Some guidelines for using iterate averaging in stochastic approximation

Averaging of the output (iterates) from a stochastic approximation (SA) recursion has been shown to be a useful technique for the gradient-based Robbins-Monro, form of SA. For the gradient-free (e.g., Kiefer-Wolfowitz) form, iterate averaging can produce an improvement in the stability of the algorithm and competitive mean-square errors relative to the standard (unaveraged) recursion. We discuss guidelines on how and when to use averaging in this context.