Terminating vs steady-state simulation
Decide whether each run represents a complete episode or long-run operation before choosing warm-up, run length, and output analysis.
Ask whether each replication has a meaningful real-world end. A shop day, evacuation, project, campaign, or outage episode can end naturally. A continuously operating factory, network, or service system may instead be studied for its long-run behavior.
That distinction changes what the start of a run means, whether warm-up is appropriate, and what one replication represents.
In a terminating study, the starting state and ending condition are part of the system you want to study. If a clinic opens empty each morning, the empty state is not initialization bias; it is part of the day.
Treat each complete episode as one replication. Compare outputs such as total completions, overtime, peak queue, or mean waiting across independently repeated episodes.
For a system that is conceptually continuous, the initial state may be arbitrary. Starting a busy service network empty can make the first part of the run unrepresentatively quiet.
Use a defensible initial state or discard an initial warm-up period, then collect output long enough to estimate the long-run measure you care about. Keep the warm-up rule and collection period separate in the study record.
A factory may stop production overnight while work in process remains on the floor. A hospital has daily staffing schedules but patients can stay across midnight. In both cases, the system state carries across the calendar boundary.
Classify the study from the state you are measuring, not from whether the clock happens to show days or shifts.
If you cannot say what one replication represents, stop before interpreting confidence intervals or comparing alternatives. The statistical analysis only makes sense after the study boundary is clear.
- Terminating: one replication is normally one complete episode with its natural start and finish.
- Steady-state: separate initialization from the measurement period and account for serial dependence within a long run.
- Mixed or cyclical systems: state explicitly what resets between cycles and what carries over.