Optimization, Simulation and Decision Models

Introduction

Logistics systems are designed and operated under scarcity, uncertainty and interdependence. A manager may need to allocate limited warehouse capacity, select transport modes, position inventory, schedule vehicles, or balance cost against service and resilience. Descriptive analytics explains what has happened; predictive analytics estimates what may happen; optimization and simulation support decisions about what should be done and what consequences may follow.

Optimization identifies the best feasible decision according to a defined objective. Simulation represents the behaviour of a system under specified policies and uncertain inputs. Decision models combine alternatives, information, preferences and uncertainty so that managers can compare actions systematically. No model is automatically “correct”: its usefulness depends on the quality of its assumptions, data, objective, constraints and validation.

This chapter develops linear, integer, nonlinear and multi-objective models, introduces heuristics and metaheuristics, explains Monte Carlo and discrete-event simulation, and shows how to interpret solutions, shadow prices and sensitivity results in a logistics context.

Learning Objectives

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

1. Formulate logistics decisions as optimization models with variables, objectives and constraints.
2. Distinguish linear, integer, nonlinear and multi-objective formulations.
3. Select exact methods, heuristics or simulation according to problem structure.
4. Build and interpret Monte Carlo and discrete-event simulation models.
5. Validate assumptions, inputs, logic and outputs.
6. Explain binding constraints, shadow prices, reduced costs and sensitivity ranges.
7. Assess ethical, governance and managerial implications of model-supported decisions.

1. Decision Models in Logistics

A decision model translates a real system into a simplified representation. A useful model normally contains:

• Decision variables: controllable quantities, such as shipment quantities, facility-opening choices or vehicle assignments.
• Parameters: known or estimated inputs, such as demand, costs, capacities and travel times.
• Objective function: a measure to minimize or maximize.
• Constraints: physical, contractual, regulatory or policy requirements.
• Uncertainty: variation in demand, processing time, disruption, fuel price or service performance.
• Boundary and timing assumptions: the entities, locations, time periods and decisions included.

A general deterministic optimization model can be written as:

Here, x is the vector of decisions, θ is the vector of parameters, f is the objective, g and h represent inequality and equality constraints, and X defines domains such as non-negativity or integrality. The formulation is applicable when the relevant relationships can be represented adequately by functions and constraints.

A model should not optimize a proxy without examining its consequences. Minimizing transportation cost alone may increase emissions, late deliveries or labour risk. Conversely, imposing every desirable goal as a hard constraint may make the model infeasible or excessively complex. Model scope is therefore a managerial as well as a mathematical decision.

2. Linear Programming

Linear programming (LP) applies when the objective and all constraints are linear in the decision variables, and variables may take continuous values. The canonical form is:

The coefficient cᵢ is the objective contribution per unit of xᵢ; aᵢⱼ is the amount of resource j used by one unit of activity i; and bⱼ is the available resource or required level. Units must be consistent: if x is pallets, a capacity coefficient should be pallet-hours per pallet and b should be pallet-hours.

Transportation example

A firm ships from two depots to three customers. Let xᵢⱼ be pallets shipped from depot i to customer j. With supply sᵢ, demand dⱼ and unit cost cᵢⱼ, a balanced transportation model is:

The model assumes divisible flow, known costs and demand, and no route-specific fixed charge. If a vehicle must be dispatched before any shipment occurs, a fixed-charge or integer model is more appropriate.

Worked example: capacity allocation

A warehouse can process 100 labour-hours. Product A yields a contribution of £40 per pallet and requires 2 hours; product B yields £30 and requires 1 hour. Demand limits are 30 pallets of A and 60 pallets of B. Let a and b denote pallets.

At a = 20 and b = 60, all 100 hours are used and contribution is £2,600. At a = 30, the remaining capacity permits b = 40, producing £2,400. The first solution is superior. The marginal contribution per processing hour is £20 for A and £30 for B, so B receives capacity first until its demand limit is reached; the remaining capacity produces A. This reasoning agrees with the LP solution.

The result is not automatically a recommendation. Management should verify whether product mix can be changed quickly, whether demand limits are reliable, and whether labour-hour capacity is truly the bottleneck.

3. Integer and Mixed-Integer Models

Many logistics decisions are indivisible. A facility is either opened or not; a vehicle is assigned or not; a route uses a link or does not. Integer programming requires some or all variables to be integers. A mixed-integer linear program (MILP) contains both continuous and integer variables.

