Originally posted by: Uonly

Hi community,

I am solving a large scale lp from a restricted master problem using primal simplex.

While I can utilize the warm-start capability, the computation from the primal simplex is slow. One of the thing I am currently trying is to tune the primal simplex method in CPLEX.

Any one can share some idea?

Thanks

Originally posted by: RWunderling

For a more to the point suggestion you would have to explain what makes primal Simplex slow, e.g. it takes lots of iterations, it looses primal feasibility, the iterations are slow,...  Anyway, let me just throw in two comments:

The simplest parameter to play with is of course the pricing strategy.

When restarting from advanced bases one typically looses the benefit of presolve.  Using CPXPARAM_Advance = 2 one can try restarting while preserving presolve.  Note, however, that the resulting starting basis in the presolve model may or may not end up being primal feasible.

Originally posted by: Uonly

Thanks Roland for your answer. Actually it was becaus of lots of iterations.

Theoretically, when doing column generation, it would be better to take advantage of warm-starting by primal simplex.

But it turns out that our testing on a class of problem is thtat it is better to use dual simplex . It was just faster.

Would you please help give some insight on how accelerate the primal simplex using warm-start?

Thanks again

Originally posted by: RWunderling

That's the thing with theory and practice in Math Programming: You never know how it works, till you try on your own model.

Lots of iteration may indicate, that you could benefit from a stronger pricing rule, e.g. Devex or Steepest Edge pricing.

If you have indications of numerical trouble (getting back into phase 1 repeatedly, messages about markowitz tolerance being increased or even hitting basis singularities) you should try turning on the numerical emphasis parameter

Roland