Optimization Under Uncertainty

Introduction and Context

Classical optimization assumes that all relevant parameters are known with sufficient precision. A firm may be modeled as choosing production quantities when prices, capacities, costs, and demand are treated as fixed. This is deterministic optimization. In practice, however, demand fluctuates, supply is disrupted, prices change, and operational data are incomplete. A solution that is optimal for one forecast may perform poorly when conditions differ.

Optimization under uncertainty incorporates imperfect knowledge into the decision model. Its purpose is not simply to identify the mathematically best action under an assumed forecast, but to design decisions that remain economically attractive when reality departs from that forecast. Four major paradigms are considered in this chapter:

• deterministic optimization;
• stochastic optimization;
• chance-constrained optimization; and
• robust optimization.

These paradigms differ in how uncertainty is represented, what information is assumed to be available, and how performance is evaluated. A sound modeling choice depends on the quality of probability estimates, the timing of decisions, the consequences of constraint violations, and the organization’s tolerance for risk.

Learning Objectives

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

1. distinguish deterministic, stochastic, chance-constrained, and robust optimization;
2. formulate scenario-based and uncertainty-set models;
3. explain here-and-now decisions, wait-and-see decisions, and recourse;
4. calculate expected value, risk measures, the value of perfect information, and the value of stochastic information;
5. interpret the trade-off between expected performance and resilience;
6. evaluate modeling assumptions, computational issues, and governance risks; and
7. select an appropriate uncertainty framework for a business analytics problem.

1. A General Decision Model

Let x denote decisions made before uncertainty is observed and let ξ denote uncertain parameters. A general problem can be written as

Here, F is the objective, G is a vector of constraints, X is the set of structural restrictions, and ξ may represent demand, prices, processing times, exchange rates, or failure events. The decision x is feasible only relative to a particular realization of ξ unless the model explicitly protects against uncertainty.

The timing of information is central. A here-and-now decision is selected before uncertainty is revealed. A recourse decision is selected after some uncertainty becomes known. If y(ξ) denotes a recourse action, a two-stage model is often expressed as

where

The vectors c and d(ξ) are first-stage and second-stage costs. Matrices A, W(ξ), and T(ξ), and vectors b and h(ξ), describe resource and balance constraints. The expectation E[·] is meaningful only when a probability distribution for ξ is specified. This formulation assumes that recourse is available and that the second-stage problem is feasible, or that infeasibility is penalized explicitly.

2. Deterministic Optimization as a Benchmark

A deterministic model replaces uncertain parameters with fixed estimates, such as expected demand or a nominal fuel price:

This model is transparent, often computationally efficient, and useful when uncertainty is small or decisions can be revised frequently. It is also a benchmark against which more sophisticated models can be compared.

Its principal weakness is parameter sensitivity. If a capacity constraint uses forecast demand rather than actual demand, the solution may be infeasible or expensive when the forecast is wrong. A deterministic model is therefore not necessarily risk-neutral; it may conceal risk inside point estimates.

3. Stochastic Optimization

Stochastic optimization represents uncertainty with a probability distribution. In a finite scenario model, scenarios s = 1,…,S have probabilities p_s satisfying

A simple two-stage stochastic linear program is

The first-stage vector x must be identical across scenarios because it is chosen before the scenario is known. The recourse vector y_s can differ by scenario. This restriction is called nonanticipativity: decisions cannot use information that was not available when they were made.

Scenario construction

A scenario should represent a coherent state of the world rather than an arbitrary combination of parameter values. Scenario generation may use:

• historical observations;
• statistical forecasting models;
• simulation;
• expert judgment;
• stress scenarios; or
• combinations of these sources.

Scenario probabilities should reflect the modeler’s beliefs and data, not merely make the scenario tree convenient. Important dependence relationships must be preserved. For example, high demand may coincide with higher prices, while a supply disruption may increase both lead time and emergency procurement cost.

Expected-value optimization and expected recourse

A model that minimizes expected cost uses

where C_s is the total cost in scenario s. Expected cost is appropriate when the organization is risk-neutral and repeated decisions permit gains and losses to average out. It may be inappropriate when a single severe loss threatens solvency, safety, service continuity, or regulatory compliance.

