The LANCELOT package uses an augmented Lagrangian approach to handle all constraints other than simple bounds. The bounds are dealt with explicitly at the level of an outer-iteration subproblem, where a bound-constrained nonlinear optimization problem is approximately solved at each iteration.

The algorithm for solving the bounded problem combines a trust region approach adapted to handle the bound constraints, projected gradient techniques, and special data structures to exploit the (group partially separable) structure of the underlying problem.

The software additionally provides direct and iterative linear solvers (for Newton equations), a variety of preconditioning and scaling algorithms for more difficult problems, quasi-Newton and Newton methods, provision for analytical and finite-difference gradients, and an automatic decoder capable of reading problems expressed in Standard Input Format (SIF).

LANCELOT A is written is standard ANSI Fortran 77. Single- and double-precision versions are available. Machine dependencies are isolated and easily adaptable. Automatic installation procedures are available for DEC VMS, DEC ULTRIX, Sun UNIX, Cray UNICOS, IBM VM/CMS, and IBM AIX.