On fraudulent stochastic algorithms

Abstract

We introduce and analyse the almost sure convergence of a stochastic algorithm for the global minimisation of smooth functions. This diffusion process is called fraudulent because it requires the knowledge of minimal value of the function. Nevertheless, its investigation is not without interest, since in particular it appears as the limit behaviour of non-fraudulent and time-inhomogeneous swarm mean-field algorithms for global optimisation or in stochastic gradient descent algorithms in over-parametrised deep learning applications. The talk is based on collaborations with Benaïm, Bolte and Villeneuve.