There are several methods for assessing the accuracy of a solution:
•Use knowledge of the problem
•Vary one or more variables and observe the behavior of the objective and constraint functions
•Try different starting points
•Observe initial and final magnitudes of the reduced gradients of the superbasic and nonbasic variables
With respect to the second method, some of the nonbasic variables can be fixed at their current values (set upper bound = lower bound = current value for the affected parameterUnknowns) and others can be varied using Embed. If this does not produce a significantly improved feasible point, confidence in the current solution is increased. Confidence is also increased if different starting points lead to nearly the same final point.
With respect to the fourth method, the best indication of accuracy occurs if the Kuhn-Tucker conditions are satisfied (inform = 0). However, if they are not, but the reduced gradient components of superbasic variables have been reduced from their initial values, by say three orders of magnitude or more, while reduced gradients of nonbasic variables at bound have the correct sign, this is symptomatic of reasonably high accuracy.
If perceived accuracy is not sufficient, reducing epstop and/or increasing nstop often helps. It is also often effective to reduce epnewt if epstop is reduced, since this increases the accuracy of the reduced gradient computation.