For LP (and RMIP) models that cause difficulty for the standard solvers, quadMINOS provides considerably greater reliability and accuracy than MINOS by using quadruple-precision floating-point arithmetic. Although it is slower by a factor of 20 or 30 for cold-start runs, you can use standard MINOS or another LP solver first and let them do their best. Warm-starts with quadMINOS may give greatly improved solutions at moderate cost.
To take advantage of quadMINOS, the main requirement is to set a few runtime options to non-default values in order to request higher accuracy. For example, it is reasonable to set
Feasibility tolerance 1e-15 Optimality tolerance 1e-15
because IEEE Quad floating-point has nearly 34 digits of precision. For an example of using Double and Quad MINOS together, see [DQQ] in the GAMS model library.
- Note
- GAMS/quadMINOS can only solve LPs and RMIPs at the moment because GAMS does not provide function evaluations in quad-precision.
The documentation of MINOS also refers to quadMINOS.