一个求解等式约束最优化的全局收敛的不精确SQP算法.pdfVIP

 SQP Æ SQP ℓ2 − ℄ ℄ ℄ KKT Fritz-John ℄ SQP i SQP A GLOBALLY CONVERGENT INEXACT SQP METHOD FOR EQUALITY CONSTRAINED OPTIMIZATION ABSTRACT We present a globally convergent inexact SQP algorithm for equality con- strained optimization. The method uses an ℓ2 − exact penalty function as the merit function. At every iteration, the algorithm generates new iter- ates by line search procedure and solves modified QP subproblem inexactly by conjugate gradient method. Under suitable conditions on inexact solu- tions, if constraint Jacobian has full row rank at a limit point of sequence of iterates, then the limit point is a KKT point of equality constrained op- timization. If constraint Jacobian does not have full row rank at the limit point and if the limit point is feasible, then the limit point is a Fritz–John point of the problem; Otherwise the limit point is infeasible and is an in- feasible stationary point of the problem. Numerical results show that the algorithm can cope with the situation that the gradients of constraints are linearly dependent at some iterates. KEY WORDS: equality constrained optimization, SQP method, inexact solution, line search method, global convergence ii SQP F g (x) ℄ A(x) c(x) ℄ . Euclidean ℄ φ( ; ) ℄ Dφ( ; )