A facility-location model uses yᵢ = 1 if facility i is opened and 0 otherwise, and xᵢⱼ = quantity shipped from facility i to customer j:

Fᵢ is the fixed opening cost and Kᵢ is facility capacity. The linking constraint ∑ⱼxᵢⱼ ≤ Kᵢyᵢ prevents shipments from an unopened facility. Omitting this constraint creates an economically meaningless solution.

Common integer formulations include assignment, set covering, vehicle routing, scheduling and lot-sizing. Binary variables can also encode logical relationships. For example, if activity A requires activity B, y_A ≤ y_B. If exactly one of two alternatives must be selected, y₁ + y₂ = 1.

MILPs are often solved using branch-and-bound or branch-and-cut. The solver repeatedly partitions the feasible region and uses linear-programming bounds to eliminate branches. A reported solution should include the objective value, feasibility status, optimality gap, computational time and any time limit. A feasible solution found within a time limit is not necessarily proven optimal.

4. Nonlinear Optimization

Nonlinear programming arises when cost, capacity, congestion, service or risk relationships are nonlinear. Examples include quantity discounts, congestion-dependent travel time, economies of scale and quadratic risk penalties.

A general nonlinear model is:

If the objective and feasible region are convex, every local optimum is a global optimum under standard regularity conditions. For a differentiable unconstrained problem, a necessary first-order condition is:

For constrained problems, Karush–Kuhn–Tucker conditions provide necessary conditions under suitable constraint qualifications. They are not, by themselves, a guarantee of global optimality in a non-convex model.

For example, a congestion cost may be represented as C(q) = αq², where q is flow and α is a non-negative coefficient. The units of α must be cost per flow-squared so that C has units of cost. Such a formulation should be used only when the quadratic relationship is empirically defensible over the operating range. Nonlinear solvers may return local optima, depend on starting points and require careful scaling.

5. Multi-Objective Optimization

Logistics decisions commonly balance cost, service, emissions, risk and resilience. These objectives may conflict. A solution is Pareto efficient if no objective can be improved without worsening at least one other objective. The set of such solutions is the Pareto frontier.

One practical approach is the weighted-sum model:

The objectives must first be normalized or expressed in comparable units. A common normalization is:

where the benchmark values define a meaningful scale. Weights represent preferences, not objective probabilities. A weighted sum may fail to reveal non-convex portions of a Pareto frontier; the ε-constraint method is often more informative. It optimizes one objective while imposing limits on others, for example minimizing cost subject to emissions ≤ ε and late deliveries ≤ δ.

Managers should examine trade-off tables rather than accept a single weighted answer. Sensitivity to weights is itself a governance question: if small changes produce radically different decisions, the recommendation is preference-sensitive and should be presented transparently.

6. Heuristics and Metaheuristics

Exact optimization is attractive but may be computationally impractical for large vehicle-routing, scheduling or network-design problems. A heuristic is a problem-solving rule that seeks a good solution without guaranteeing optimality. Examples include nearest neighbour, savings methods, insertion heuristics, local search and greedy assignment.

A metaheuristic provides a broader search strategy. Simulated annealing accepts some worsening moves early to escape local optima; genetic algorithms evolve populations of candidate solutions; tabu search prevents cycling through recently visited solutions; and ant-colony methods use artificial pheromone trails. Their quality depends on representation, feasibility repair, parameter tuning, stopping rules and random seeds.

A professional comparison should report:

• objective value and constraint violations;
• runtime and computational resources;
• gap from a known optimum or lower bound where available;
• results across repeated runs;
• robustness under changed data;
• operational feasibility and explainability.

A heuristic is not validated merely because it produces a visually plausible route. It must be compared with benchmarks, historical practice or exact solutions on smaller instances.

7. Monte Carlo Simulation

Monte Carlo simulation propagates uncertain inputs through a model by repeated random sampling. If Y = g(X₁,…,X_p), then simulation generates observations Y₁,…,Y_N and estimates the expected output with:

The sample standard deviation is:

The estimated standard error of the mean is s/√N when replications are independent and identically distributed. A normal-approximation 95% confidence interval is approximately Ȳ ± 1.96(s/√N), provided the approximation is reasonable. More replications reduce sampling error, but they do not correct biased distributions, incorrect correlations or flawed system logic.

A simulation study should specify input distributions, parameters, dependence structure, random-number seeds, number of replications, warm-up treatment where relevant, performance measures and confidence intervals. Triangular distributions can be useful when only minimum, most-likely and maximum values are available, but they should not be used automatically when empirical data exist.

8. Discrete-Event Simulation

Discrete-event simulation (DES) models a system as a chronological sequence of events that change state at distinct times. Typical logistics events include order arrival, vehicle arrival, picking completion, loading completion and departure. The event calendar advances the clock to the next event; state variables are updated; and future events are scheduled.

