I'm using pyomo 6.6.4 and several solvers (Ipopt, Bonmin, Cbc,...) to solve different relatively complicated models (MINLP in general). Since my models have known infeasibility points I am solving them using several solvers or heuristics. In order to evaluate the quality of the results between different solvers or self defined solving heuristics I'm trying to build a solver comparison tool and noticed that when solving several time a model with a solver in a loop I'm always getting the same results whereas when using single runs, the results often differ.
I have tried to use different solver instances, model instances, to change the randomness generation seeds between iterations (for solvers which allow it) but nothing seems to do the trick. Any ideas?