Uncertainty, Simulation and Resilient Logistics

Introduction

Logistics systems operate in environments where demand, travel time, processing capacity, asset availability, weather, regulation, labour supply and infrastructure conditions are uncertain. A transport plan that is optimal under average conditions may perform poorly when congestion increases, a supplier fails, a port closes or demand shifts unexpectedly. Graduate-level logistics management therefore requires more than deterministic routing and scheduling. It requires models that represent variability, methods that test performance under alternative futures, and strategies that preserve essential service during disruption.

This chapter examines five complementary approaches: queuing theory, discrete-event simulation, stochastic programming, robust optimization and scenario analysis. It then connects these methods to digital twins and to the managerial design of resilient logistics networks. The central argument is that resilience is not simply the possession of spare capacity. It is the capability to anticipate disruption, absorb shocks, maintain critical functions, recover efficiently and adapt after learning.

Learning Objectives

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

1. Explain how congestion and variability influence logistics performance.
2. Apply basic queuing models and interpret their assumptions.
3. Design a discrete-event simulation for a logistics process.
4. Distinguish stochastic programming, robust optimization and scenario analysis.
5. Evaluate continuity, redundancy, flexibility, recovery and resilience strategies.
6. Explain how digital twins support operational and strategic decision-making.
7. Critically assess model validity, data requirements, governance and ethical risks.

1. Uncertainty and the Logistics System

Uncertainty can be classified according to its source and its managerial treatment.

A useful distinction is between aleatory uncertainty, arising from inherent variability, and epistemic uncertainty, arising from incomplete knowledge. Repeated observations can estimate a distribution for loading time, but they cannot eliminate uncertainty caused by an unprecedented regulatory closure. The distinction matters because statistical forecasting is more suitable for aleatory uncertainty, whereas stress testing, expert judgment and robust decisions are often more suitable for epistemic uncertainty.

Operational performance is commonly evaluated through service level, throughput, cycle time, lateness, utilization, cost, emissions and recovery time. These measures may conflict. Increasing utilization can reduce idle resources but may sharply increase waiting time. Minimizing transport cost may increase exposure to a single corridor. A resilient design therefore optimizes a portfolio of outcomes rather than a single average cost.

2. Congestion and Queuing Theory

2.1 The basic structure of a queue

A queuing system consists of arriving entities, a waiting line, one or more servers, a service discipline and a departure process. In logistics, entities may be trucks, containers, orders, pallets or customers; servers may be gates, docks, cranes, inspection stations or vehicles.

The principal variables are:

• λ: mean arrival rate, measured in entities per unit of time;
• μ: mean service rate of one server, measured in entities per unit of time;
• c: number of parallel servers;
• ρ: utilization, or the proportion of capacity being used;
• W: mean time in the system, including waiting and service;
• W_q: mean waiting time before service;
• L: mean number in the system;
• L_q: mean number waiting.

For a stable system with identical parallel servers, a necessary utilization condition is:

This condition is necessary but not sufficient for acceptable service. As ρ approaches 1, waiting time can increase disproportionately because random arrivals and service-time variation create temporary demand peaks.

2.2 Little’s Law

Little’s Law is one of the most useful relationships in operations management:

Similarly, for the waiting portion:

The variables must use consistent time units. Little’s Law applies to a stable system over a sufficiently long observation period, including systems with arbitrary arrival and service distributions. It does not require Poisson arrivals or exponential service times. It relates long-run averages; it does not by itself describe the distribution of waiting times or guarantee that the system is stable.

Worked example. A cross-dock processes an average of 18 trucks per hour. The average time a truck spends in the gate-and-dock system is 20 minutes, or 1/3 hour. By Little’s Law:

The system contains six trucks on average, including those waiting and being processed. If managers observe an average of nine trucks, the implied average time is:

2.3 The M/M/1 model

The M/M/1 queue assumes Poisson arrivals, exponentially distributed service times, one server, first-come-first-served discipline and an infinite waiting space. Its utilization is ρ = λ/μ, with stability requiring ρ < 1. The principal results are:

Here λ and μ must be expressed in the same units. The model is analytically convenient but often unrealistic for logistics because service times may be scheduled, batch-based, correlated or affected by shipment complexity. It should therefore be treated as a benchmark rather than an unquestioned representation.

