×

Do stochastic algorithms circumvent traps? (Les algorithmes stochastiques contournent-ils les pièges?) (French) Zbl 0841.68048

Summary: We are considering the validity of commonly held claims: “a stochastic gradient algorithm only converges towards one of local minima” or “a stochastic algorithm only converges towards an asymptotically stable solution of the associated differential equation”. We prove that a stochastic algorithm does not fall into a regular trap if the noise is exciting in a repulsive direction.

MSC:

68W10 Parallel algorithms in computer science
PDFBibTeX XMLCite