Optimization, Simulation and Decision Analysis

Introduction

Operations and supply chain managers routinely make decisions under scarcity, uncertainty and competing objectives. A manufacturer must allocate limited capacity among products; a retailer must determine replenishment quantities when demand is uncertain; a hospital must schedule staff while maintaining service quality; and a logistics provider must balance cost, speed and reliability. Optimization, simulation and decision analysis provide complementary approaches to these problems.

Optimization identifies the best feasible decision according to a specified objective. Simulation experiments with a model of a system under alternative policies, especially when uncertainty or operational detail makes direct optimization difficult. Decision analysis structures choices, probabilities, preferences and consequences so that managers can make transparent decisions under risk. These methods are not substitutes for managerial judgment. They are disciplined ways to make assumptions explicit, compare alternatives and understand trade-offs.

Learning Objectives

By the end of this chapter, you should be able to:

1. Formulate linear, integer and nonlinear optimization models.
2. Interpret feasible solutions, objective values, shadow prices and reduced costs.
3. Conduct basic sensitivity analysis and recognize its limits.
4. Build and interpret Monte Carlo and discrete-event simulation models.
5. Apply basic queuing models to capacity and service decisions.
6. Use decision trees and multicriteria methods for structured evaluation.
7. Communicate assumptions, uncertainty, limitations and managerial implications.

1. Modeling Operational Decisions

A quantitative model is a simplified representation of a real system. A useful model specifies:

• Decision variables: controllable quantities, such as units produced or vehicles assigned.
• Parameters: known or estimated inputs, such as demand, labor hours and costs.
• Objective function: the measure to maximize or minimize.
• Constraints: physical, contractual, regulatory or policy restrictions.
• Uncertainty: unknown quantities represented by scenarios, probability distributions or stochastic processes.

A model is credible when its structure reflects the decision context, its data are defensible and its results are interpreted within the assumptions used. Mathematical sophistication cannot compensate for poor data or an incorrectly defined objective.

1.1 Linear Programming

A linear programming (LP) model assumes that contributions and resource requirements change proportionally with activity levels, and that effects are additive. A general maximization model is:

Here, xⱼ is the level of activity j (for example, units produced), cⱼ is its contribution per unit, aᵢⱼ is the amount of resource i used per unit, bᵢ is the available amount of resource i, and Z is total contribution. The coefficients must use consistent units: if cⱼ is dollars per unit, Z is dollars; if a labor coefficient is hours per unit, bᵢ must be hours.

LP applies when divisibility is acceptable or when fractional results can be rounded safely. It is inappropriate when fixed charges, indivisible items, nonlinear costs or strong interactions are material.

Worked Example: Product Mix

A plant produces products A and B. Each A contributes $40 and uses 2 machine hours and 1 labor hour. Each B contributes $30 and uses 1 machine hour and 2 labor hours. The plant has 100 machine hours and 80 labor hours.

Let x₁ and x₂ denote units of A and B. The model is:

At the intersection of both binding constraints:

Solving gives x₁ = 40 and x₂ = 20. Contribution is:

The solution uses all 100 machine hours and 80 labor hours. However, this conclusion is valid only if demand is sufficient, products are divisible, contribution values are accurate and no omitted constraint—such as materials or minimum production quantities—changes the result.

2. Integer and Mixed-Integer Optimization

Some decisions must be whole numbers. A truck, warehouse, employee or production run cannot generally be represented by 0.4 units. Integer programming requires selected variables to take integer values:

Binary variables take only 0 or 1 and are useful for yes/no decisions. For example, yₖ = 1 if facility k is opened and 0 otherwise. A fixed-charge facility model may include:

where xₖ is flow through facility k, Mₖ is a valid upper bound on that flow, and yₖ indicates opening. This constraint prevents flow through a closed facility. Mₖ should be as small as logically valid; excessively large values can cause numerical instability and weak computational performance.

Mixed-integer linear programming (MILP) combines continuous and integer variables. It can represent routing, scheduling, lot sizing, workforce planning, supplier selection and network design. Common formulations include:

• Assignment: each job is assigned to one resource.
• Transportation: flows move from origins to destinations.
• Facility location: facilities are opened and customer demand is allocated.
• Scheduling: jobs are sequenced subject to precedence and capacity.

