An analytic center cutting plane method in conic programming

Vasile (Luc) Basescu

PhD Thesis, August 2003

Mathematical Sciences

Rensselaer Polytechnic Institute

Advisor: John Mitchell

Download the thesis, in pdf


Conic programming has been lately one of the most dynamic area of the optimization field. Although a lot of attention was focused on designing and analyzing interior-point algorithms for solving optimization problems, the class of analytic center cutting plane methods was less investigated. These methods are designed to solve feasibility problems by finding points which are interior to different sets of interest. Although these methods can be used by themselves to solve optimization problems, most of the time they are used as an initial step in a larger interior-point scheme employed in solving optimization problems.

There are many advantages in using this class of algorithms. For these methods to work there is no need to have before hand a complete description of the set of interest. All we need is an oracle that describes the set. This feature is especially useful when such a description is either missing or it is too large to be practical.

In this thesis we present a general analytic center cutting plane method for solving feasibility problems in the context of conic programming. The set of interest is convex, bounded, fully dimensional. It is described by an oracle that either recognizes that a point is interior to the set or returns a set of constraints violated by the current point but verified by all the points of the set of interest. These violated constraints are also known as cuts.

Our approach is an extension to the analytic center methods used in linear programming, second order cone programming or semidefinite programming. We prove that our algorithm can solve any feasibility problem with a convex, bounded, fully dimensional set of interest. We derive an upper bound for the total number of iterations the algorithm requires to get the solution. Also, we analyze how expensive each iteration is.

The performance of the algorithm is analyzed by solving some feasibility problems derived from the set of problems proposed in ``The Seventh DIMACS Implementation Challenge Semidefinite and Related Optimization Problems". We consider feasibility problems with the sets described only by linear and second order conic constraints. We will also present an algorithm for solving optimization problems that incorporates our analytic center cutting plane method. In the last part of this thesis we analyze the linear programming version of this algorithm and prove that it converges. Complexity results are also presented.

Library listing.

Download the thesis, in pdf.