Consistent, asymptotically efficient and asymptotically normal stepwise estimators are given for a subclass of the uniparametric and multiparametric exponential families. Estimators are derived from the Robbins-Monro stochastic approximation procedure with certain families of random variables arising from the normalized log-likehood. Three multivariate normal examples where the maximum likelihood estimators are not tractable are considered. (Author/RC)