How to pre-check constraints conflict in timefold?

[hangocanh2303]

Hi everyone
I have a question
When the user enters multiple constraints that conflict with each other such that it is certain that no solution can satisfy both, how can we pre-check this, explain it, and notify the user early?

[triceo]

Hello @hangocanh2303,
the short answer is that you can't. To do that, you'd have to have the capability to understand the semantics of Java code. For example:

.forEach(Shift.class)
.filter(shift -> shift.isAssigned())

and:

.forEach(Shift.class)
.filter(shift -> !shift.isAssigned())

These two pieces of constraint streams would do the exact opposite, and if they are both penalized as a hard constraint, then the solution is guaranteed to break hard constraints.

The only way to figure this out reliably would be to parse the Java code, somehow figure out what the filters do, and realizing that they are in conflict. We do not have any tools for this, and I am a bit skeptical that it is feasible. Especially considering that the above example is only the most simplistic one, and real world use cases would be much harder to introspect.

[hangocanh2303]

@triceo thank you for your response. I mean:
for example in timetable scheduling
if teacher A request constraint: teach at Monday, period 1
but still this teacher A, after that request a constraint like: unavailable at Monday, period 1
how can I pre-check constraint conflict before pass this to solver?

[triceo]

As I already said - to do this generically, your code would have to perfectly understand the semantics of constraint implementation, incl. deep introspection of Java code.

The important question is: how would such a situation even happen? Why would you have conflicting constraints in your code? I'm assuming you're building a system which lets users choose their own constraints from a library. If that is the case, mark constraints as mutually conflicting, and don't let your users choose both at the same time.

If instead this conflict is at the dataset level, then you have the option of rejecting the dataset - you understand the semantics of your dataset, the solver does not. Write a series of validations to check if the dataset makes sense, before you put it into the solver.

But to reiterate - to have the solver do this generically for any problem, we would have to have a perfect understanding of user code, and that is unlikely to ever happen.