© 1969 by British Computer Society
| ||||||||||||||||||||||||||||||||||||||||||||||||||
Variable metric methods of minimisation
Research Analysis Corporation, McLean, Virginia, USA
Two basic approaches to the generation of conjugate directions are considered for the problem of unconstrained minimisation of quadratic functions. The first approach results in a projected gradient algorithm which gives n step convergence for a quadratic. The second approach is based on the generalised solution of a set of underdetermined linear equations, various forms of which generate various new algorithms also giving n step convergence. One of them is the Fletcher and Powell modification of Davidon's method.
Results of an extensive numerical comparison of these methods with the NewtonRaphson method, the FletcherReeves method, and the FletcherPowellDavidon method are included, the test functions being non-quadratic.
Received October 1968.