I am using GLPK to solve a number of LP's in succession in a single program. The LP is modified after each glp_simplex_solve. New constraints are added to it. After some 10's of such solves I start to get warnings like this in the output :
And then one of the solve returns NaN as the objective value. When the model reports NaN I dump out the model and exit the program. This model (which reported NaN in GLPK) solves to optimality when solved using Gurobi, which means the Model was actually feasible and bounded. That means the NaN was because of the numerical instability in the glpk model.
Is there any way the numerical stability of the model can be reinstated e.g. resetting the model after every few solves?? Are there any other robust open-source solvers that can solve thousands of LP's in succession without numerical instability. How does glpk compare to lp-solve or COIN OR in such a situation?