Experience with a primal presolve algorithm.pdf

A TT Bell Laboratories Numerical Analysis Manuscript 93–06 Experience with a Primal Presolve Algorithm Robert Fourer ? David M. Gay ? April 23, 1993 ? Dept. of Industrial Engineering and Management Sciences, Northwestern Uni- versity, Evanston, IL 60201-3119, U.S.A.; 4er@ . ? ATT Bell Laboratories, Room 2C-463, 600 Mountain Avenue, Murray Hill, New Jersey 07974-0636, U.S.A.; dmg@ . Experience with a Primal Presolve Algorithm Robert Fourer Dept. of Industrial Engineering and Management Sciences Northwestern University Evanston, IL 60201 USA David M. Gay ATT Bell Laboratories Murray Hill, New Jersey 07974 ABSTRACT Sometimes an optimization problem can be simplified to a form that is faster to solve. Indeed, sometimes it is convenient to state a problem in a way that admits some obvious simplifications, such as eliminating fixed variables and removing constraints that become redundant after simple bounds on the variables have been updated appropriately. Because of this convenience, the AMPL modeling system includes a ‘‘presolver’’ that attempts to simplify a problem before passing it to a solver. The current AMPL presolver carries out all the primal simplifications described by Brearely et al. in 1975. This paper describes AMPL’s presolver, discusses reconstruction of dual values for eliminated constraints, and presents some computational results. Introduction Consider the constrained optimization problem find x ∈ I R n to minimize f (x) (1) subject to b ≤ g(x) ≤ d (2) and ≤ x ≤ u (3) in which I R k is the set of vectors having k real components and g: I R n → I R m . This paper is about simpli- fying the constraints (2) and (3), primarily when the general constraints (2) are linear, i.e., they have the form b ≤ Ax ≤ d (4) for some matrix A ∈ I R m×n . Sometimes the general constraints (2) may imply bounds (3) on the variables or may imply the only values that certain variables can assume. This may let us remove some variables and constraints, a process sometimes calle