Mathematics and Computer Science Division (MCS)
Argonne National Laboratory
Argonne, IL 60439
To induce global convergence and to unify the approaches used in LANCELOT and MINOS, we work with an elastic form of the linearized constraints (by adding an L1 penalty term to the augmented Lagrangian). Our stabilized LCL method possesses two important properties: the linearly constrained subproblems are always feasible, and they may be solved inexactly.
The current implementation is efficient on large problems, using MINOS to solve the subproblems. Only first derivatives are required. We present numerical results on the COPS and CUTE test problems.