Partenaires

CNRS
Logo tutelle
Logo tutelle
Logo tutelle

Rechercher

Sur ce site

Sur le Web du CNRS

Accueil du site > Séminaires > Probabilités Statistiques et réseaux de neurones > Recurrent Artificial Neural Networks for Optimization

Vendredi 21 juin 2002 à 11h00

Recurrent Artificial Neural Networks for Optimization

Gonzalo Joya (Université de Malaga)

Résumé : Recurrent artificial neural networks (RANNs) are fundamentally defined by its dynamics - expressed with a system of ordinary differential equations- and an associated energy function. The existence of this energy function allows this paradigm for the application to optimization problems, which are relevant from both a theoretical and practical perspective. From a theoretical point of view, because optimization problems are frequently NP-complete, thus providing a good benchmark for comparison with other optimization methods. From a practical point of view, because these problems frequently describe real problems, which are not efficiently solved by classical techniques. Moreover, other interesting problem classes such as control and parameter estimation can be described in terms of optimization. The most important limitations of RANNs regarding this field are two : existence of local minima and slow convergence. Besides, the association between the diverse dynamical equations and the corresponding energy functions is often carried out with insufficient rigor. Several methods have been proposed to face the problem of local minima : on the one hand, strategies for both local minima avoidance and global minimum search have been established, mainly based on the variation of the neuron gain parameters. On the other hand, new energy functions that possess only one global minimum have been explored, resulting in new conditions on the network weights. The slow convergence problem may be approached by either a parallel implementation or searching new numerical methods for solving the system of differential equations that describe the network dynamics. In this course, a review of the previous questions is carried out. Thus, we describe the process of obtaining the network structure that solves each particular optimization problem, and we analyze each of the above mentioned applicability limitations and some of the proposed solutions.

Dans la même rubrique :