A new quasi-Newton algorithm (Q1095608): Difference between revisions

The authors give a new algorithm which belongs to the one parameter Broyden's family for unconstrained optimization problems min f(x), \(x\in R^ n\). This algorithm uses, unlike other methods, the negative values for the parameter of the Hessian updating formula. A global convergence theorem and a superlinear convergence theorem, when it is applied to convex functions, are given. In their algorithm the authors not only use the negative value for the parameter, but they vary the parameter such that the corresponding descent search direction spans a wedge. The algorithm chooses in the set of the descent quasi-Newton directions (the wedge) the one in which the objective function locally descends most rapidly. As the authors remark, for test problems the new algorithm consistently outperforms BFGS in terms of the iteration number and the effective function evaluation number.