A Non-Rectangular Sampling Plan for Estimating Steady-State Means

P. W. Glynn

Proceedings of the 6th Army Conference on Applied Mathematics and Computing, 965-978 (1988)

The method of multiple replicates is frequently used by simulators to estimate the steady-state mean of a stochastic simulation. One important advantage of this approach is that it is easily adapted to a parallel computer. Unfortunately, the method of multiple replicates is quite sensitive to contamination by initial bias. In this paper, a new type of sampling plan is described. It retains the replication flavor, yet attenuates the bias problem. It is shown that the new method reduces mean square error relative to conventional multiple replicates for problems in which the "initial transient" decays slowly.