AI Glossary / Mathematics and optimization / Intermediate

Simulation-based optimization

Definition

Simulation-based optimization combines a simulation model with a procedure for searching for better decisions. The simulation evaluates candidate settings, and the optimizer uses those evaluations to choose what to try next. It is useful when a system's performance is difficult to express as a simple analytical formula. The simulation may be deterministic or stochastic; random simulations introduce additional uncertainty into comparisons.

What it means

A simulator can answer what happens under a proposed staffing level, inventory rule or equipment configuration. Optimization adds the question of which setting should be selected under a stated objective and constraints. The distinction matters: running a few scenarios is simulation analysis, but it is not necessarily a systematic search for a good decision. Conversely, the best setting within a simulator is only a recommendation under that model, not proof that the same setting will be best in the real system.

How it works

Specify decision variables, feasible choices and the performance measure to optimize. Run the simulator at selected candidates, collect outputs and provide them to a search method. Depending on the problem, that method might use statistical selection, gradient estimates, local search, a surrogate model or another optimization strategy. For stochastic outputs, allocate replications to estimate performance and uncertainty rather than treating one run as the true objective. Continue within a declared computational budget, then evaluate promising candidates with fresh runs or scenarios. Verification of simulation code and validation against the real system remain separate tasks from optimizing its settings.

Three practical examples

A warehouse study can search over staffing and dispatch rules using a discrete-event model of queues and work times.
An engineering design study can choose parameter settings using a deterministic physics simulator when each design evaluation is computationally expensive.
An inventory experiment can compare replenishment policies under simulated demand, including both operating cost and service requirements rather than minimizing one convenient output in isolation.

Think of it this way

Think of testing possible schedules in a detailed rehearsal before choosing one for an event. The rehearsal estimates consequences, and the search decides which schedules deserve another trial. The analogy captures the feedback loop. A simulator follows explicit assumptions and may omit important behavior, so repeated rehearsals of the same imperfect model do not automatically establish real-world reliability.

Worked example

Consider a deliberately simple one-period inventory simulator. Choose q from {80, 100, 120, 140}. Each unit ordered costs 1, and each unit of unmet demand costs an additional 4. For demand D, simulated cost is q + 4 max(D − q, 0). Use four equally weighted illustrative demand scenarios: 70, 90, 110 and 130. For q = 80, costs are [80, 120, 200, 280], averaging 170. For q = 100, they are [100, 100, 140, 220], averaging 140. For q = 120, they are [120, 120, 120, 160], averaging 130. For q = 140, all four costs are 140. Thus q = 120 has the smallest average among these four candidates on this scenario set. The check is exhaustive here, so it establishes the sample-set optimum. It does not establish the optimum for an unknown real demand distribution. If the four scenarios are intended as an exact discrete distribution, the same arithmetic gives its expected costs; if they are sampled observations, it is only an estimate. A separate service constraint could change the selected decision even when these cost averages remain unchanged.

Illustrative case study

In a fictional project, Harbor Dispatch uses a queueing simulation to explore shift schedules. The team first compares simulated arrival and completion patterns with historical observations. It searches using one set of random streams and reevaluates finalists on separate streams, recording both estimated waiting time and staffing cost. When a candidate improves average waiting but worsens a high-percentile delay measure, managers review the trade-off. The case illustrates model validation and decision comparison without inventing operational savings or a deployment result.

Common misconceptions and mistakes

The best observed simulation run is not necessarily the best expected policy. Random variation can make weak candidates look attractive, especially when many are tried. More replications reduce sampling uncertainty under the model but do not remove model bias. A fixed random seed helps reproducibility; it does not make one scenario representative. Shared random inputs can sometimes improve paired comparisons, but the variance effect depends on how outcomes co-vary. An optimizer can also exploit unrealistic simulator behavior, so an unexpectedly strong result should prompt inspection of the underlying simulated events.

Business use

Use simulation-based optimization when decisions are consequential enough to justify building and evaluating a credible model. Separate uncertainty about random outcomes, estimated input distributions and model structure. Record the simulation horizon, initialization or warm-up policy, replication design, objective and constraints. Budget both search and final evaluation so every computational resource is not spent on discovering a candidate with no time left to assess it. Report uncertainty and sensitivity alongside a recommended setting, and distinguish the best candidate found from a mathematically certified global optimum.

Check your understanding

The inventory example selects q = 120 on four scenarios. Does that prove q = 120 is optimal for next year's demand?

Answer: No. The conclusion is conditional on the tested candidates, objective and demand representation. Future demand and omitted service requirements can change it.

Related terms

Surrogate model
discrete-event simulation
stochastic optimization
sample-average approximation
ranking and selection
simulation validation

Learn next

reproduce the four inventory cost vectors, then change the shortage penalty and scenario weights separately. Observe how the preferred quantity changes. For a stochastic simulator, compare rankings from a few replications with rankings from larger independent replication sets before trusting small apparent differences.

Sources

Sources: Sigurdur Ólafsson and Jumi Kim, Simulation Optimization, Proceedings of the 2002 Winter Simulation Conference, https://www.informs-sim.org/wsc02papers/011.pdf ; Michael C. Fu, Handbook of Simulation Optimization, publisher overview, https://link.springer.com/book/10.1007/978-1-4939-1384-8 . The inventory scenario calculation and fictional dispatch study are original; no actual simulation experiment or real operating improvement is claimed.

Previous: Simulated annealing Next: Simulation component metadata