Ways to set a "good" time-based termination policy based on problem complexity?

[greyhairredbear]

As already hinted in another discussion, I have a question regarding termination configurations.

Our users might try to solve problems of different complexity - from a few vehicles/loads to hundreds of loads/dozens of vehicles. For more complex problems, users are also prepared to wait for a longer time (e.g. an hour) for planning to finish, while this is not the case for simpler problems (only a couple of loads on one vehicle).

Is there a way to set a timeout based on the complexity of the problem instance? I know that the benchmarker has some value for problem size - would it be the way to go to determine "good" timeouts for various problem sizes and then interpolate or use a step function for different complexity levels? Also, is the problem size part of the public API of Timefold?

[triceo]

The short answer is that there is no way to set a timeout based on the solution itself. Not directly.
The unimproved time termination is closest to what you need - at least you can benchmark and guesstimate how much time it would take for a particular dataset to flatline.

Arguably if you already know how long they're willing to wait for, just set that as the maximum time spent, and set unimproved time spent to a fraction of that.

[greyhairredbear]

The unimproved time termination is closest to what you need - at least you can benchmark and guesstimate how much time it would take for a particular dataset to flatline.

Ok, so basically what I've guessed. Just thought that the ProblemSizeStatistics record could have been helpful for that. But I guess calculating the search space size manually beforehand and then adjusting the timeout depending on that should be possible in my case (though it will require a different data type than Int 😅 - still might be able to scale the timeout based on that).

Arguably if you already know how long they're willing to wait for, just set that as the maximum time spent, and set unimproved time spent to a fraction of that.

As mentioned, this is where my main problem lies. Depending on the problem, this value might just be a few seconds (e.g. just plan a single vehicle with 10 loads to 10 different locations) or even an hour (hundreds of loads to different locations).