Xpress global solver

Support
1

Hello,

I am having problems to use Xpress global solver on MINLP problems. For some reason, when called from AMPL, the nonlinearities of the problem are approximated with piece-wise linear functions before being solved and, thus, I do not obtain the same high quality solutions that I can obtain with other global solvers. Is there any way to force XPRESS to use the global solver without having AMPL reformulating the problem first as a MILP? I have seen in past posts that you were working on this issue, so I wonder whether there is already some way to get it done.

For instance, if I run:

var x>=0;
minimize objfun: exp(x)+100*cos(x);
option solver xpress;
solve;

I get

XPRESS 9.4.2 (43.01.03): 
------------ WARNINGS ------------
WARNING.  2 case(s) of "PLApprox". One of them:
  An expression of type 'Exp' has been
piecewise-linearly approximated. Set cvt:plapprox:reltol
to control precision (currently 0.010000).

WARNING:  "PLApproxDomain"
  Argument domain of a 'Exp'
has been reduced to [0.000000, 13.815511] for numerical reasons
(partially controlled by cvt:plapprox:domain.)

XPRESS 9.4.2 (43.01.03): optimal solution; objective -78.93351152
12 simplex iterations
1 branching nodes
absmipgap=7.89335e-05, relmipgap=1e-06

And the obtained solution is 2.97505 with objective value -79.0258, where as the optimal solution is 2.94946 with objective value -79.0643.

Any suggestion or workaround would be much appreciated.

Thanks in advance.
Best regards,
Julio.

2

Currently, adding acc:_all=2 to the xpress_options string triggers global optimization and gets a much more accurate result (x = 2.95302, objfun = -79.0635).

We’re still working on a fix that can replace this workaround.

3

Thanks a lot Bob for the answer.
Best,
Julio.