On stochastic mirror descent with interacting particles: convergence properties and variance reduction (Q2077867)

Optimization models often include noisy estimates of the function and it is beneficial to control the fluctuations around the true minimum. The conventional way to control the error is to decrease the step size. Another option is to run independent replicas of the algorithm. Each of these runs is referred to as a particle. The question addressed in this paper is whether it is beneficial to allow these particles to interact with each other and study this question using the general framework of stochastic mirror descent (SMD). The main focus is on the convergence properties of the stochastic mirror descent with interacting particles (ISMD), on demonstrating the performance of ISMD in practice and on demonstrating the effect of interaction on decreasing the variance in a variety of examples. Theoretical and numerical evidence is provided to suggest that the interaction helps to improve the convergence and to reduce the variance of the estimate.