A warehouse DES may contain entities (orders and pallets), resources (pickers, docks and forklifts), queues, routing rules, processing-time distributions and shift calendars. Outputs may include throughput, waiting time, resource utilization, queue length and service-level attainment.

DES is particularly useful when congestion, priority rules, blocking, batching or variability make analytical formulas unrealistic. It can compare policies such as first-come-first-served versus priority dispatching. It does not automatically optimize a policy: simulation evaluates specified policies. Optimization over simulation outputs requires experimental design, ranking-and-selection methods, response-surface methods or simulation-based optimization.

A key distinction is between terminating systems, such as a daily delivery operation with a defined end, and steady-state systems, such as a continuously operating distribution centre. Steady-state studies commonly require a warm-up period to remove initialization bias and multiple independent replications to estimate uncertainty.

9. Validation and Verification

Verification asks whether the model was implemented correctly: are equations, event logic, units and algorithms coded as intended? Validation asks whether the model is an adequate representation for its intended use. A model can be verified yet invalid if its assumptions do not reflect reality.

A disciplined process includes:

1. Face validation with subject-matter experts.
2. Dimensional and logical checks, including conservation of flow.
3. Extreme-condition tests, such as zero demand or unlimited capacity.
4. Historical back-testing where suitable data exist.
5. Comparison with analytical benchmarks or small exact instances.
6. Sensitivity and scenario analysis.
7. Independent review of data, code and assumptions.

Calibration should not be confused with validation. Fitting parameters to historical data may improve in-sample fit while leaving the model unable to predict changed conditions. Validation data, where possible, should be separate from calibration data. For stochastic simulation, observed and simulated distributions can be compared using confidence intervals, quantile plots or appropriate statistical tests, while recognizing that statistical similarity does not prove causal validity.

10. Interpreting Optimization Results

A solver output is a decision aid, not a managerial conclusion. Important interpretations include:

• Feasible solution: satisfies every constraint.
• Binding constraint: holds at equality at the optimum and limits improvement locally.
• Slack: unused amount of a ≤ resource constraint; for a constraint aᵀx ≤ b, slack is b − aᵀx.
• Shadow price: the marginal change in optimal objective value per one-unit increase in a right-hand-side resource, within the valid sensitivity range.
• Reduced cost: for a variable at its bound, the marginal objective penalty associated with forcing it into the solution, under the model’s conventions.
• Sensitivity range: the parameter interval over which the reported marginal interpretation remains valid, usually without a change in the optimal basis for LP.

For a maximization model with a resource constraint, a shadow price of £12 per labour-hour means that, locally and under unchanged assumptions, one additional hour could improve the objective by approximately £12. It does not mean that every additional hour is worth £12 indefinitely. Once another constraint becomes binding, the value may change. Shadow prices from continuous LPs should not be interpreted uncritically in MILPs, where discrete changes can make marginal values discontinuous.

Post-optimal analysis should vary demand, costs, capacities, service requirements and disruption assumptions. Scenario analysis changes a coherent set of parameters; sensitivity analysis usually changes one parameter or a small group to examine local effects. Both are needed for strategic decisions.

11. Integrated Illustrative Case: Regional E-Commerce Fulfilment

The following is a fictional illustrative case.

Background and problem

Northstar Retail serves three markets from two candidate fulfilment centres. Centre East has a fixed annual cost of £480,000 and capacity of 18,000 orders; Centre West costs £360,000 and has capacity of 14,000 orders. Annual demand is 25,000 orders. Variable fulfilment and transport costs differ by market. Management must decide which centres to open and how to allocate orders. Demand may vary by ±20%, and processing time is uncertain during peak weeks.

Optimization approach

A MILP uses binary opening variables and continuous order-allocation variables. It includes demand satisfaction, capacity linking and fixed costs. The objective minimizes fixed plus variable cost. Because demand is uncertain, the deterministic solution is tested under low, base and high demand scenarios. A multi-objective extension adds emissions and a penalty for late orders.

The base model selects the lower-cost configuration subject to capacity. Under high demand, opening only West becomes infeasible because 25,000 × 1.20 = 30,000 orders exceeds its 14,000-order capacity. Opening both centres provides 32,000 units of capacity, leaving 2,000 units of nominal headroom. That headroom is not necessarily adequate: variability in pick times and temporary equipment failure can create queues even when average capacity is sufficient.

