How to configure SelectionFilter for list-variable models?

[holgerbrandl]

Hi there,

With chained models, one can configure a SelectionFilter via moveSelectorConfigList.filterClass for both local search and construction phase. However, with list-variable models, this is not possible:

java.lang.IllegalArgumentException: Construction Heuristic phase with a list variable does not support moveSelectorConfigList configuration. Remove the moveSelectorConfigList ([ChangeMoveSelectorConfig(null, ValueSelectorConfig(previousTaskOrTool))]) from the config.

What is the intended migration path?

Thanks & Kind regards,
Holger

[triceo]

Hello Holger,

the list variable and the chained variable are not functionally equivalent. Not yet anyway. But maybe there are better ways now of accomplishing those same things.

Would you mind sharing what you were attempting to do?

[holgerbrandl]

Thanks Christopher.

How can we disable the CH (and initialize the model separately)?

Is a such filter planned for CH (given the use-cases described above) to bring back the more consistent design option from chain models? (without an additional hard-constraint which only affects CH and is clutter during LS)

I don't agree with the assumption that "make sure equipment matches task" constraint is fast. In particular, in Job-Shop-Scheduling we very very often encounter a complete schedule invalidation after moves because of the interdependencies between operations of each work item. This makes it actually quite costly to evaluate a schedule vs evaluating a selection-filter.

Kind regards
Holger

[Christopher-Chianelli]

I am referring to the fact this code will be fast to evaluate:

Constraint makeSureEquipmentMatchesTask(ConstraintFactory cf) {
     return cf.forEach(Task.class)
                     .filter(task -> task.getRequiredEquipment().equals(task.getEquipment()))
                     .penalize(HardSoftScore.ONE_HARD)
                     .asConstraint("Matches equipment");
}

The complete schedule invalidation would only occur in CH (which should only be running for a short time), since LS would not pick moves that invalidate the schedule because of an invalid equipment being assigned.

To disable CH/initialize the model separately create a Custom Solver Phase and add it to your solver-config.xml:

<solver xmlns="https://timefold.ai/xsd/solver" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
    xsi:schemaLocation="https://timefold.ai/xsd/solver https://timefold.ai/xsd/solver/solver.xsd">
  <!-- ... -->
  <customPhase>
    <customPhaseCommandClass>...MyCustomPhase</customPhaseCommandClass>
  </customPhase>
  <localSearch>
    <unionMoveSelector>
      <listChangeMoveSelector>
        <filterClass>...Filter</filterClass>
          <!-- ... -->
      </listChangeMoveSelector>
      <!-- ... -->
    </unionMoveSelector>
  </localSearch>
</solver>

In regards to plans, I agree it is a valid issue; I created #823 to track it.

[holgerbrandl]

Thanks Christopher for your kind support and guidance.

[terbrack]

Hi everyone,
only a short additional note on the above stated blocker:
in a JS scheduling problem with list variables, we model the Variable Listener with the assumption of a valid plan, i.e. task-tool-assignment.
Due to the missing move selector config, invalid task-tool-assignments occur in the construction phase, leading to a score corruption.
Based on that, I am looking forward to a progress on #823.
Best,
Hajo

[triceo]

@terbrack If you're seeing move corruptions, I doubt that it is because of missing move filters. Score corruptions come from wrong implementations of variable listeners, custom moves, or constraints. A correct implementation of either of those does not cause score corruptions.

Also, to set the correct expectations here - we are aware of #823, but it is currently not on the roadmap. Timefold Solver is an open source project though; pull requests are welcome.