CONTROL VARIATE SELECTION FOR MULTIRESPONSE SIMULATION

A solution is offered to the general problem of optimal selection of control variates. Solutions are offered for two different cases of the general problem: (a) when the covariance matrix of the controls is unknown, and (b) when the covariance matrix of the controls is known and is incorporated into point and confidence region estimators. For the second case a new estimator is introduced. Under the assumption that the responses and the controls are jointly normal, the unbiasness of this new estimator is established, and its dispersion matrix is derived. A selection algorithm is implemented which locates the optimal subset of controls. The algorithm is based on criteria derived for the two cases listed above. A promising new class of controls is introduced which are called routing variables. The asymptotic distribution of these controls is derived as well as their asymptotic mean and variance. Finally, the performance of the selection algorithm is investigated and the new estimator is contrasted with the classical estimator