Recourse and flexibility

Recourse gives a model the ability to respond after uncertainty is revealed. Examples include buying emergency inventory, assigning temporary labor, rerouting shipments, or outsourcing production. Recourse is valuable only if it is operationally available, timely, and correctly priced. A model that permits unlimited cheap recourse can produce an unrealistic plan.

4. Perfect Information and the Value of Flexibility

The expected value of perfect information (EVPI) measures the maximum amount a decision maker would pay to know the scenario before selecting x. Let z_SP be the objective value of the stochastic program and z_WS be the expected value of solving each scenario after it is known:

for a minimization problem. Since advance information cannot worsen the decision, EVPI is nonnegative under consistent formulations.

The expected value of the perfect information solution is

where z_s is the optimal cost if scenario s were known in advance.

The expected value of the stochastic solution (VSS) compares a stochastic model with a policy based on the expected value of uncertain parameters. Let x_EEV be the solution obtained by replacing uncertain parameters with their expectations, and let z_M(x_EEV) be its expected cost when evaluated across scenarios. Then

For minimization, VSS is nonnegative when both models are correctly evaluated. VSS measures the benefit of using a stochastic policy rather than a single expected-value policy; EVPI measures the additional value of knowing the future perfectly. Usually,

These values support investment decisions about forecasting, sensing, data acquisition, and flexibility.

5. Chance-Constrained Optimization

A chance constraint limits the probability of violating a requirement. For example, a service-level requirement may be written as

where α is the permitted violation probability. If α = 0.05, the constraint must hold with probability at least 95% under the assumed distribution.

A common inventory formulation requires sufficient stock to meet random demand D:

where Q is the stocking quantity. If D is normally distributed with mean μ and standard deviation σ, then

where z_(1−α) is the (1−α)-quantile of the standard normal distribution. For a 95% service probability, z_0.95 ≈ 1.645. This formula assumes a valid normal approximation, known or reliably estimated μ and σ, and a service definition based on a single-period demand event.

Chance constraints are attractive because they express reliability directly. Their limitations include:

• probabilities may be poorly estimated;
• multiple chance constraints may not control the probability of joint feasibility;
• distributions may be non-normal or heavy-tailed; and
• a permitted violation may be unacceptable in safety-critical settings.

A joint requirement such as P(G_i(x,ξ) ≤ 0 for every i) ≥ 1−α is different from imposing each individual constraint at probability 1−α. The latter can produce a much lower joint reliability unless dependence is considered.

6. Robust Optimization

Robust optimization does not require precise probabilities. Instead, uncertain parameters are assumed to belong to an uncertainty set U. A robust counterpart requires feasibility for every realization in U:

For a linear constraint with uncertain coefficients,

the robust constraint controls the worst permitted coefficient combination.

Types of uncertainty sets

A box set allows each parameter to vary independently:

This set is easy to use but may be excessively conservative because it assumes all adverse deviations occur simultaneously.

An ellipsoidal set captures correlated or aggregate deviations more smoothly. A budgeted set limits the number or total magnitude of simultaneous deviations. If Γ is an uncertainty budget, Γ = 0 gives the nominal model, while larger Γ gives greater protection and usually a more conservative solution.

Robust optimization is valuable when probability estimates are unreliable, historical data are scarce, or constraint violation is costly. Its central trade-off is conservatism: a robust solution may sacrifice expected profit to reduce worst-case exposure. Robustness must therefore be calibrated rather than maximized blindly.

7. Comparing the Four Paradigms

The approaches are not mutually exclusive. A robust-stochastic model can use distributions within a set of plausible distributions. A stochastic model can include chance constraints or coherent risk measures. The correct choice depends on the decision context, not on the perceived sophistication of the technique.

8. Risk Measures and Performance–Resilience Trade-offs

Expected value alone does not describe the distribution of outcomes. For a loss L, the variance is

Variance penalizes both unusually high and unusually low losses and is not always suitable when only downside risk matters.

The Value at Risk (VaR) at level α is the α-quantile of loss:

The Conditional Value at Risk (CVaR), also called expected shortfall, is the expected loss in the worst tail beyond VaR, subject to technical treatment of discontinuous distributions:

when the loss distribution is continuous. CVaR is generally more informative than VaR about the severity of tail outcomes and is a coherent risk measure under standard conditions.

