Recursive Stochastic Algorithms for Global Optimization in $\mathbb{R}^d $ (Q3978271)

This paper deals with a class of algorithms for finding a global minimum of a smooth function \(U(x)\), \(x\in \mathbb{R}^ d\). More precisely the authors analyze the convergence of a modified stochastic gradient algorithm: \(X_{k+1}=X_ k-a_ k(\nabla U(X_ k)+\xi_ k)+b_ kW_ k\), where \(\xi_ k\) is a sequence of \(\mathbb{R}^ d\) valued random variables, \(W_ k\) is a sequence of standard \(d\)-dimensional independent Gaussian random variables, and \(a_ k\), \(b_ k\) are sequences of positive numbers with \(a_ k\), \(b_ k\to 0\). This algorithm is obtained by adding the \(b_ kW_ k\) term to a standard stochastic gradient algorithm. The ideal behind the additional \(b_ kW_ k\) term is that if \(b_ k\) tends to zero slowly enough then possibly \(X_ k\) will avoid getting trapped in a strictly local minimum of \(U\). Conditions and relations on the different parameters are given for ensuring convergence in probability towards a global minimum. No evaluation of the convergence speed is given but the paper is interesting.