Worked example. A weighbridge receives 12 trucks per hour and processes 15 trucks per hour. Thus ρ = 12/15 = 0.8. The expected number waiting is:

The expected waiting time is:

The result illustrates congestion sensitivity: increasing utilization from 0.8 to 0.9 does not merely increase workload by 12.5%; under the model, waiting increases substantially.

2.4 Variability and capacity decisions

A practical approximation for a single-server queue with general interarrival and service distributions is Kingman’s formula:

where E[S] is mean service time, c_a is the coefficient of variation of interarrival times, c_s is the coefficient of variation of service times, and c² denotes variance divided by the squared mean. The approximation assumes a stable, work-conserving, single-server queue and is most useful for insight rather than exact prediction.

The formula shows three managerial levers: reduce utilization, reduce arrival variability and reduce service variability. Appointment systems, pre-clearance, standardized loading, workload smoothing and reliable information can be as valuable as adding physical capacity.

3. Discrete-Event Simulation

3.1 Purpose and logic

Discrete-event simulation (DES) represents a system as a sequence of events occurring at particular points in time. Typical events include arrival, queue entry, service start, service completion, vehicle breakdown, replenishment and disruption onset. Between events, the state remains unchanged or changes according to a specified rule.

DES is valuable when a system contains interacting queues, finite buffers, priorities, shift calendars, batch operations, resource constraints or complex disruption logic that is difficult to capture analytically. It estimates performance through repeated replications rather than producing a closed-form solution.

A basic simulation clock advances to the time of the next scheduled event. The model then updates the system state, records performance measures and schedules any resulting events. A logistics model may include entities such as orders and trucks, resources such as docks and drivers, and rules such as priority for urgent medical shipments.

3.2 Building a credible simulation

A rigorous DES study usually follows these steps:

1. Define the decision and boundary. Specify whether the question concerns dock capacity, fleet size, appointment policy or emergency recovery.
2. Map the process. Identify arrivals, activities, queues, resources, routing logic and constraints.
3. Collect and clean data. Estimate arrival patterns, service-time distributions, travel times, failures and repair durations.
4. Select distributions cautiously. Use empirical distributions when data are sufficient; do not impose normality or exponential assumptions without evidence.
5. Implement the model. Encode events, priorities, calendars, capacities and random-number streams.
6. Verify. Check whether the code implements the conceptual model correctly.
7. Validate. Compare outputs with observed system behaviour and consult process experts.
8. Experiment. Test policies, capacity levels and disruption scenarios across multiple replications.
9. Analyze uncertainty. Report confidence intervals, not only point estimates.
10. Communicate limitations. State which conclusions are robust and which depend on assumptions.

For a performance estimate Y from n independent replications, the sample mean is:

If the replication results have sample standard deviation s, an approximate 95% confidence interval for the mean is:

The t critical value is appropriate when the replication observations are independent and the sample mean approximation is suitable. More replications reduce sampling error, but they do not correct a structurally invalid model.

3.3 Warm-up, initialization and experimental design

A terminating model, such as a one-day emergency operation, may begin empty and end at a defined time. A steady-state model, such as a continuously operating distribution centre, can be biased by its empty initial condition. A warm-up period allows transient effects to dissipate; its length should be assessed using plots, batch means or other statistical diagnostics rather than selected arbitrarily.

Common experimental designs include:

• scenario comparison: compare existing and proposed policies;
• factorial experiments: estimate the effects and interactions of several decision factors;
• common random numbers: use matched random streams to reduce comparison noise;
• variance reduction: improve precision without proportionally increasing replications;
• sensitivity analysis: vary uncertain assumptions to test conclusion stability.

Simulation should not be used to create false precision. If input data are poor, outputs should be expressed as ranges, conditional findings or confidence intervals.

4. Stochastic Programming and Robust Optimization

4.1 Stochastic programming

Stochastic programming models decisions made before and after uncertainty is revealed. Let x represent first-stage decisions, y_s represent recourse decisions in scenario s, p_s the probability of scenario s, c the first-stage cost vector and q_s the recourse cost vector. A two-stage formulation is:

subject to:

The first constraints describe decisions that must be made before the scenario is known, such as opening facilities or reserving capacity. The second set describes corrective actions after demand or disruption is observed, such as reallocating inventory or hiring emergency transport.

