Simulation warm-up and initialization bias

Suppose a hospital is normally busy, but the simulation starts with no patients and every bed free. Early waiting times will be unusually low because the model has not yet reached its normal workload.

That early period is called the initial transient. A warm-up period removes it from the measurements while leaving the model itself running normally.

Do not add warm-up automatically. A terminating model may need to start empty because the empty state is part of the real problem, such as a shop opening for the day.

Use warm-up when you want long-run behavior and the chosen starting state is artificial.

Run several long replications and plot the measure you care about over time. Welch's graphical method averages the replications and smooths the result so the initial drift is easier to see.

Pick a cutoff after the initial trend has settled, then check that a nearby cutoff does not change the conclusion.

The small example below shows the same run summarized with different warm-up cutoffs. The model state is unchanged; only the part used for measurement changes.