DO STOCHASTIC ALGORITHMS FALL INTO TRAPS

Authors
BRANDIERE O DUFLO M
Citation
O. Brandiere et M. Duflo, DO STOCHASTIC ALGORITHMS FALL INTO TRAPS, Annales de l'I.H.P. Probabilites et statistiques, 32(3), 1996, pp. 395-427
Citations number
34
Categorie Soggetti
Statistic & Probability","Statistic & Probability
Journal title
Annales de l'I.H.P. Probabilites et statistiques → ACNP
ISSN journal
02460203
Volume
32
Issue
3
Year of publication
1996
Pages
395 - 427
Database
ISI
SICI code
0246-0203(1996)32:3<395:DSAFIT>2.0.ZU;2-4
Abstract
Do stochastic algorithms fall into traps? We are considering the valid ity of commonly held claims: ''a stochastic gradient algorithm only co nverges towards one of the local minima'' or ''a stochastic algorithm only converges towards an asymptotically stable solution of the associ ated differential equation''. We prove that a stochastic algorithm doe s not fall into a regular trap if the noise is exciting in a repulsive direction.