A crucial modelling principle is non-anticipativity: first-stage decisions cannot depend on information that is not yet available. A model that allows a separate first-stage decision for each future scenario incorrectly assumes perfect foresight.

Expected cost can be insufficient where severe service failures are unacceptable. A risk-averse formulation may add a penalty for tail outcomes, such as conditional value at risk (CVaR), or impose service constraints in selected scenarios.

4.2 Robust optimization

Robust optimization does not require reliable probabilities for every future state. It seeks a decision that performs acceptably for all values in a defined uncertainty set. A generic robust problem is:

where x is the decision, u is the uncertain parameter and U is the uncertainty set. For example, demand may be assumed to lie within a range, or travel time may be allowed to increase by a specified percentage.

Robustness has a price. A highly conservative solution may use excess inventory, capacity or suppliers and therefore cost more under normal conditions. The uncertainty set must therefore be calibrated. If it is too narrow, the solution is fragile; if too broad, the solution becomes unnecessarily expensive or infeasible.

4.3 Comparing decision approaches

These methods are complementary. Optimization may identify a policy, simulation may test its operational feasibility, and scenario analysis may challenge its assumptions.

5. Scenario Analysis and Stress Testing

Scenario analysis constructs internally consistent descriptions of alternative futures. It differs from forecasting: the purpose is not to predict one most likely future but to explore how decisions perform under materially different conditions.

A disciplined process includes:

1. Define the focal decision and time horizon.
2. Identify driving forces, such as demand growth, fuel prices, climate exposure, regulation and geopolitical risk.
3. Separate predetermined elements from critical uncertainties.
4. Build a small number of distinctive, plausible scenarios.
5. Quantify operational implications where possible.
6. Test current policies and identify trigger points.
7. Develop no-regret actions, contingent actions and signposts.

A useful stress test may combine several adverse conditions: a 30% demand surge, a 20% travel-time increase and loss of a major supplier. Such combinations should be justified rather than selected solely to produce dramatic results. The purpose is to reveal vulnerabilities, not to manufacture an implausible catastrophe.

6. Resilient Logistics Design

Resilience has four connected capabilities:

• anticipation: sensing threats and preparing options;
• absorption: maintaining essential operations during a shock;
• recovery: restoring performance within an acceptable time;
• adaptation: changing processes or network design after learning.

6.1 Continuity

Business continuity planning identifies critical logistics functions, dependencies, maximum tolerable disruption and recovery priorities. A continuity plan should specify:

• critical products, customers and services;
• minimum operating capacity;
• alternative sites, routes, carriers and suppliers;
• communication and decision authority;
• data backup and manual workarounds;
• recovery time and recovery point objectives;
• testing, training and review procedures.

Continuity is not equivalent to maintaining every activity. During a severe disruption, the objective may be to preserve life-critical, contractual or strategically essential flows while temporarily reducing low-priority service.

6.2 Redundancy, flexibility and visibility

Redundancy provides additional resources or pathways: safety stock, spare vehicles, multiple suppliers, parallel facilities and alternative ports. It improves robustness but can increase cost, emissions and complexity.

Flexibility allows resources to be reconfigured. Examples include cross-trained workers, multi-purpose facilities, interchangeable packaging, postponement and contracts with flexible carrier capacity.

Visibility provides timely information about inventory, location, condition and capacity. Visibility does not itself create resilience, but it improves detection and enables earlier intervention. Excessive data without decision rights can create noise rather than capability.

Modularity separates network components so that failure in one part does not propagate through the entire system. Standard containers, interoperable information systems and regional fulfilment nodes can support modularity.

6.3 Recovery and resilience metrics

Useful measures include time to detect, time to respond, time to recover, service loss, backlog clearance, fill rate during disruption and financial loss. A simple resilience curve represents performance P(t) before, during and after disruption. Cumulative performance loss over a horizon [t₀,t₁] can be represented as:

where P₀ is the target or baseline performance. The integral has units of performance multiplied by time; for example, percentage-points of service lost over days. The measure is meaningful only when P(t) is defined consistently and the horizon includes comparable recovery periods.