A risk-adjusted objective may be written as

where λ ≥ 0 expresses the weight placed on tail risk. Units are consistent because both expected loss and CVaR are measured in currency, while λ is dimensionless. Increasing λ usually improves resilience but may reduce expected profit.

An alternative is a constrained formulation:

where K is an acceptable tail-loss limit. This formulation is often easier to communicate to managers because K has a direct interpretation.

9. Worked Example: Inventory with Uncertain Demand

A retailer must choose inventory Q before observing demand. Demand is 100 units with probability 0.6 and 180 units with probability 0.4. Each unit purchased costs 4 currency units, each unit sold earns 10, and unsold units have salvage value 1. Assume no stockout penalty beyond lost contribution.

For demand D, profit is

Consider Q = 100:

• If D = 100, profit = 10(100) − 4(100) = 600.
• If D = 180, profit = 10(100) − 4(100) = 600.
• Expected profit = 0.6(600) + 0.4(600) = 600.

Consider Q = 180:

• If D = 100, profit = 10(100) + 1(80) − 4(180) = 380.
• If D = 180, profit = 10(180) − 4(180) = 1,080.
• Expected profit = 0.6(380) + 0.4(1,080) = 660.

Thus the expected-profit stochastic decision is Q = 180, with expected profit 660. The deterministic expected-demand model uses E[D] = 0.6(100) + 0.4(180) = 132 units. If only Q = 100 or Q = 180 are permitted, it may select 180 because expected demand exceeds 100; if continuous ordering is allowed, the objective must be evaluated carefully because the discrete demand distribution creates a piecewise-linear expected-profit function.

The Q = 180 policy has higher expected profit but loses 220 when demand is low, whereas Q = 100 earns 600 in both scenarios. This illustrates that expected performance and resilience are distinct. If the retailer has a strict downside-risk limit, a robust or risk-constrained model may prefer Q = 100.

10. Applied Case: Resilient Distribution Planning

Background

This illustrative case concerns a regional healthcare distributor serving hospitals from one central warehouse. Demand for emergency supplies is uncertain, and severe weather can interrupt the main transport corridor.

Problem

A deterministic model based on average demand minimizes distribution cost but schedules nearly all shipments through the lowest-cost corridor. Historical evidence suggests that weather disruption is infrequent but produces severe service failures. The organization must decide how much inventory to hold, whether to reserve a secondary carrier, and how much capacity to contract in advance.

Analysis

The analytics team constructs six scenarios from demand forecasts and weather states. Each scenario includes demand, transport capacity, lead time, and emergency procurement cost. Probabilities are estimated from five years of weather records and updated forecasts. The team compares:

1. a deterministic expected-demand model;
2. a two-stage stochastic model with emergency procurement recourse;
3. a chance-constrained model requiring at least 98% probability of hospital service; and
4. a robust model with a budgeted disruption set.

The stochastic model provides the lowest expected cost but occasionally violates the service target in the severe-disruption scenario. The chance-constrained model reserves additional secondary capacity. The robust model protects against a broader set of disruptions and has the highest nominal cost, but it avoids dependence on potentially inaccurate disruption probabilities.

Approach and solution

Management selects a hybrid policy: a stochastic model with a CVaR constraint and a limited robust capacity requirement. The first-stage decision reserves secondary transport and establishes minimum inventory. Recourse decisions allocate emergency shipments after demand and weather information are observed. Scenario probabilities are stress-tested, and the model is re-solved monthly.

Outcome

The resulting plan has slightly higher expected operating cost than the deterministic plan but materially lower tail losses and fewer projected service failures. The improvement does not come from a single algorithm. It comes from explicitly modeling timing, recourse, correlated disruptions, and the organization’s service obligation.

Learning points

• Rare events can dominate operational risk.
• Recourse must reflect actual contracts and lead times.
• A probability-based model should be stress-tested against probability misspecification.
• Resilience has an economic value that may not appear in average-cost comparisons.
• Governance requires documenting scenario definitions, probabilities, and approved risk limits.

