Control of Initialization Bias in Queueing Simulations Using QueueingApproximations

摘要

The most widely used methodology of initialization bias control is that of datatruncation. The idea is that the stochastic nature of the random variables will ultimately produce observations which are more representative of the steady-state of the system after a period of transience. The initial transient observations which bias an estimate are not representative of the steady state characteristics and can be deleted or thrown away. From that point in the output sequence to the end, the estimate will be less biased than taking the entire output sequence average. While these control methodologies are commonly known as heuristics, they do serve a purpose to provide the decision maker with the best possible information by doing the best they can with an output sequence. This research focuses on producing a good estimate from sequentially correlated simulation output data. I evaluate the use of proven accurate queueing approximations to stochastically set the initial queue length from the approximated steady state distribution to derive a better estimate than empty and idle, without the use of pilot runs. I also evaluate how point approximations can assist in controlling the bias in output estimates of a desired performance parameter through four truncation heuristics. The end result is a less biased and more accurate