triceo
commented
1 week ago Thanks for the report, @jason076!
We are aware that the time gradient makes Simulated Annealing painful to work with. It is one of the reasons why Late Acceptance is the default - I recommend using LA instead of SA, since tweaking SA is not on our roadmap at the moment.
Even when using LA though, time-based terminations are problematic because they aren't reproducible on different hardware etc. We're currently in the process of designing a termination that doesn't require a time at all.
Is your feature request related to a problem? Please describe. Finding the right termination config for time gradient based algorithms like simulated annealing is pretty hard especially for benchmarks. If I configure a maximum solving time of 10 minutes and a unimproved seconds spent limit of 30 seconds the 30 seconds spent limit is influencing the progress of the time gradient. The same is true if I configure a best score limit. It is not possible to configure an unimproved seconds and best score spent limit without influencing the time gradient.
This is especially problematic for benchmarking. I want to determine how long my algorithm takes until it reaches a specific score and I want it to behave like the production environment. But if I define a best score limit this influences the time gradient and therefore the algorithm behaves different than without the limit.
Describe the solution you'd like I like to be able to configure a time based spent limit together with unimproved seconds or best score limit without effecting the time gradient. E.G. if I configure 10 minutes spent limit with 30 seconds unimproved spent limit, the time gradient should only depend on the progress of the 10 minutes. Maybe you could add a boolean flag to the termination configs to let the user decide if a termination criterion should be considered by the time gradient calculation.
Describe alternatives you've considered I don't know any working alternative.