11. Implementation, Technology, and Governance

Modern optimization systems combine forecasting, simulation, optimization solvers, and monitoring dashboards. Machine-learning forecasts can improve scenario distributions, but prediction accuracy does not automatically produce good decisions. A forecast may be well calibrated on average yet fail during structural breaks or rare crises.

Best practices include:

• separate training, validation, and decision-evaluation data;
• preserve temporal ordering when testing forecasts;
• use out-of-sample simulation to evaluate policies;
• document all probability and uncertainty-set assumptions;
• test sensitivity to scenario probabilities and model parameters;
• include operational constraints that reflect real contracts and capacities;
• monitor realized outcomes and recalibrate models;
• retain human review for safety-critical or high-impact decisions.

Ethically and legally, organizations should consider whether uncertainty models distribute shortages or service failures unfairly across customers, regions, or vulnerable groups. A cost-minimizing policy may systematically deprioritize low-power communities. In regulated industries, auditability, explainability, data governance, and evidence of compliance may be as important as objective value. Optimization should support accountable decision-making, not obscure responsibility behind a solver.

12. Limitations and Contemporary Developments

No uncertainty paradigm eliminates model risk. Scenario models may omit important states. Probability estimates may be nonstationary. Robust sets may be too narrow to protect against reality or so broad that decisions become uneconomical. Chance constraints can create false confidence when the distribution is misspecified. Recourse models may assume information arrives sooner than it does.

Current research increasingly addresses distributionally robust optimization, which protects against a family of probability distributions rather than one estimated distribution. A generic distributionally robust objective is

where 𝒫 is an ambiguity set of plausible distributions. The model guards against distributional misspecification while avoiding the extreme conservatism of unrestricted worst-case analysis.

Other developments include multistage stochastic programming, adaptive robust optimization, data-driven uncertainty sets, online optimization, and decision-focused learning. These methods are powerful but require careful validation, computational resources, and transparent communication of assumptions.

Chapter Summary

Optimization under uncertainty designs decisions for environments in which parameters are not known in advance. Deterministic optimization uses point estimates and is useful as a benchmark, but it may conceal sensitivity to forecast error. Stochastic optimization uses scenarios and probabilities, allowing expected performance and recourse to be modeled while enforcing nonanticipativity. Chance constraints express reliability requirements probabilistically, whereas robust optimization protects against every realization in a specified uncertainty set.

EVPI measures the value of knowing the future before acting, while VSS measures the value of using a stochastic policy instead of an expected-value policy. Risk measures such as variance, VaR, and CVaR extend analysis beyond expected performance. Robustness and resilience generally require additional cost, creating a trade-off that should be evaluated through risk limits, scenario analysis, and sensitivity testing.

Effective practice combines mathematically valid formulations with realistic information timing, credible data, operationally feasible recourse, governance, and continuous monitoring. The most sophisticated model is not automatically the best model; the appropriate model is the one whose uncertainty assumptions, risk preferences, and decision timing match the managerial problem.

Key Terms

Deterministic optimization: Optimization using fixed, known parameter values.

Uncertainty: Imperfect knowledge about future parameters or events.

Scenario: A coherent possible realization of uncertain conditions.

Scenario probability: The assigned likelihood of a scenario.

Stochastic optimization: Optimization using probability distributions or scenarios.

Here-and-now decision: A decision made before uncertainty is observed.

Recourse: An action taken after some uncertainty has been revealed.

Nonanticipativity: The requirement that decisions cannot use future information.

Chance constraint: A constraint required to hold with a specified probability.

Robust optimization: Optimization that enforces feasibility for all values in an uncertainty set.

Uncertainty set: The collection of parameter values treated as plausible.

Expected value: Probability-weighted average outcome.

EVPI: Expected value of perfect information.

VSS: Expected value of the stochastic solution.

VaR: A specified quantile of a loss distribution.

CVaR: Expected loss in the worst tail beyond a selected quantile.

Resilience: The ability to maintain or recover performance under disruption.

Risk appetite: The level and type of risk an organization is willing to accept.

Distributionally robust optimization: Optimization over a family of plausible probability distributions.

Model risk: Risk arising from incorrect assumptions, data, formulation, or implementation.

