Discrete-Event Simulation (DES)
- Introduced by
- Banks, Carson, Nelson & Nicol (textbook lineage); foundational work by Tocher & Conway (1960s)
- Year
- 1960s (formalized); modern computational form from 1970s onward
- Type
- Stochastic process simulation
What does it do?
Discrete-Event Simulation (DES) is a computational modeling paradigm in which the state of a system changes only at a countable sequence of points in time — the events.
Between events nothing changes, so the simulation clock jumps directly from one event to the next. Formalized through the foundational textbooks of Banks, Carson, Nelson and Nicol and of Law in the 1960s–2000s, DES has become the standard tool for analyzing queuing systems, healthcare patient flows, manufacturing lines, and logistics networks where entities move through resources over time.
The intuition behind the method
Imagine a hospital emergency department: patients arrive, wait for a nurse, receive treatment, and leave. Real life does not update every millisecond — it changes only when something happens: a patient arrives, a nurse becomes free, or a patient is discharged. DES captures exactly this logic. The simulation maintains a sorted list of future events and processes them one by one, advancing a virtual clock to each event time rather than grinding through every millisecond. This makes it fast even for complex systems, and it lets analysts ask 'what if we add one more nurse?' or 'what if arrivals double on Mondays?' without touching the real system.
History
The foundations of DES were laid in the late 1950s and early 1960s by researchers including K.D. Tocher and R.W. Conway, who developed the first general-purpose simulation languages (GSP, GPSS). The event-scheduling paradigm — maintaining a future-event list sorted by time — became standard by the 1970s. Banks, Carson, Nelson and Nicol's textbook (first edition 1984, now in its fifth edition, 2010) and Law's companion volume (first edition 1982, now in its fifth edition, 2015) consolidated the methodology, output analysis procedures, and input-modeling techniques that define modern practice. DES spread rapidly through operations research, industrial engineering, and healthcare management, and is today implemented in both open-source libraries (SimPy, simmer) and commercial platforms.
When to use it
DES fits whenever a system can be described as discrete entities moving through resources governed by stochastic arrival and service times. It is the appropriate choice for queuing systems, emergency department and hospital ward flow, manufacturing and supply chain throughput analysis, call center staffing, and logistics routing — wherever the analyst needs to evaluate capacity, bottlenecks, or 'what-if' scenarios without experimenting on the real system. No normality assumption is required. The number of replications must be sufficient (typically at least 30, or ≥1000 for narrow confidence intervals) — with too few replications DES results are unreliable and an analytical queuing model may be preferable. Distributional inputs (arrival and service time distributions) must be estimated from historical data; when no data exist, expert elicitation or pilot runs are needed.
Limitations and considerations
- Input uncertainty: results are only as reliable as the fitted arrival and service distributions.
- Warm-up bias: ignoring the warm-up period inflates or deflates steady-state estimates.
- Replications: too few replications (fewer than ~30) produce wide confidence intervals that make scenario comparisons unreliable.
- Computational cost: complex models with many entities and resources can require substantial run time to achieve narrow confidence intervals.
- Model validation requires historical data for verification; poorly mapped processes lead to misleading outputs.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Handles complex, nonlinear processes — resource contention, variable routing, priority queues — that analytical queuing models cannot represent.
- Produces realistic confidence intervals through independent replications, making uncertainty explicit.
- Supports 'what-if' experimentation and scenario comparison without disrupting the real system.
- No distributional assumptions on outputs; any input distribution can be accommodated.
- Applicable across a wide range of domains — healthcare, manufacturing, logistics, service operations — without changing the core paradigm.
Applications
- Emergency department patient flow: evaluating staffing levels, triage policies, and bed capacity to reduce waiting times.
- Manufacturing throughput: identifying bottleneck workstations, sizing buffers, and planning shift schedules.
- Supply chain and logistics: modeling warehouse operations, delivery routing, and inventory replenishment under stochastic demand.
- Call center staffing: determining agent counts and routing rules to meet service-level agreements.
