A score corruptoin from the Constraint - As identical finish time as possible

[kentbill]

Hi, everyone.

I'm working on a JobShop scheduling case where it assigns jobs to appropriate production lines. One of the constraints is to ensure that the finish times of jobs across all production lines are as similar as possible(identical is perfect).

The constraint is:

// Make the end times of each production line as identical as possible
Constraint balanceProductionLine_endTime(ConstraintFactory factory) {
	return factory.forEach(ProductionLine.class)
		.groupBy(ConstraintCollectors.max(ProductionLine::getEndTimeDuration))
		.penalizeBigDecimal(HardMediumSoftBigDecimalScore.ONE_SOFT, maxEndTime -> BigDecimal.valueOf(maxEndTime))
		.asConstraint("Identical end time");
}

The approach is to use the number of minutes between the end time of a production line (the end time of the last job) and 00:00:00 on 2025-01-01 as the production line minutes, obtain the maximum production line minutes in the constraints using the ConstraintCollectors.max method, and use that number of minutes as the penalty score.

However, it triggered a Score Corruption exception. Does anyone know why?

15:17:34.899 [main        ] INFO  Solving started: time spent (23), best score (0hard/0medium/0soft), environment mode (FULL_ASSERT), move thread count (NONE), random (JDK with seed 0).
15:17:34.901 [main        ] INFO  Problem scale: entity count (3), variable count (3), approximate value count (8), approximate problem scale (20,160).
15:17:34.919 [main        ] DEBUG     CH step (0), time spent (43), score (0hard/0medium/-6316140soft), selected move count (2), picked move (JobO1 {null -> Line2[0]}).
15:17:34.923 [main        ] DEBUG     CH step (1), time spent (47), score (0hard/0medium/-12666540soft), selected move count (1), picked move (JobO2 {null -> Line1[0]}).
15:17:34.933 [main        ] DEBUG     CH step (2), time spent (57), score (0hard/0medium/-6316140soft), selected move count (2), picked move (JobO3 {null -> Line1[0]}).
15:17:34.939 [main        ] DEBUG     CH step (3), time spent (63), score (0hard/0medium/-6316140soft), selected move count (2), picked move (JobO4 {null -> Line2[0]}).
15:17:34.951 [main        ] DEBUG     CH step (4), time spent (75), score (0hard/0medium/-6921060soft), selected move count (4), picked move (JobO5 {null -> Line2[0]}).
15:17:34.955 [main        ] DEBUG     CH step (5), time spent (79), score (0hard/0medium/-26635080soft), selected move count (1), picked move (JobO6 {null -> Line3[0]}).
15:17:34.958 [main        ] INFO  Construction Heuristic phase (0) ended: time spent (82), best score (0hard/0medium/-26635080soft), move evaluation speed (210/sec), step total (6).
15:17:34.958 [main        ] DEBUG Constraint weights for solution (1):
  Constraint (com.easyplan.domain/Job has looped shadow variables.) weight set to (0hard/-1medium/0soft).
  Constraint (com.easyplan.domain/As identical finish time as possible) weight set to (0hard/0medium/-1soft).
  Constraint (com.easyplan.domain/Minimize make span) weight set to (0hard/0medium/-1soft).
15:17:34.968 [main        ] DEBUG     LS step (0), time spent (92), score (0hard/0medium/-25177080soft), new best score (0hard/0medium/-25177080soft), accepted/selected move count (1/1), picked move (JobO1 {Line2[2] -> Line1[2]}).
Exception in thread "main" ai.timefold.solver.core.impl.solver.exception.ScoreCorruptionException: Score corruption (120soft): the workingScore (0hard/0medium/-26818680soft) is not the uncorruptedScore (0hard/0medium/-26818800soft) after completedAction (2-Opt(firstEntity=Line2, secondEntity=Line2, firstEndpointIndex=0, secondEndpointIndex=1)):
Score corruption analysis:
  The corrupted scoreDirector has 1 ConstraintMatch(es) which should not be there:
    com.easyplan.domain/As identical finish time as possible/[265080]=0hard/0medium/-265080soft
  The corrupted scoreDirector has 1 ConstraintMatch(es) which are missing:
    com.easyplan.domain/As identical finish time as possible/[265200]=0hard/0medium/-265200soft
  Maybe there is a bug in the score constraints of those ConstraintMatch(s).
  Maybe a score constraint doesn't select all the entities it depends on,
    but discovers some transitively through a reference from the selected entity.
    This corrupts incremental score calculation,
    because the constraint is not re-evaluated if the transitively discovered entity changes.
Shadow variable corruption in the corrupted scoreDirector:
  None
	at ai.timefold.solver.core.impl.score.director.AbstractScoreDirector.assertScoreFromScratch(AbstractScoreDirector.java:745)
	at ai.timefold.solver.core.impl.score.director.AbstractScoreDirector.assertWorkingScoreFromScratch(AbstractScoreDirector.java:719)
	at ai.timefold.solver.core.impl.score.director.AbstractScoreDirector.lambda$executeTemporaryMove$1(AbstractScoreDirector.java:333)
	at ai.timefold.solver.core.impl.move.director.MoveDirector.executeTemporary(MoveDirector.java:225)
	at ai.timefold.solver.core.impl.score.director.AbstractScoreDirector.executeTemporaryMove(AbstractScoreDirector.java:328)
	at ai.timefold.solver.core.impl.localsearch.decider.LocalSearchDecider.doMove(LocalSearchDecider.java:117)
	at ai.timefold.solver.core.impl.localsearch.decider.LocalSearchDecider.decideNextStep(LocalSearchDecider.java:96)
	at ai.timefold.solver.core.impl.localsearch.DefaultLocalSearchPhase.solve(DefaultLocalSearchPhase.java:89)
	at ai.timefold.solver.core.impl.solver.AbstractSolver.runPhases(AbstractSolver.java:82)
	at ai.timefold.solver.core.impl.solver.DefaultSolver.solve(DefaultSolver.java:196)
	at com.easyplan.Main.main(Main.java:206)

