Several extensions reduce the variance of the gradient estimate or adapt the step size per parameter: