A nonmonotone globalization algorithm with preconditioned gradient path for unconstrained optimization (Q618130)

Nonmonotone technique in optimization is useful, because, at each iteration, the method with enforcing monotonicity of the objective function values may cause a series of very small steps if the contours of the objective function are a family of curves with deep valleys. In this paper a nonmonotone stabilization algorithm with the preconditioned gradient path for unconstrained optimization is presented. The global convergence and locally superlinear convergence are established for this class of algorithms. Numerical results show that the strategy proposed in this paper allows considerable computational savings, especially for solution of ``difficult'' test problems.