Gradient/sensitivity Estimation in Discrete-Event Simulation

We consider methods for simultaneously estimating the gradient (resp. sensitivity) of J with respect to a continuous (resp. discrete) parameter /spl theta/. While one can always estimate a derivative (or sensitivity) by performing two distinct simulations at different values of /spl theta/ (and forming the estimate (J/spl theta/', - J/spl theta/2)/(/spl theta/' - /spl theta/)), here we focus on methods which compute estimates from a single sample path (though multiple replications can be used to reduce variance and compute confidence intervals). When applicable, these single run methods are often computationally more efficient. They also have applications beyond simulation (e.g. in on-line optimization and control).