Occasionally the NLP solver CONOPT will terminate with:
| Iter Phase Ninf Objective RGmax NSB Step InItr MX OK
823 3 1.6227533982E+01 3.0E-02 407 0.0E+00 F F
824 3 1.6227533982E+01 3.0E-02 401 0.0E+00 F F
825 3 1.6227533982E+01 3.0E-02 405 0.0E+00 F F
826 3 1.6227533982E+01 3.0E-02 407 0.0E+00 F F
827 3 1.6227533982E+01 3.0E-02 407 ** Feasible solution. Convergence too slow. The change in objective
has been less than 4.8683E-11 for 20 consecutive iterations |
In some cases we can reach a (local) optimum just by solving again:
| solve ramsey maximizing AggW using nlp;
solve ramsey maximizing AggW using nlp; |
Indeed this trick actually works in this case and the second model solves quickly and terminates with:
| Iter Phase Ninf Objective RGmax NSB Step InItr MX OK
193 4 1.6948156578E+01 2.7E+00 342 2.0E+00 1 F T
194 4 1.6948221678E+01 1.4E-03 441 1.0E+00 1 F T
195 4 1.6948492083E+01 1.9E-03 441 1.0E+00 6 T T
196 4 1.6949366978E+01 1.3E-03 440 1.0E+00 24 F T
197 4 1.6950307538E+01 3.1E-03 440 1.0E+00 113 F T
198 4 1.6950345143E+01 1.4E-04 440 1.0E+00 171 F T
199 4 1.6950345450E+01 1.6E-05 440 1.0E+00 136 F T
200 4 1.6950345453E+01 1.6E-06 440 1.0E+00 101 F T
201 4 1.6950345453E+01 1.0E-07 440 1.0E+00 7 F T
202 4 1.6950345453E+01 5.6E-08 440 ** Optimal solution. Reduced gradient less than tolerance. |
But in order to solve it again, all .l variables should be re-initialized to their initial (e.g. pre-solve) values, otherwise it would use the values obtained in the first solve as a starting values for the second solve, isn't it?
ReplyDeleteNo, purpose is to do a restart with optimal levels of first solve. Not just repeat the first solve (that would give exactly the same results). You can see the second solve only takes 202 iterations compared to 827 for the first.
Delete