A study of hybrid conjugate gradient methods
The main subject of the research in this thesis is the study of conjugate gradient methods for optimization and the development of improved algorithms. After an introductory first chapter, Chapter 2 contains a background of numerical methods for optimization in general and of conjugate gradient-type algorithms in particular. In Chapter 3 we study the convergence properties of conjugate gradient methods and discuss Powell's (1983) counter example that proves that there exist twice continuously differentiable functions with bounded level sets for which the Polak-Ribiere method fails to achieve global convergence whereas the Fletcher-Reeves method is shown to be globally convergent, despite the fact that in numerical computations the Polak-Ribiere method is far more efficient than that of Fletcher-Reeves. Chapters 4 and 5 deal with the development of a number of new hybrid algorithms, three of which are shown to satisfy the descent property at every iteration and achieve global convergence regardless of whether exact or inexact line searches are used. A new restarting procedure for conjugate gradient methods is also given that ensures a descent property to hold and global convergence for any conjugate gradient method using a non negative update. The application of these hybrid algorithms and that of the new restarting procedure to a wide class of well-known test problems is given and discussed in the final Chapter "Discussions and Conclusions". The results obtained, given in the appendices, show that a considerable improvement is achieved by these hybrids and by methods using the new restarting procedure over the existing conjugate gradient methods and also over quasi-Newton methods.