No label defined (Q4676006)

This paper proposes a new quasi-Newton approach for the unconstrained optimization problem of a quadratic function in a finite number of iterations. An approximating matrix of the Hessian function is used and updated throughout the iteration to incorporate the value changes in the function and its gradient. The exposed method is compared and shown to be similar to the algorithm of \textit{L. C. W. Dixon} [J. Inst. Math. Appl. 11, 317--328 (1973; Zbl 0259.65060); ibid. 16, 9--18 (1975; Zbl 0294.90076)], providing significant improvements in accuracy and computational storage savings. The advantages of the new technique over the standard ones based on the Broyden-Fletcher-Goldfarb-Shanno formula, applied to general functions, are detected especially when the new algorithm moves closer to the minimum, where near quadratic behaviour is displayed. The algorithm has been implemented in C++ (run on a PIP 200 processor using double precision), and tested on 17 standard (benchmark) functions. The numerical results indicate a clear advantage of the new method, especially when applied on large dimensions and non-quadratic functions.