|
Nonlinear terms merit special scaling efforts. Drud, in his CONOPT implementation, pays particular attention to scaling diagnostics. His recommendations for scaling are essentially those above and he also emphasizes scaling the model so that the gradients of all nonlinear terms so exhibit absolute values close to one at the starting point and at optimality.
Thus, in nonlinear models scaling should be done so that the absolute value of the nonlinear terms as reported back by GAMS LIMROW/LIMCOL and GAMSCHK are close to one.
Several actions are then in order
| • | One needs to use a starting point as close to the optimal solution point as possible. |
| • | Use solver scaling. MINOS5, and CONOPT2 have special options controlling whether nonlinear scaling is done. CONOPT3 does default scaling. |
| • | If you are having problems then scale the model using the user defined GAMS scaling options. I have worked with cases where MINOS5 would not solve the model unless user defined scaling was done. |
|