Integer models are usually harder to solve than LP models because the feasible region is discrete. Solvers commonly use branch-and-bound, cutting planes and presolve techniques. A solver’s optimality gap should be reported when a proven optimum is not available.

3. Nonlinear Optimization

Nonlinear programming (NLP) is appropriate when costs, returns or constraints are nonlinear. Examples include quantity discounts, congestion, risk penalties and economies of scale. A general formulation is:

The functions f, gᵢ and hₖ may be nonlinear. A smooth unconstrained local optimum often satisfies the first-order condition:

where ∇f is the gradient and x* is a candidate solution. For constrained problems, Karush–Kuhn–Tucker conditions provide necessary conditions under suitable constraint qualifications; they do not automatically guarantee a global optimum.

If a minimization objective is convex and the feasible region is convex, any local optimum is also global. Without convexity, algorithms may find only a local optimum, and multiple starting points or global optimization methods may be needed. Analysts should therefore distinguish “solver-reported optimum” from “globally verified optimum.”

4. Sensitivity and Post-Optimal Analysis

An optimal solution is conditional on model inputs. Sensitivity analysis asks how the solution changes when coefficients or resource limits change.

The shadow price of a binding resource is the marginal change in the optimal objective value from increasing that resource by one unit, within the allowable range over which the current basis remains valid. Its units are objective units per resource unit—for example, dollars per machine hour.

In the product-mix example, the shadow prices can be found from:

where a₁ is the value of a machine hour and a₂ is the value of a labor hour. Solving gives a₁ = 20/3 ≈ $6.67 per machine hour and a₂ = 50/3 ≈ $16.67 per labor hour. Thus, within the valid sensitivity range, one additional labor hour increases contribution by approximately $16.67, while one additional machine hour increases it by approximately $6.67.

Sensitivity analysis should examine:

• changes in objective coefficients, such as product margins;
• changes in right-hand-side resources, such as capacity;
• changes in demand, lead times and costs;
• alternative constraints or policy assumptions.

Shadow prices are not permanent economic values. They may change when the active constraints or optimal basis changes. Sensitivity analysis also does not replace scenario analysis when several assumptions change simultaneously or when uncertainty is substantial.

5. Simulation as an Experimental Method

Simulation imitates the operation of a system through time or repeated random sampling. It is especially useful when a system contains variability, queues, complex rules or interactions that are difficult to express analytically.

5.1 Monte Carlo Simulation

Monte Carlo simulation repeatedly samples uncertain inputs and calculates the resulting output. If demand D is modeled by a probability distribution and profit is:

then repeated draws D₁, D₂,...,Dₙ produce simulated profits Π₁, Π₂,...,Πₙ. The estimated expected profit is:

where N is the number of replications. The estimated standard error of the mean is:

where s is the sample standard deviation of simulated profits. Independent replications and a sufficiently large N are assumed. A simulation estimate is not exact; a confidence interval or uncertainty statement should accompany it.

Random numbers must be generated using a reliable pseudorandom generator, and the analyst should document distributions, parameters, seeds and validation procedures. Correlations—such as demand and price moving together—must not be ignored without justification.

5.2 Discrete-Event Simulation

Discrete-event simulation represents a system as a sequence of events that change its state at particular times. Typical events include an arrival, service completion, machine failure, replenishment receipt or departure. The model maintains a future-event list and advances the simulation clock to the next event.

A basic procedure is:

1. Define entities, resources, queues, states and events.
2. Initialize the clock and system state.
3. Generate or schedule the next event.
4. Advance time to that event.
5. Update the state and collect statistics.
6. Continue until the stopping rule is reached.
7. Repeat with independent replications and analyze outputs.

Discrete-event simulation can evaluate warehouse picking policies, production bottlenecks, emergency-department flow and transport terminals. It does not automatically identify the best policy; alternatives must be designed and compared. Optimization and simulation can be combined in simulation-optimization, where an optimizer searches policy choices and simulation evaluates uncertain performance.

5.3 Verification, Validation and Experimental Design

Verification asks whether the model was implemented correctly. Techniques include code review, extreme-value tests and conservation checks. Validation asks whether the model represents the real system sufficiently for its intended use. Historical comparison, expert review and face validation are useful, but no model is valid for every purpose.

