Minimizing initialization bias in simulation output using a simple heuristic

We present a rule for determining the number of observations to delete from the beginning of an output sequence generated by a steady-state, discrete-event simulation. This rule is easy to implement and has strong intuitive appeal. Given a finite sequence with an arbitrary initial condition, the rule is also optimal, in the sense that it minimizes the width of the marginal confidence interval about the truncated sample mean. We illustrate the performance of the rule by applying it to output generated by multiple runs of four queueing simulations. Models and run conditions provide the test sequences for which truncation is contraindicated, test sequences exhibiting positive and negative initialization bias, and test sequences which are entirely transient. Results demonstrate that the rule is efficient, effective, and consistent with the conflicting objectives of mitigating initialization bias while preserving information