Affordable Access

Stochastic Optimization with Unadjusted Kernel: the SOUK Algorithm

Authors
  • De Bortoli, Valentin
  • Durmus, Alain
  • Pereyra, Marcelo
  • Fernandez Vidal, Ana
Publication Date
Jan 12, 2019
Source
Acervo Digital da Unesp
Keywords
Language
English
License
Unknown
External links

Abstract

In this paper we propose a stochastic approximation algorithm to minimize functions for which the gradient writes as the expectation of some integrable features. Considering the case where computing feature moments or sampling from the underlying probability distribution is not feasible we mix a Markov chain dynamic with the stochastic gradient descent dynamic, thus combining optimization tools with statistics controls. This approach is motivated with examples from Empirical Bayesian statistics. We assess the convergence as well as rates of convergence for various objective functions.

Report this publication

Statistics

Seen <100 times