The CH phase requires a very long time to complete, is there a method to improvement it ?

[kentbill]

I have a dataset to solver that uses PlanningListVariable. The number of planning entities is quite large, such as over 5000, and the CH phase getting slower and slower. The number of moves per step is about between 10 and 60. When more than 2000 steps are completed, it even takes a few seconds to complete one step. Is there any way to shorten it?

[Christopher-Chianelli]

Greetings @kentbill ! Without seeing your domain and constraints, I cannot give specific advice. But since only 10-60 moves are evaluated per step, I suspect you have a large cross product in the join of one of your constraints. Try using Joiners where possible, and do filtering before the join where possible. For example,

constraintFactory.forEach(Visit.class)
    .join(Visit.class)
    .filter((a,b) -> b.getVisitType() == VisitType.PRIORITY)
    .filter((a, b) -> a.getArrivalTime().isBefore(b.getArrivalTime())
    .filter((a, b) -> a.getPriority() < b.getPriority())
    .filter((a,b) -> Objects.equal(a.getVehicle(), b.getVehicle()) )
    .penalize(HardSoftScore.ONE_HARD)
    .asConstraint("...")

can be rewritten as

constraintFactory.forEach(Visit.class)
    .join(constraintFactory.forEach(Visit.class).filter(b ->  b.getVisitType() == VisitType.PRIORITY),
             Joiners.lessThan(Visit::getArrivalTime),
             Joiners.lessThan(Visit::getPriority),
             Joiners.equal(Visit::getVehicle)
    )
    .penalize(HardSoftScore.ONE_HARD)
    .asConstraint("...")

For other performance tips, see https://docs.timefold.ai/timefold-solver/latest/constraints-and-score/performance

[kentbill]

Thank you! @Christopher-Chianelli