Any idea why the score stop improving

[Magnusrk]

Here is a picture of my score over time. The pink is the most optimal score and the blue is using the default solver. Does the solver get stuck and is there ways to avoid it? I am using timefold version 19

[triceo]

Metaheuristic-based solvers such as Timefold Solver don't give a guarantee of reaching the optimal solution and what you are showing here is nothing out of the ordinary.

Typically, you can help the solver by:

• eliminating score traps from constraints,
• introducing more constraints to help the solver distinguish solutions from one another,
• and writing custom moves to help the solver overcome situations where domain-specific understanding is necessary.

Also, as a side note - if you can write a hauristic to find the optimal solution, and the optimal solution doesn't take an eternity to find, then you don't need the solver. The solver is useful in situations where exhaustive search no longer suffices.

[Magnusrk]

Thank you for the answer :) My custom heuristic phase only works on simple problems a human could also do in a few seconds. But i will try some of your suggestions. Do you think the issue also could arise from complex problems with more determining variables?

[triceo]

Typically these problems call for custom moves.

The solver has no way of understanding, for example, that worker A and worker B must always work together, one being the apprentice of the other. It will move one worker to a different time slot, only to find out it broke hard constraints, and ignore that option. Therefore:

• A score is unlikely to improve by moving just one worker.
• The solver may move one worker, score the move, and move the other worker speculatively. This can happen, but it is statistically unlikely.
• Therefore the solver is kinda stuck, because there are productive parts of the search space not accessible to it.

With domain understanding, you know about this aspect of your problem, and can write a custom move that always moves the two workers together. That, in turn, allows the solver to work efficiently.

This is obviously a simplified description. I'm sure you can imagine that, with growing complexity of the problem, situations like these can become numerous.