A simulation study should specify warm-up periods for terminating bias in steady-state systems, run length, number of replications, random-number management and performance measures. Comparing policies requires common random numbers or other variance-reduction methods only when implemented correctly; otherwise apparent differences may be sampling noise.

6. Queuing Models

Queuing theory analyzes waiting lines formed when demand for service is variable and service capacity is limited. Important measures include arrival rate λ, service rate μ, utilization ρ, average number in the system L, average number waiting Lq, average time in the system W and average waiting time Wq.

For a stable single-server M/M/1 queue, arrivals are Poisson, service times are exponentially distributed, one server operates, the queue is typically first-come-first-served and λ < μ. Then:

Rates must use the same time unit. Little’s Law applies broadly under steady-state conditions:

and, for the queue alone, Lq = λWq. As utilization approaches 1, waiting time increases sharply. This explains why operating a service system at 100% nominal capacity is generally impractical.

For multiple identical servers, an M/M/c model may be appropriate, but its formulas and assumptions differ. Real systems may have finite capacity, priority rules, abandonment, batch arrivals or non-exponential service. Simulation is often preferable when these details materially affect decisions.

7. Decision Analysis Under Risk

Decision analysis separates choices from uncertain states of nature. A decision tree contains decision nodes, chance nodes and outcome branches. If decision d has outcomes j with probabilities pⱼ and monetary payoffs vⱼ, expected monetary value is:

The rational choice under risk neutrality is the decision with the greatest EMV for a maximization problem. However, managers may be risk-averse. Expected utility replaces monetary payoff with utility u(v):

The utility function should reflect the decision maker’s risk preferences, not be selected merely to produce a preferred answer.

The value of perfect information is the expected value with perfect information minus the best EMV without it. Information is valuable only if its cost is lower than the value it creates, and real information is imperfect. Bayesian updating can revise probabilities when new evidence is received, provided the evidence model is credible.

8. Multicriteria Decision Methods

Operational decisions often involve cost, service, resilience, emissions, safety and strategic fit. A weighted-sum model may score alternative k as:

where wᵢ is the weight for criterion i, rᵢₖ is the normalized score of alternative k on criterion i, and ∑wᵢ = 1. Benefit criteria should be scaled so higher values are better; cost criteria must be reversed or transformed consistently.

A transparent linear normalization for a benefit criterion is:

For a cost criterion, one may use:

These formulas require a nonzero range and can be sensitive to the alternatives included. Weights should be elicited from stakeholders and tested through sensitivity analysis. Methods such as AHP, ELECTRE, TOPSIS and goal programming offer alternatives, but no method removes value judgments. A ranking that changes under small weight changes should be reported as unstable rather than presented as definitive.

9. Integrated Case: Regional Distribution Network

The following is an illustrative case, with simplified data.

Background and Problem

Northstar Foods serves 600 retail stores from two possible distribution centers. Management must decide whether to open a second center, how much inventory to hold and which transportation policy to use. The current center has limited throughput; demand varies weekly; delivery delays affect customer service.

Analysis

The team first formulated a MILP. Binary variables represented whether each candidate center was opened, continuous variables represented store flows, and integer variables represented weekly truck departures. Fixed facility costs, transportation costs, capacity limits and service requirements were included.

Because demand and travel times were uncertain, the team then built a Monte Carlo model. Weekly store demand was sampled from historical distributions, while travel times were sampled by lane and day type. The model estimated total cost, stockout probability and the probability of meeting a two-day delivery promise.

A discrete-event model represented order arrivals, picking, loading, truck departure and delivery. It revealed that the apparent bottleneck was not storage capacity but a narrow loading window. An M/M/1 approximation initially suggested severe congestion because utilization was near 0.90, but the actual process had batch departures and priority orders; therefore, simulation was used for the final comparison.

Finally, management used a multicriteria score with weights of 0.45 for cost, 0.30 for service reliability, 0.15 for resilience and 0.10 for emissions. Weight sensitivity showed that the preferred network changed when service reliability received more than approximately 0.40 weight. This result was more useful than a single ranking because it identified the strategic trade-off.

