In Timefold for Python I use a solving configuration with three phases: First Fit, then Simulated Annealing, then Tabu Search. The termination criteria are in terms of steps (I use both setCountLimit and unimprovedStepCountLimit). I also specify the random seed (or, if I don't specify the seed, it appears to always use seed 0). With this, should different runs of the solver on the same input give precisely the same result? Because unfortunately, that's not what happens. At the end of the First Fit phase, I already get different amounts of hard constraints between runs. There are some advantages to deterministic runs, so does anyone have ideas on how to fix this? Continue reading...
