Algorithms with conic termination for nonlinear optimization. (English) Zbl 0663.65062
The authors describe the implementation of algorithms for unconstrained optimization which have the property of minimizing conic objective functions in a finite number of steps, when line searches are exact. The basic properties of the conic function \(f(x)=f(x_ 0+s)=f_ 0+(1-a^ Ts)^{-1}g_ 0^ Ts+2^{-1}(1-a^ Ts)^{-2}s^ TAs\) are described, where \(g_ 0\), \(a\in {\mathbb{R}}^ n\) and A is a positive definite and symmetric \(n\times n\) matrix. The new presented algorithms are designed to minimize conic functions in n steps.
Therefore the authors begin by reviewing the three algorithms, 1) the conjugate gradient method, 2) the BFGS quasi-Newton method and 3) the limited BFGS method. These methods are translated into algorithms for the minimization of the conic function f, and a transition from conic objective functions to general nonlinear functions is given. Finally a general algorithm is discussed and numerical results obtained with the preferred implementation of the three new classes of algorithms are presented.
Therefore the authors begin by reviewing the three algorithms, 1) the conjugate gradient method, 2) the BFGS quasi-Newton method and 3) the limited BFGS method. These methods are translated into algorithms for the minimization of the conic function f, and a transition from conic objective functions to general nonlinear functions is given. Finally a general algorithm is discussed and numerical results obtained with the preferred implementation of the three new classes of algorithms are presented.
Reviewer: H.Benker