A tolerant algorithm for linearly constrained optimization calculations (Q910337)

An algorithm is presented for solving smooth nonlinear programming problems with linear constraints. The main difference to existing algorithms is the definition of active constraints, which is replaced by so-called `small residual' constraints. The intention is to prevent small steps forced by nearly active constraints when some of them are degenerate. Moreover the BFGS-formula is used to update an approximation of the Hessian matrix of the objective function. Convergence is ensured by performing a line search. Critical computational details of the algorithm are outlined and global convergence is proved, i.e., convergence from an arbitrary starting point towards a Kuhn-Tucker point. The software was tested on six examples, the results are analyzed in detail.