Simulation approach

A DES represents hourly order arrivals, picking times, packing stations, dock availability and dispatch cut-offs. Arrival counts follow a fitted hourly distribution, and picking times are sampled from observed order classes. Five warm-up days are discarded for a steady-state peak-week experiment, followed by 30 independent replications. Each policy is evaluated using mean throughput, 95th-percentile order waiting time, utilization and the proportion dispatched before cut-off.

Outcome and interpretation

The lowest-cost optimization plan is not automatically the best operating plan if it produces excessive congestion. Simulation may show that opening both centres and shifting a modest share of orders to West reduces late dispatches at a small incremental cost. Management can then examine a Pareto set rather than a single cost-minimizing answer. A recommended decision should state the demand scenario, service target, confidence interval, capacity assumptions and contingency trigger.

Learning points

The case demonstrates that optimization allocates resources under stated assumptions, while simulation tests dynamic performance and variability. The two methods are complementary. Optimization may identify where to operate and how much to allocate; DES may reveal queues, congestion and service failures hidden by aggregate constraints. Neither method replaces managerial judgment or data governance.

12. Technology, AI and Governance

Modern solvers use presolve, parallel computation, decomposition and advanced cutting planes. Cloud platforms make large scenario studies accessible, while digital twins connect operational data with simulation models. Machine learning can forecast demand, estimate travel times or classify order types, but a forecast is not an optimization model. Forecast errors, distribution shift and biased historical data can propagate into decisions.

AI-assisted optimization should therefore maintain human oversight, version-controlled data, reproducible experiments and clear decision rights. Sensitive variables should be examined for discriminatory effects, especially in workforce scheduling, delivery access and service prioritization. Explainability is essential when decisions affect employees, suppliers or customers. Models should protect personal and commercially sensitive data, comply with applicable privacy and sector regulations, and retain audit trails of inputs, model versions, overrides and outcomes.

13. Best-Practice Workflow

A robust logistics analytics project follows this sequence:

1. Define the decision, time horizon, stakeholders and success measures.
2. Map the real process and identify controllable decisions.
3. Collect, clean and document data; record units and provenance.
4. Build the simplest adequate model before adding complexity.
5. Formulate objectives, constraints and uncertainty explicitly.
6. Establish feasibility checks and benchmark solutions.
7. Solve, simulate or search using reproducible settings.
8. Verify implementation and validate against expert knowledge and data.
9. Conduct sensitivity, scenario and stress testing.
10. Present alternatives, trade-offs, uncertainty and implementation conditions.
11. Monitor actual performance and update the model after deployment.

The final step is important: logistics systems change. A model that was valid before a network redesign, labour agreement or demand shock may no longer be valid afterward.

Challenges and Limitations

Optimization can produce false precision when costs or capacities are poorly estimated. Linear models may omit indivisibilities; integer models may be computationally difficult; nonlinear models may be non-convex; and multi-objective models may conceal subjective weights. Simulation can be expensive, statistically noisy and difficult to validate. Heuristics provide no universal optimality guarantee. All approaches can suffer from incomplete data, omitted constraints, inappropriate distributions and implementation resistance.

A technically sophisticated model can therefore be less useful than a simpler transparent model that stakeholders understand and can maintain. The appropriate standard is not maximum complexity but fitness for the decision, evidence and governance requirements.

Chapter Summary

Optimization models represent controllable logistics decisions through variables, objectives and constraints. LP is appropriate for continuous linear relationships; integer and mixed-integer models represent indivisible choices and fixed charges; nonlinear models capture curved relationships subject to additional assumptions; and multi-objective models expose trade-offs among cost, service, emissions and risk. Heuristics and metaheuristics provide practical solutions when exact optimization is difficult, but their performance must be benchmarked.

Monte Carlo simulation estimates the effects of uncertain inputs through repeated sampling. Discrete-event simulation represents queues, resources and chronological operational events, making it valuable for congestion and service analysis. Verification checks implementation, whereas validation assesses whether the model represents the real system adequately. Solver results must be interpreted through feasibility, binding constraints, slack, shadow prices, reduced costs and sensitivity limits. Effective practice combines optimization, simulation, data governance, expert judgment and ongoing monitoring.

Key Terms

Optimization: Selecting the best feasible decision according to a defined objective.

Decision variable: A controllable quantity determined by the model.

Parameter: An input treated as known or estimated in a model.

Constraint: A mathematical restriction on feasible decisions.

Linear programming: Optimization with linear objective and constraint relationships and continuous variables.

Mixed-integer linear programming: Linear optimization containing continuous and integer variables.

Binary variable: An integer variable restricted to 0 or 1.

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

Convexity: A structural property supporting global-optimum interpretations in suitable models.

Pareto efficiency: A state where improving one objective requires worsening another.

