We present a Modeler-Optimizer Interface (MOI) for general closed form Non-Linear Programs (NLP), which can be used to to transfer NLPs in a clear and simple manner between optimization components in a distributed environment. We demonstrate how this interface allows first order derivative information to be easily calculated on the optimizer's side, using automatic differentiation, hence removing the bottleneck of communicating derivative information between the modeler and the optimizer.
We also show how this interface directly corresponds to a file format for NLPs, the extended MPS format (xMPS). This format directly extends the standard MPS file format for linear and mixed integer programs to include NLPs and permits a standardized way of transferring benchmark problems. The format spares the modeler the tedious task of calculating derivative information with minimal extra work required by the optimizer and thus increases efficiency. This work was originally done at Maximal Software in order to connect the MPL modeling language with non-linear solvers.
Bjarni V. Halldorsson, Dept. of Mathematical Sciences, Carnegie Mellon Univ.
Bjarni Kristjansson, Maximal Software, Inc.
Erlendur S. Thorsteinsson, Graduate School of Industrial Admin., Carnegie Mellon Univ.
You can view the rest of the article in: