IIRC, only CPLEX could work with infeasible starting point. If feeding infeasible starting point into Gurobi or COPT or other MIP solvers, the starting point would not be respected thus those solvers would still start with their own initial guesses. So my question is, for a large and hard MIP problem (like SCUC), how to make sure or are we sure that the ML learner would always give the feasible solution?
Great work by the way.