Stochastic approximation algorithms: Nonasymptotic estimation of their convergence rates (Q1281080)

Stochastic approximation algorithms are well--known from optimization theory. However, to apply them properly, the convergence properties must be investigated first. It is known (from the literature) that the technique of Lyapunov functions is here very often employed. However, to obtain ``good'' convergence results (by Lyapunov approaches) some rather strong conditions are usually assumed. The aim of this paper is to investigate the convergence rate without these assumptions. To obtain more general convergence results the authors replace the investigation of mathematical expectations (of the corresponding Lyapunov functions) by direct investigation in the points of parameter estimates. By this approach they try to omit an ``unpleasant'' assumption on concavity. Moreover, they extend a class of algorithms that can be properly applied. At the end of the paper the authors mention also the fact that the new results can be successfully employed in the case of adaptive dynamic systems.