 |
Center for Nonlinear Studies |
There is a gap between what we want to solve and what we can solve. In other words, the model that the optimization user requires to be solved does not always match the model that compatible solvers were designed to solve. Indeed, the user typically want the represent in the model as most structure about its problem as possible while the solver is usually dedicated to a class of optimization problem and only support the structure that it actually exploits.
Implementing the support for every possible structure a solver is compatible with in a solver-specific manner is unscalable in addition to preventing solvers from supporting structures that it does not know about. MathOptInterface bridges this gap in a solver-independent way that fully supports a decentralized development of the modeling framework. Problem reformulations can be defined independently and appropriate transformations are automatically selected using an algorithm derived from Bellman-Ford.
Host: Tillmann Weisser