A probe about the unified construction of gradient projection type methods (Q1210243)

The problem of minimizing a continuously differentiable objective function \(f(x)\) under inequality constraints \(g_ j(x)\leq 0\) \((j=1,\dots,m)\), where \(x\in E^ n\) and the \(g_ j\)'s are continuously differentiable is considered. It is shown that the iterative directions of many well-known gradient projection type methods at an arbitrary point can be obtained through solving a linear system constructed on the basis of the original optimization problem. Further, some new gradient projection type algorithms can be derived using this linear system.