Computational experience with a safeguarded barrier algorithm for sparse nonlinear programming

Abstract. We describe an enhanced version of the primal-dual interior point algorithm in Lasdon, Plummer, and Yu (ORSA Journal on Computing, vol. 7, no. 3, pp. 321–332, 1995), designed to improve convergence with minimal loss of efficiency, and designed to solve large sparse nonlinear problems which may not be convex. New features include (a) a backtracking linesearch using an L1 exact penalty function, (b) ensuring that search directions are downhill for this function by increasing Lagrangian Hessian diagonal elements when necessary, (c) a quasi-Newton option, where the Lagrangian Hessian is replaced by a positive definite approximation (d) inexact solution of each barrier subproblem, in order to approach the central trajectory as the barrier parameter approaches zero, and (e) solution of the symmetric indefinite linear Newton equations using a multifrontal sparse Gaussian elimination procedure, as implemented in the MA47 subroutine from the Harwell Library (Rutherford Appleton Laboratory Report RAL-95-001, Oxfordshire, UK, Jan. 1995). Second derivatives of all problem functions are required when the true Hessian option is used. A Fortran implementation is briefly described. Computational results are presented for 34 smaller models coded in Fortran, where first and second derivatives are approximated by differencing, and for 89 larger GAMS models, where analytic first derivatives are available and finite differencing is used for second partials. The GAMS results are, to our knowledge, the first to show the performance of this promising class of algorithms on large sparse NLP’s. For both small and large problems, both true Hessian an