Recovery time alone can be misleading. A network that experiences a small service reduction for a long period may be preferable to one that fails completely and recovers quickly. Managers should therefore examine both depth and duration of performance loss.

7. Digital Twins for Uncertain Logistics

A digital twin is a continuously or periodically updated digital representation of a physical logistics system, linked through data and used for monitoring, prediction, experimentation or control. It may represent a warehouse, fleet, port, distribution network or end-to-end supply chain.

A mature digital twin typically integrates:

1. physical assets and processes;
2. sensors, enterprise systems, telematics and external data;
3. a state-estimation layer that reconciles inconsistent observations;
4. simulation, optimization and predictive analytics;
5. decision interfaces and authorized control actions.

The twin can test a new dock schedule, estimate the effect of a road closure, predict maintenance needs or compare inventory policies. It is especially powerful when simulation is connected to live conditions, but it should not be confused with a dashboard. A dashboard reports information; a twin represents system behaviour and supports counterfactual analysis.

Important governance issues include data ownership, cyber security, model drift, access control, auditability and human override. An algorithm that reroutes vehicles automatically may optimize cost while violating driver safety rules, delivery commitments or local restrictions. Digital-twin deployment therefore requires clear accountability and controls for automated decisions.

8. Illustrative Case: Resilience at NorthRiver Medical Distribution

Background

NorthRiver Medical Distribution is an illustrative case involving a regional wholesaler serving hospitals from one central distribution centre. The centre processes 240 orders per hour during the evening peak. Orders arrive unevenly, and picking time varies according to product mix. A single outbound sortation line creates a queue before dispatch. The firm has also experienced road closures during severe storms.

Problem

Managers proposed increasing average utilization by reducing scheduled labour and using the existing sortation line more intensively. However, hospitals reported late deliveries during peak periods. The firm needed to decide whether to add a second sortation module, cross-train staff, hold regional emergency inventory or contract an alternate carrier.

Analysis

Initial observations showed an average arrival rate of 240 orders per hour and a mean processing capacity of 300 orders per hour. A simplified M/M/1 benchmark gives ρ = 0.8 and:

This appears acceptable, but the benchmark understates reality because arrivals are batchy, service times are not exponential, urgent orders receive priority and the line stops during replenishment. A DES was therefore constructed with empirical order arrivals, product-specific pick times, shift calendars, priority rules and storm-related transport failures.

The model was verified against process maps and validated against two weeks of observed queue lengths and dispatch lateness. Five hundred replications were run for each policy, with separate random streams and a warm-up period for the continuously operating process. The policies were:

• Policy A: current staffing and one sortation line;
• Policy B: additional peak-period labour;
• Policy C: second sortation module;
• Policy D: cross-trained staff plus emergency inventory at a regional partner;
• Policy E: Policy D plus an alternate carrier activated when storm indicators exceed a threshold.

The simulation measured hospital order fill rate, 95th-percentile dispatch delay, normal operating cost and disruption-period service. Policy C performed well in normal peaks but required high capital expenditure. Policy D produced better flexibility at lower capital cost but depended on reliable inventory visibility. Policy E had the strongest disruption performance but increased contract and monitoring costs.

Approach and solution

NorthRiver adopted a layered design. It cross-trained employees, established a small regional emergency inventory, negotiated an alternate carrier and retained a staged option to add sortation capacity. A storm trigger was defined using weather alerts and road-closure information. Activation authority was assigned to the control-tower manager, while decisions affecting clinical priorities required an operations-and-clinical governance group.

The company also implemented a digital-twin pilot that combined order status, dock capacity, carrier position and weather data. The twin did not autonomously change hospital priorities; it generated recommended actions for an authorized manager.

Outcome and learning points

The case demonstrates why average utilization alone is inadequate. A queue can be mathematically stable yet operationally unacceptable when variability and priority interactions are ignored. It also shows that resilience is often achieved through a portfolio: modest redundancy, flexible labour, information quality, alternative transport and staged investment. Finally, the case illustrates the value of combining methods: analytical queuing exposed congestion sensitivity, simulation represented operational detail, scenarios tested storms and governance controls protected clinical priorities.

9. Managerial, Ethical and Regulatory Considerations

Resilience investments create distributional questions. A company may protect high-value customers while leaving vulnerable communities with reduced service. Prioritization rules should therefore be explicit, defensible and consistent with contractual, safety and public-interest obligations.

