Originally posted by: Christian Bliek
Hello,
I'm not exactly sure what you mean by well centered dual values and how sub optimality could help you get that.
I suspect that you do not want a simplex type solution where some dual values are zero and others are not, and that instead you are looking for solutions which are centered. For example solutions in the middle of a facet rather than at a vertex, so that the dual values are not zero if they don't have to be. If you run barrier without crossover you will find this type of solutions. Barrier attempts to follow the central path, and hence has a tendency to find centered solutions.
Make sure that you turn off crossover, otherwise cplex will continue, and use the barrier solution to find a vertex solution. The convergence tolerance controls when the barrier algortihm stops. Increasing it will give you less accurate solutions; primal and dual errors will be larger as well as complementarity. I'm not sure that this is what you want and I don't think this will affect centrality much.
The only thing that comes to mind is to increase the number of corrections you allow. By default corrections are used for overall solution speed. If they are too expensive cplex may do only very few. However, if you wish to closely follow the central path, you may want to do more. This way the final solution may be more centered.
Christian.
#CPLEXOptimizers#DecisionOptimization