Heuristic: A practical solution rule without a general optimality guarantee.

Metaheuristic: A general search framework for exploring difficult solution spaces.

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

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

Verification: Checking whether a model has been implemented as intended.

Validation: Assessing whether a model is adequate for its intended purpose.

Shadow price: Local marginal value of relaxing a resource or right-hand-side constraint.

Binding constraint: A constraint that is active at the solution.

Sensitivity analysis: Examination of how results change when inputs vary.

Digital twin: A data-connected digital representation of a physical system.

Review Questions

Short Answer Questions

1. What are the principal components of an optimization model?
2. Why are binary variables required in facility-location models?
3. Distinguish verification from validation.
4. What does a shadow price represent in a continuous LP?
5. Why may a DES require a warm-up period?

Long Answer Questions

1. Compare LP, MILP and nonlinear programming for logistics applications, including assumptions and computational implications.
2. Explain how a multi-objective logistics model can reveal trade-offs among cost, service and emissions.
3. Describe a rigorous workflow for validating a Monte Carlo or discrete-event simulation.
4. Discuss the strengths and limitations of heuristics compared with exact optimization.
5. Explain how optimization and simulation can be integrated in fulfilment-centre design.

Analytical / Application Questions

1. Formulate a facility-location MILP for three candidate depots and four customer zones, including fixed costs, capacities and demand satisfaction. State the meaning and units of every variable and parameter.
2. A warehouse LP has a labour-capacity shadow price of £9 per hour over a stated sensitivity range. Explain what management may infer, what it may not infer, and how the interpretation changes if the model is converted to a MILP.
3. Design a simulation experiment to compare two dispatching policies during a peak week. Specify input data, replications, warm-up treatment, output measures, confidence reporting and validation tests.

Multiple Choice Questions with Answers

Question 1

Which feature most directly requires a mixed-integer model?

A. A continuous transportation quantity

B. A facility that is either opened or closed

C. A linear processing cost

D. A known customer demand

Correct Answer: B. A facility that is either opened or closed

Question 2

In an LP resource constraint, slack is calculated as:

A. Used resource minus available resource

B. Available resource minus used resource

C. Objective value divided by resource use

D. Fixed cost minus variable cost

Correct Answer: B. Available resource minus used resource

Question 3

A shadow price is best interpreted as:

A. The total cost of a constraint

B. The local marginal value of relaxing a right-hand-side constraint

C. The probability that a constraint is violated

D. The reduced cost of every nonbasic variable

Correct Answer: B. The local marginal value of relaxing a right-hand-side constraint

Question 4

Which statement about Monte Carlo simulation is correct?

A. More replications eliminate biased input distributions

B. It estimates outcomes by repeated random sampling

C. It always finds the globally optimal decision

D. It does not require uncertainty assumptions

Correct Answer: B. It estimates outcomes by repeated random sampling

Question 5

What is the principal purpose of a warm-up period in a steady-state DES?

A. To increase the number of decision variables

B. To remove initialization effects from performance estimates

C. To guarantee an optimal solution

D. To convert continuous variables to integers

Correct Answer: B. To remove initialization effects from performance estimates

Question 6

A Pareto-efficient solution is one for which:

A. All objectives have identical numerical values

B. No objective can improve without worsening at least one other objective

C. The cost objective is always minimized alone

D. Every constraint has positive slack

Correct Answer: B. No objective can improve without worsening at least one other objective

Question 7

Which condition supports interpreting a local optimum of a convex optimization model as a global optimum?

A. Convex feasible structure and appropriate convex objective conditions

B. Use of binary variables only

C. Random initialization without validation

D. A non-convex objective with no constraints

Correct Answer: A. Convex feasible structure and appropriate convex objective conditions

Question 8

What does verification examine?

A. Whether the model is implemented according to its specification

B. Whether stakeholders prefer the result

C. Whether the market will remain unchanged

D. Whether all objectives have equal weights

Correct Answer: A. Whether the model is implemented according to its specification

Question 9

Why can a heuristic be appropriate for vehicle routing?

A. It always proves optimality faster than an LP

B. Large combinatorial problems may make exact optimization impractical

C. It removes the need for feasibility checks

D. It guarantees identical solutions across all runs

Correct Answer: B. Large combinatorial problems may make exact optimization impractical

Question 10

In a facility-location model, the constraint ∑ⱼxᵢⱼ ≤ Kᵢyᵢ primarily ensures that:

A. Customer demand is ignored when a facility opens

B. An unopened facility cannot supply orders

C. All facilities must open

D. Variable shipments equal fixed costs

Correct Answer: B. An unopened facility cannot supply orders