Approach and Outcome

The recommended policy was to open the second center only if demand exceeded a specified threshold, while first expanding loading capacity at the existing center. The contingent policy reduced the risk of premature fixed investment and improved delivery performance in high-demand scenarios. The analysis did not claim certainty: results depended on demand distributions, correlations, labor availability and the assumed service penalty.

Learning Points

1. Optimization is useful for structural allocation and network design.
2. Simulation captures operational variability and process interactions.
3. Queuing analysis highlights the cost of high utilization.
4. Multicriteria analysis makes strategic preferences explicit.
5. A recommendation should include trigger conditions, uncertainty and implementation risks.

10. Governance, Ethics and Contemporary Practice

Optimization systems can encode unfair priorities. For example, a routing model that minimizes cost may systematically provide slower service to remote or disadvantaged communities. Workforce schedules may optimize coverage while creating unsafe fatigue levels. Data-driven models can reproduce historical discrimination or conceal accountability behind technical language.

Best practice includes documenting data sources, objective functions, constraints, model versions and approval authority. Sensitive decisions require human oversight, audit trails, access controls and monitoring for drift. Privacy laws and employment, safety, environmental and transportation regulations may constrain data collection and feasible decisions.

AI and machine learning can improve demand forecasts, anomaly detection and surrogate modeling, but forecast accuracy does not guarantee decision quality. An optimization model using biased forecasts can produce highly efficient but inequitable outcomes. Generative AI may assist with code or scenario documentation, but analysts must independently verify formulas, data transformations and solver outputs.

11. Best Practices and Limitations

A reliable analytics project should:

• define the decision and success measures before choosing a method;
• separate inputs, assumptions, model logic and outputs;
• use dimensional checks and plausibility tests;
• validate against historical or operational evidence;
• test extreme cases and alternative scenarios;
• report uncertainty, sensitivity and infeasibility;
• involve process owners and affected stakeholders;
• distinguish mathematical optimality from practical implementability;
• monitor results after implementation and update the model.

Common failures include optimizing an inappropriate objective, omitting constraints, treating estimates as facts, using simulation without enough replications, interpreting correlation as causation and presenting a precise ranking when the result is highly assumption-dependent. Models should support learning and disciplined judgment, not create false confidence.

Chapter Summary

Optimization formulates operational decisions using variables, objectives and constraints. Linear programming is appropriate for proportional, additive relationships; integer and mixed-integer models represent indivisible and yes/no decisions; nonlinear programming represents curved objectives or constraints and requires careful attention to local versus global optimality. Sensitivity analysis explains how resource values and solutions respond to changing inputs, but shadow prices apply only within relevant ranges.

Monte Carlo simulation propagates input uncertainty through repeated sampling, while discrete-event simulation represents time-based system behavior, queues and operational rules. Verification concerns implementation correctness; validation concerns representation of the real system. Queuing models provide useful analytical insight, particularly into the effects of utilization, but simulation may be necessary for complex processes. Decision trees and expected value organize choices under risk, while expected utility reflects risk preferences. Multicriteria methods expose trade-offs among cost, service, resilience and other objectives. Across all methods, transparent assumptions, ethical governance, sensitivity analysis and managerial interpretation are essential.

Key Terms

Optimization: Selecting the best feasible decision under a defined objective.

Decision variable: A controllable quantity determined by the model.

Constraint: A restriction defining feasible decisions.

Linear programming: Optimization with linear objective and constraint functions.

Integer programming: Optimization requiring selected variables to be whole numbers.

Binary variable: A variable restricted to 0 or 1.

Mixed-integer programming: A model containing both continuous and integer variables.

Nonlinear programming: Optimization with at least one nonlinear objective or constraint.

Feasible solution: A decision satisfying every constraint.

Shadow price: Marginal objective value of relaxing a resource constraint within a valid range.

Sensitivity analysis: Examination of how results respond to input changes.

Monte Carlo simulation: Repeated random sampling to estimate uncertain outcomes.

Discrete-event simulation: Simulation in which state changes occur at distinct events.

Verification: Checking whether a model is implemented correctly.

Validation: Assessing whether a model is suitable representation of the intended system.

Queuing theory: Analysis of waiting lines and service systems.

Utilization: The proportion of service capacity being used.

Expected monetary value: Probability-weighted average monetary payoff.

Multicriteria decision analysis: Evaluation of alternatives across multiple criteria.

