The stability-capacity diagram of a fully connected neural network, whose bonds take the values 1 and -1 with equal probability, is determined numerically. Two different optimization methods (simulated annealing and tabu search) are used and their relevant features are discussed. The results indicate the existence of a region, in the stability-capacaty plane, where replica-symmetry should be broken, within the replica-symmetric phase found by previous, analytical, computations. The critical capacity, consistent with simple arguments on information storage, is found to be less than one Le diagramme stabilite-capacite d'un reseau de neurones totalement connecte, dont les liens synaptiques ne peuvent prendre que deux valeurs 1 ou -1, avec la meme probabilite, est determine par simulations numeriques. Deux differentes methodes d'optimisation ont ete utilisees (recuit simule et recherche tabou) et nous discutons leurs aspects les plus importants. Les resultats indiquent l'existence d'une region, dans le plan stabilite-capacite, ou la symetrie des repliques devrait etre brisee, a l'interieur de la phase symetrique trouvee precedemment par des calculs analytiques. La capacite critique, en accord avec des arguments simples, est trouvee inferieure a un
[1]
E. Gardner.
The space of interactions in neural network models
,
1988
.
[2]
W. Krauth,et al.
Learning algorithms with optimal stability in neural networks
,
1987
.
[3]
Sompolinsky,et al.
Information storage in neural networks with low levels of activity.
,
1987,
Physical review. A, General physics.
[4]
E. Gardner,et al.
Optimal storage properties of neural network models
,
1988
.
[5]
Fred Glover,et al.
Interactive decision software and computer graphics for architectural and space planning
,
1985
.
[6]
C. D. Gelatt,et al.
Optimization by Simulated Annealing
,
1983,
Science.
[7]
M. Mézard,et al.
Spin Glass Theory and Beyond
,
1987
.