O. Epelly, J. Gondzio and J.-P. Vial
Abstract
Solving large-scale optimization economic models such
as Markal-Macro models proves to be difficult or even out
of reach for state-of-the-art solvers.
We propose an optimizer which takes advantage of their possible
special structure: a large dynamic linear block on one side,
a small nonlinear convex block on the other one. This framework
favors the use of interior point methods which are efficient for
large-scale linear programs and which can handle convex programs.
NLPHOPDM is an implementation of an interior point method built upon
the HOPDM
code for linear and convex quadratic optimization
The method combines ideas of a globally convergent algorithm
and the extension of multiple centrality correctors technique
to nonlinear convex programming. It is designed for being hooked
to modeling languages such as GAMS and AMPL. We present in this
paper preliminary results relative to our research code NLPHOPDM
and to commercial nonlinear solvers. Our approach achieves
a significant computational speed-up. This is performed via
the use of a library which computes exact second derivatives.
Key words: Interior Point Method, Economic Model, Smooth Convex Optimization.