Optimality gap: Difference between an incumbent integer solution and a bound on the best possible solution.

Review Questions

Short Answer Questions

1. What are the four basic components of an optimization model?
2. Why are binary variables useful in supply chain models?
3. Distinguish simulation verification from simulation validation.
4. State Little’s Law and define its variables.
5. Why should sensitivity analysis accompany a multicriteria ranking?

Long Answer Questions

1. Explain the assumptions of linear programming and discuss situations in which they may be inappropriate.
2. Compare linear programming, mixed-integer programming and nonlinear programming in terms of representation and computational difficulty.
3. Explain how Monte Carlo and discrete-event simulation differ and when each is useful.
4. Discuss the meaning and limitations of shadow prices in managerial decision-making.
5. Explain how ethical and governance concerns can affect optimization and AI-enabled operational decisions.

Analytical / Application Questions

1. A service desk receives 18 requests per hour and one agent can serve 24 requests per hour. Assuming an M/M/1 model, calculate utilization, average time in the system and average waiting time. Interpret the result and discuss whether the assumptions are realistic.
2. Formulate a mixed-integer model for selecting distribution centers, assigning customer demand and enforcing a fixed opening cost. Define variables, objective and at least four constraints.
3. A company ranks suppliers on cost, quality, resilience and emissions. Design a normalized weighted-sum method, explain how weights should be obtained and describe a sensitivity analysis that would test ranking stability.

Multiple Choice Questions with Answers

Question 1

In a linear programming model, what does a decision variable represent?

A. A controllable quantity whose value is determined by the model

B. A fixed historical observation

C. A probability distribution only

D. A constraint that cannot be changed

Correct Answer: A. A controllable quantity whose value is determined by the model

Question 2

Which variable type is most appropriate for representing whether a warehouse is opened?

A. A nonnegative continuous variable

B. A binary variable

C. A normally distributed variable

D. A shadow-price variable

Correct Answer: B. A binary variable

Question 3

A shadow price is best interpreted as:

A. The total fixed cost of a resource

B. The average cost of all products

C. The marginal objective improvement from one additional resource unit within a valid range

D. The probability that a constraint is binding

Correct Answer: C. The marginal objective improvement from one additional resource unit within a valid range

Question 4

For an M/M/1 queue with arrival rate λ = 6 per hour and service rate μ = 8 per hour, utilization is:

A. 0.25

B. 0.50

C. 0.75

D. 1.33

Correct Answer: C. 0.75

Question 5

Which statement correctly describes Monte Carlo simulation?

A. It removes all uncertainty from a model

B. It repeatedly samples uncertain inputs to estimate outcome distributions

C. It guarantees a mathematically optimal decision

D. It applies only to queueing systems

Correct Answer: B. It repeatedly samples uncertain inputs to estimate outcome distributions

Question 6

In discrete-event simulation, the simulation clock typically advances to:

A. The next scheduled event time

B. The average of all historical times

C. The maximum possible time

D. A randomly selected calendar date unrelated to events

Correct Answer: A. The next scheduled event time

Question 7

Expected monetary value is calculated by:

A. Selecting the largest payoff only

B. Multiplying every payoff by the same cost

C. Taking the probability-weighted sum of possible payoffs

D. Subtracting probability from payoff

Correct Answer: C. Taking the probability-weighted sum of possible payoffs

Question 8

Which condition supports the conclusion that a local optimum is global in a minimization problem?

A. The objective is convex and the feasible region is convex

B. At least one variable is binary

C. The model has no constraints regardless of the objective

D. The solver uses a random starting point

Correct Answer: A. The objective is convex and the feasible region is convex

Question 9

In multicriteria analysis, why must cost criteria often be transformed before combining them with benefit scores?

A. To ensure higher normalized scores consistently indicate better performance

B. To eliminate the need for weights

C. To make every criterion measured in dollars

D. To guarantee that all alternatives tie

Correct Answer: A. To ensure higher normalized scores consistently indicate better performance

Question 10

What is the primary purpose of simulation validation?

A. To prove that every future observation will be predicted exactly

B. To determine whether the model is a suitable representation for its intended use

C. To reduce all model inputs to integers

D. To replace stakeholder judgment

Correct Answer: B. To determine whether the model is a suitable representation for its intended use