Using Global Line Searches for Nding Global Minima of Mlp Error Functions

In this paper we consider a possible improvment of conjugate gradient methods commonly used for training multilayer perceptrons. These methods rely on a line search procedure which is partially responsible for the fact that only local minima are found. We propose to perform line searches along very long lines in order to nd global minima; as usual line search procedures are not suitable in this context, we propose to use powerful algorithms from the eld of global optimization, and in particular Zilinskas's P algorithm. We show that P can not replace the usual line search directly, but can be coupled with a conjugate gradient algorithm in a diierent way. The new algorithm we propose outperforms several related algorithms.