- Airport and port operations: simulating gate assignments, ground handling, and cargo throughput.
How it works
How it works
The analyst maps the real process: entities (patients, jobs, vehicles), resources (nurses, machines, trucks), queues, and routing logic. This conceptual model is the blueprint for everything that follows — vague process maps produce unreliable simulations.
Arrival inter-times and service durations are fitted to probability distributions — typically exponential, Erlang, log-normal, or empirical — using historical data and goodness-of-fit tests. Distributional input is the primary source of uncertainty in DES; if data are scarce, sensitivity analyses across plausible distribution families are essential.
The event calendar (future-event list) is initialized. Each event type (arrival, service-start, service-end, departure) is coded with the state transitions it causes. A random-number stream drives each stochastic element, and separate streams are assigned to arrival and service processes so that variance reduction techniques such as common random numbers can be applied later.
The simulation is run for a warm-up period before statistics are collected. At time zero the system is usually empty — an artificial initial condition that biases steady-state estimates. The warm-up period (estimated by Welch's method or similar) lets the system reach statistical equilibrium before measurement begins.
Multiple independent replications are run, each using a different random seed. Replications, not observations within a single run, are the units of statistical analysis. At least 30 replications are recommended for reliable confidence interval estimation; the number required can be determined by a sequential pilot procedure.
Summary statistics — mean waiting time, throughput, resource utilization — are computed across replications and reported with 95% confidence intervals. If the confidence interval is too wide, more replications or variance-reduction techniques are applied. Scenario comparisons use the same random seeds (common random numbers) to reduce between-scenario noise.
Common pitfalls
- Running fewer than 1000 replications when narrow confidence intervals are needed — use a sequential pilot procedure to determine the minimum required number.
- Omitting or under-sizing the warm-up period — steady-state metrics collected from an empty system at time zero are biased downward for utilization and upward for wait times.
- Using a single long run instead of multiple independent replications — within-run observations are autocorrelated and cannot be treated as independent data points for confidence interval estimation.
- Fitting the wrong distribution family to arrival or service times — always perform formal goodness-of-fit tests and inspect the fitted distribution against empirical data before accepting it.
- Conflating DES with continuous simulation — DES is appropriate for systems where state changes are event-driven; for continuously evolving stocks and flows, system dynamics is the correct paradigm.
Frequently asked questions
How many replications do I need?
A minimum of 30 independent replications is a common starting point for preliminary analysis, but narrow confidence intervals for decision-making typically require 100–1000 or more. Use a sequential pilot procedure: run an initial batch of 10–30 replications, estimate the variance, then compute the additional replications needed to achieve your target half-width.
How do I determine the warm-up period?
Welch's graphical method is the most widely used approach: plot a moving average of the performance metric across time, and identify the point at which it stabilizes. Alternatively, use the Mecklenburg-White test or simply run the model long enough that the warm-up represents less than 10% of total run length. Omitting the warm-up biases steady-state estimates.
When should I use DES rather than an analytical queuing model?
Analytical queuing models (M/M/c and variants) give exact or approximate closed-form results for simple, idealized systems. Use DES when the process has complex routing, multiple competing resource types, non-exponential service times, batch arrivals, or time-varying arrival rates — situations where analytical solutions are unavailable or require unrealistic simplifying assumptions.
What if I have no historical data for fitting distributions?
Expert elicitation (e.g., minimum, most-likely, maximum estimates to fit a triangular distribution) is an accepted fallback for preliminary models. Conduct sensitivity analyses across plausible distribution families to understand how input uncertainty propagates to output uncertainty. Collect real data as the model matures.
Related methods
Sources
Banks, J., Carson, J.S., Nelson, B.L. & Nicol, D.M. (2010). Discrete-Event System Simulation (5th ed.). Pearson.
ISBN: 978-0136062127Copied.Law, A.M. (2015). Simulation Modeling and Analysis (5th ed.). McGraw-Hill.
ISBN: 978-0073401324Copied.