P.S.: Due to the Predecessor/Successor job, I used DSV.

The following is the complete code
https://github.com/kentbill/dsv-testing.git

Thank you in advanced!

[kentbill]

I think I got the cause now. It's similar to the discussion #1958. For a ProductionLine, its JobList may not have changed at all, but because the predecessor job of a Job on that ProductionLine has changed, the end time has changed, and this change has not been detected, so a Score Corruption has occurred.

If that's the case, how should I implement this constraint? That is, I want workers on each production line to finish work and be off duty at the same time as much as possible. For example, workers on production line 1 work all day, while workers on production line 2 take a break in the morning. Therefore, I want to assign more work to production line 1 so that they can leave at the same time.

[Christopher-Chianelli]

penalizedLoopedJobs is incorrect:

    Constraint penalizeLoopedJobs(ConstraintFactory factory) {
        return factory.forEach(Job.class)
                .filter(Job::isInconsistent)
                .penalizeBigDecimal(HardMediumSoftBigDecimalScore.ONE_HARD)
                .asConstraint("Job has looped shadow variables.");
    }

It must use forEachUnfiltered, since forEach does not include inconsistent entities (since otherwise, you would need to include an inconsistent check in basically every constraint).

The correct constraint would be

    Constraint penalizeLoopedJobs(ConstraintFactory factory) {
        return factory.forEachUnfiltered(Job.class)
                .filter(Job::isInconsistent)
                .penalizeBigDecimal(HardMediumSoftBigDecimalScore.ONE_HARD)
                .asConstraint("Job has looped shadow variables.");
    }

[kentbill]

Thank you! @Christopher-Chianelli
Use forEach, the startTime and endTime in some jobs are null.

On the other hand, I want the end time of the last job on each production line to be as similar as possible. Do you think this constraint is correct?

    Constraint balanceProductionLine_endTime1(ConstraintFactory factory) {
        return factory.forEach(Job.class)
                .groupBy(Job::getProductionLine, ConstraintCollectors.max(Job::endTimeDuration))
                .penalizeBigDecimal(HardMediumSoftBigDecimalScore.ONE_SOFT, (productionLine, maxEndTime) -> BigDecimal.valueOf(maxEndTime))
                .asConstraint("As identical finish time as possible");
    }

[Christopher-Chianelli]

I solved with these constraints and did not get any Jobs with null start/end times.
How would you want these two schedules to be penalizes?

• Line 1 ends at 10, Line 2 ends at 8, Line 3 ends at 6
• Line 1 ends at 10, Line 2 ends at 9, Line 3 ends at 1

With the above constraint, these two schedules would be penalized the same, despite the first one being more "fair".
You might want to consider ConstraintCollectors.loadBalance (you would need to convert the duration to a long).

[kentbill]

Yep, the following constraint seems to work.

Constraint balanceProductionLine_endTime(ConstraintFactory factory) {
        return factory.forEach(Job.class)
                .groupBy(Job::getProductionLine, ConstraintCollectors.max(Job::getEndTimeDuration))
                .complement(ProductionLine.class, pl -> 0)
                .groupBy(ConstraintCollectors.loadBalance((productionLine, jobMaxEndTime) -> productionLine,
                        (productionLine, jobMaxEndTime) -> jobMaxEndTime))
                .penalizeBigDecimal(HardMediumSoftBigDecimalScore.ONE_MEDIUM, LoadBalance::unfairness)
                .asConstraint("Balance the end time");
    }

Thanks!

[kentbill]

I solved with these constraints and did not get any Jobs with null start/end times. How would you want these two schedules to be penalizes?

• Line 1 ends at 10, Line 2 ends at 8, Line 3 ends at 6
• Line 1 ends at 10, Line 2 ends at 9, Line 3 ends at 1

With the above constraint, these two schedules would be penalized the same, despite the first one being more "fair". You might want to consider ConstraintCollectors.loadBalance (you would need to convert the duration to a long).

em, using the end time as an unfair calculation field does not work in some cases, and I need to find other ways. The goal is to make the end time of all production lines as similar as possible after taking into account the different maintenance times on each production line.