The Euclidean Steiner Tree Problem in dimension greater than 2 is notoriously difficult. Successful methods for exact solution are not based on mathematical-optimization - rather, they involve very sophisticated enumeration. There are two types of mathematical-optimization formulations in the literature, and it is an understatement to say that neither scales well enough to be useful. We focus on a known nonconvex MINLP formulation. Our goal is to make some first steps in improving the formulation so that large instances may eventually be amenable to solution by a spatial branch-and-bound algorithm. Along the way, we developed a new feature which we incorporated into the global-optimization solver SCIP and made accessible via the modeling language AMPL, for handling piecewise-smooth univariate functions that are globally concave.
展开▼