EE8405 Chap.10 Simulation, Correlation and Model Checking
Simulation, Correlation and Model Checking
Discrete-event simulation advances from one arrival or departure to the next and records both time-weighted states and customer delays. The chapter controls empty-start bias with warm-up, quantifies precision through independent replications and confidence intervals, and validates output against analytical cases. Correlation is treated as a limited linear measure rather than proof of independence or causation.
What this chapter covers
- 01
Event calendars
- 02
State updates
- 03
Warm-up bias
- 04
Independent replications
- 05
Confidence intervals
- 06
Correlation and validation
Worked example · free
Put precision around replication means
- 2Calculate the replication mean.
- 3Use sample variance with the n minus one divisor.
- 3Convert to standard error and choose a t critical value.
Key terms
- Warm-up Period
- The initial simulated interval excluded to reduce bias from the starting state.
- Standard Error
- The estimated standard deviation of a sample-mean estimator.
Simulation, Correlation and Model Checking FAQ
Why discard a warm-up period?
A queue started empty does not initially represent steady state. Warm-up removal reduces the bias that this artificial initial condition creates in long-run estimates. Justify the cutoff from diagnostics.
Why use independent replications?
They provide transparent observations for sample variance and confidence intervals, expose seed sensitivity and make the precision calculation reproducible. Each replication should use a distinct random stream and identical observation rules.
Does zero correlation prove independence?
No. It rules out measured linear association but nonlinear dependence may remain. Independence is a stronger property of the full joint distribution. Inspect mechanisms and lag structure before concluding more.
What belongs in a reproducible simulation record?
Record input parameters and units, the seed or random-stream policy, event-order rules, warm-up, observation horizon, replication count, estimator definitions and the version of the model logic. Include the achieved confidence half-width rather than only a point estimate.
A saved chart is insufficient because it cannot show which population was averaged, how transient data was handled or whether another student can recreate the same event sequence.
How can an analytical queue validate a simulation?
Configure the model as a simple case with a known result, such as a stable M/M/1 queue, and compare utilization, mean number and delay within simulation uncertainty. Also trace early events by hand and test conservation and impossible-state invariants. Agreement at several levels is stronger than matching one mean.
Only after those checks should the same code be trusted for a richer routing, finite-capacity or correlated-traffic scenario.
Why report a confidence half-width with the estimate?
The half-width shows attainable precision and whether differences between designs exceed simulation noise. A point estimate alone cannot support that comparison or guide more replications. Report the confidence level too.
Exam move
Document inputs, units, seed policy, warm-up, run length, replications and estimator. Compare with an analytical special case before trusting a novel scenario. Begin a discrete-event model with an explicit state and event calendar. For a single queue, track current time, number in system, server status, next arrival and next departure.
On each event, update time-weighted statistics before changing the state, then schedule only events that can occur. Trace the first several events by hand with fixed random inputs and record a tie rule for simultaneous events. Save the expected state sequence as a regression test. Define every estimator separately from the trace.
Time-average queue size weights each state by its duration, while mean customer delay averages completed customer observations. State warm-up, observation horizon and the treatment of customers still present at the end. Inspect a cumulative mean or time series to justify removing the artificial empty-start period. Use the same warm-up and run length across designs being compared.
Build confidence intervals from independent replication estimates rather than treating every customer inside one run as independent. For each replication, record the seed or random stream, estimate and any diagnostic counts. Calculate sample mean, sample variance, standard error and the appropriate critical value.
If the half-width is too large, use its inverse square-root relation with replication count to plan additional runs, then report the achieved precision. Validate at several levels.
Check conservation, dimensions and impossible states; compare with an analytical M/M/1 or other special case; test light-load and deterministic limits; and reproduce one result with a different implementation or calculation path when stakes are high. For traffic correlation, report sampling interval and several lags. Zero contemporaneous correlation rules out only measured linear association and does not prove independence.
Connect dependence to mechanisms such as sessions, retransmissions or daily cycles, and compare aggregation windows before choosing a traffic model. On the handwritten page, keep the event-update order, time-weighted accumulator, replication interval and validation ladder. A final result should include the estimate, uncertainty, seed policy, warm-up and version of the event logic so another student can reproduce it.
Working through Simulation, Correlation and Model Checking in EE8405? Sia is AskSia’s AI Electrical Engineering tutor — ask any EE8405 Simulation, Correlation and Model Checking question and get a clear, step-by-step explanation grounded in how EE8405 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
