBuilder()
.append(getTimeslot()) // <== must be Family
.append(getTimeslot())
.toHashCode();
}
On 19-07-13 11:56, Geoffrey De Smet wrote:
5:05 p.m.
ge0ffrey wrote
Thanks for looking into it, and you're right. Looks like I made an error in
the reproducer project, sorry... The original code didn't have this
equals/hashcode problem. After fixing this in the reproducer, the example
schedule works fine though, so I'll have to have another look to see if I
can get it to reproduce again with another example dataset.
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
2:57 p.m.
ge0ffrey wrote
Yes, please reopen the jira issue if you're able to reproduce it
with
6.0.0.Beta5 or higher.
Done, I've modified the dataset, so the issue's reproducible again in
6.0.0.Beta5 and 6.0.0.SNAPSHOT, with the corrected equals/hashcode for
FamilyStart.
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
10:20 p.m.
You're welcome, good to hear that it's fixed already!
--
View this message in context:
http://drools.46999.n3.nabble.com/OptaPlanner-score-corruption-when-using...
Sent from the Drools: User forum mailing list archive at Nabble.com.
4239
days inactive
4258
days old
14 comments
2 participants
participants (2)
-
Geoffrey De Smet -
pvandenbrink