Data-driven logistics raises privacy concerns when worker or driver location is continuously monitored. Data minimization, purpose limitation, retention controls and access restrictions should be applied. Cyber security is a resilience requirement: a connected fleet or digital twin can become a pathway for disruption if credentials, interfaces and recovery procedures are weak.

Regulation may govern hazardous materials, driver hours, accessibility, medical products, environmental emissions and automated decision-making. Models must include regulatory constraints rather than treating them as post-processing checks. Sustainability also matters. Redundancy through duplicate facilities or emergency air freight can improve continuity while increasing emissions. Multi-objective evaluation should consider service, cost, risk, worker welfare and environmental impact.

10. Best Practices and Limitations

Best practice begins with decision clarity. A sophisticated model cannot compensate for an undefined problem or poor data. Managers should:

• model the mechanism of congestion rather than only its average;
• use empirical data and document assumptions;
• validate inputs and outputs with domain experts;
• combine expected performance with tail-risk measures;
• test policies across scenarios, not only one forecast;
• distinguish operational flexibility from unused capacity;
• design recovery procedures before disruption occurs;
• assign decision rights and escalation thresholds;
• use pilots and staged investment for uncertain technologies;
• report uncertainty and sensitivity transparently.

Limitations remain important. Queuing models may omit correlations and finite buffers. Simulation can be expensive, opaque and sensitive to input distributions. Stochastic programming depends on scenario probabilities and may become computationally large. Robust optimization can be excessively conservative. Scenario analysis can reflect managerial bias. Digital twins can reproduce inaccurate processes at high speed. None of these methods eliminates uncertainty; they improve the quality and transparency of decisions under uncertainty.

Chapter Summary

Uncertainty is a defining feature of logistics, arising from demand, processes, supply, travel, disruption and incomplete knowledge. Queuing theory explains how utilization and variability generate congestion; Little’s Law links average flow, time and inventory, while M/M/1 and Kingman-style relationships provide useful but conditional benchmarks. Discrete-event simulation represents interacting queues, priorities, calendars and disruptions and requires careful verification, validation, warm-up treatment and replication analysis.

Stochastic programming separates pre-uncertainty decisions from recourse decisions and requires non-anticipativity. Robust optimization protects against specified uncertainty sets when probabilities are unreliable, although it may impose additional cost. Scenario analysis and stress testing explore coherent alternative futures and identify triggers, vulnerabilities and no-regret actions.

Resilient logistics combines anticipation, absorption, recovery and adaptation. Continuity plans, redundancy, flexibility, visibility, modularity and alternative pathways should be evaluated as a portfolio. Digital twins connect operational data with simulation and optimization but require governance, cyber security and human accountability. Effective resilience is not the cheapest design under average conditions; it is a defensible design that preserves critical service across plausible and adverse futures.

Key Terms

Uncertainty: Lack of complete knowledge about future conditions or system behaviour.

Aleatory uncertainty: Inherent random variability in a process or environment.

Epistemic uncertainty: Uncertainty caused by incomplete knowledge or an imperfect model.

Queue: A waiting line formed when demand for a resource exceeds immediate capacity.

Utilization: The proportion of available service capacity being used.

Little’s Law: The relationship L = λW between average number, flow rate and time in a stable system.

M/M/1 queue: A single-server queue with Poisson arrivals and exponential service times.

Coefficient of variation: Standard deviation divided by the mean.

Discrete-event simulation: A model that represents system changes as events occurring over time.

Verification: Checking whether a model has been implemented according to its design.

Validation: Assessing whether a model adequately represents the real system for its purpose.

Stochastic programming: Optimization with uncertain parameters represented through scenarios and probabilities.

Recourse: Corrective action taken after uncertainty is revealed.

Non-anticipativity: The requirement that decisions cannot use information unavailable when they are made.

Robust optimization: Optimization designed to perform acceptably across a defined uncertainty set.

Scenario analysis: Exploration of alternative coherent future conditions.

Resilience: The capability to anticipate, absorb, recover from and adapt to disruption.

Redundancy: Additional resources or pathways that provide alternatives when one fails.

Digital twin: A data-linked digital representation used to monitor, analyze or experiment with a physical system.

Recovery time objective: The target time within which a critical function should be restored.

Review Questions

Short Answer Questions

1. Why can a logistics system with utilization below 100% still experience severe congestion?
2. State Little’s Law and explain the units of its variables.
3. What is the difference between verification and validation in simulation?
4. Define non-anticipativity in stochastic programming.
5. Distinguish redundancy from flexibility in resilient logistics.

Long Answer Questions

1. Explain the assumptions and managerial uses of the M/M/1 queue. What limitations arise when it is applied to a distribution centre?
2. Describe a rigorous process for developing and validating a discrete-event simulation of a transport terminal.
3. Compare stochastic programming and robust optimization for network design under uncertain demand and disruption.
4. Develop a resilience framework that integrates continuity, visibility, redundancy, recovery and adaptation.
5. Critically evaluate the role of digital twins in logistics decision-making, including data, cyber security and accountability issues.

Analytical / Application Questions

1. A service station receives 20 vehicles per hour and can serve 25 vehicles per hour. Using the M/M/1 model, calculate utilization, expected number waiting and expected waiting time. Interpret the result and identify assumptions that should be tested before implementation.
2. Design a discrete-event simulation experiment to decide whether a distribution centre should add a second loading dock. Specify entities, events, inputs, performance measures, replications and validation procedures.
3. A firm faces uncertain demand, a possible port closure and unreliable probability estimates. Recommend a combined use of stochastic programming, robust optimization, simulation and scenario analysis. Justify which decisions should be strategic, contingent or operational.

Multiple Choice Questions with Answers

Question 1

Which condition is necessary for stability in a parallel-server queue?

A. λ = cμ

B. λ > cμ

C. λ < cμ

D. μ = 0

Correct Answer: C. λ < cμ

Question 2

A system has an average flow of 30 orders per hour and an average time in system of 0.2 hours. According to Little’s Law, what is the average number of orders in the system?

A. 6 orders

B. 15 orders

C. 30 orders

D. 150 orders

Correct Answer: A. 6 orders

Question 3

Which feature most clearly distinguishes discrete-event simulation from a simple average-based calculation?

A. It ignores time entirely

B. It represents state changes at specified events

C. It always produces an exact optimum

D. It requires no input data

Correct Answer: B. It represents state changes at specified events

Question 4

In a two-stage stochastic program, what does non-anticipativity require?

A. All scenarios must have equal probability

B. Recourse decisions must be made before uncertainty occurs

C. First-stage decisions cannot depend on unrevealed scenario information

D. Every scenario must have identical costs

Correct Answer: C. First-stage decisions cannot depend on unrevealed scenario information

Question 5

Robust optimization is particularly appropriate when:

A. All future probabilities are known precisely

B. Uncertainty can be represented by a credible set of possible values

C. There is no uncertainty in the system

D. Only historical averages matter

Correct Answer: B. Uncertainty can be represented by a credible set of possible values

Question 6

Which action primarily reduces service-time variability?

A. Standardizing loading procedures

B. Adding an unrelated supplier

C. Increasing product variety without training

D. Removing all performance measurements

Correct Answer: A. Standardizing loading procedures

Question 7

What is the main purpose of a simulation warm-up period in a steady-state model?

A. To increase the arrival rate

B. To remove the influence of the artificial initial condition

C. To guarantee zero waiting time

D. To replace model validation

Correct Answer: B. To remove the influence of the artificial initial condition

Question 8

Which statement best describes resilience?

A. Minimizing normal-period cost only

B. Maintaining, recovering and adapting critical performance under disruption

C. Eliminating every possible risk

D. Operating permanently at maximum utilization

Correct Answer: B. Maintaining, recovering and adapting critical performance under disruption

Question 9

A digital twin differs from a basic dashboard because it:

A. Only displays historical financial data

B. Represents system behaviour and can support experiments or predictions

C. Cannot use live operational data

D. Automatically removes the need for human governance

Correct Answer: B. Represents system behaviour and can support experiments or predictions

Question 10

Which measure captures both the magnitude and duration of performance degradation?

A. Cumulative area between target performance and actual performance over time

B. Number of suppliers only

C. Average purchase price only

D. Maximum warehouse size only

Correct Answer: A. Cumulative area between target performance and actual performance over time