Review Questions

5 Short Answer Questions

1. What distinguishes a stochastic model from a deterministic model?
2. Define a here-and-now decision and a recourse decision.
3. Why is nonanticipativity required in scenario-based optimization?
4. What does a chance constraint control?
5. Why can a robust optimization model be conservative?

5 Long Answer Questions

1. Compare deterministic, stochastic, chance-constrained, and robust optimization in terms of assumptions, strengths, and limitations.
2. Explain EVPI and VSS, including their managerial interpretation.
3. Discuss how scenario generation can introduce model risk.
4. Explain why CVaR may be preferable to variance for disruption planning.
5. Analyze the role of information timing in a two-stage stochastic program.

3 Analytical / Application Questions

1. A retailer faces three demand scenarios with known probabilities and can order before demand is observed. Formulate a two-stage inventory model with emergency procurement recourse.
2. A service provider requires at least 97% probability of meeting demand. Explain when a chance constraint is preferable to a robust constraint and identify the data assumptions required.
3. Design a sensitivity analysis for a distribution plan that compares expected cost, CVaR, service reliability, and the effect of enlarging a budgeted uncertainty set.

Multiple Choice Questions with Answers

Question 1

Which feature distinguishes a two-stage stochastic program from a deterministic program?

A. It permits recourse decisions after uncertainty is observed.

B. It assumes all parameters are known exactly.

C. It eliminates the need for feasibility constraints.

D. It requires every decision to be made after the scenario is known.

Correct Answer: A. It permits recourse decisions after uncertainty is observed.

Question 2

In a finite scenario model, what condition must scenario probabilities satisfy?

A. Every probability must equal 0.5.

B. Their sum must equal 1 and each must be nonnegative.

C. Their sum must equal the number of scenarios.

D. They must be independent of all decisions.

Correct Answer: B. Their sum must equal 1 and each must be nonnegative.

Question 3

A chance constraint P(G(x,ξ) ≤ 0) ≥ 0.95 directly requires what?

A. The constraint to hold in every conceivable state.

B. The expected violation to equal zero.

C. The constraint to hold with probability at least 95%.

D. The uncertain parameter to be normally distributed.

Correct Answer: C. The constraint to hold with probability at least 95%.

Question 4

For a minimization problem, EVPI is generally calculated as:

A. The stochastic-program value minus the wait-and-see value.

B. The wait-and-see value minus the stochastic-program value.

C. The expected-value policy value plus the stochastic value.

D. The nominal cost multiplied by the number of scenarios.

Correct Answer: A. The stochastic-program value minus the wait-and-see value.

Question 5

If demand is normally distributed with mean μ and standard deviation σ, which quantity satisfies P(Q ≥ D) ≥ 1−α?

A. Q = μ − z_(1−α)σ

B. Q = μ + z_(1−α)σ

C. Q = αμ + σ

D. Q = μσ − z_(1−α)

Correct Answer: B. Q = μ + z_(1−α)σ

Question 6

What is the central requirement of robust optimization?

A. Optimality under the average scenario only.

B. Feasibility for every realization in the specified uncertainty set.

C. A normally distributed uncertainty variable.

D. A zero probability of all adverse events.

Correct Answer: B. Feasibility for every realization in the specified uncertainty set.

Question 7

Which measure focuses on the average loss in the worst tail of a distribution?

A. Mean

B. Variance

C. CVaR

D. Expected demand

Correct Answer: C. CVaR

Question 8

What does nonanticipativity prevent?

A. The use of recourse decisions.

B. Decisions from depending on information unavailable when they are made.

C. Scenario probabilities from being estimated.

D. Constraints from being linear.

Correct Answer: B. Decisions from depending on information unavailable when they are made.

Question 9

In the inventory example, what was the expected profit for ordering 180 units?

A. 380

B. 600

C. 660

D. 1,080

Correct Answer: C. 660

Question 10

Increasing the weight λ in E[L] + λCVaR typically does what, subject to other conditions?

A. Places greater emphasis on tail risk.

B. Removes all uncertainty from the model.

C. Guarantees maximum expected profit.

D. Makes CVaR irrelevant.

Correct Answer: A. Places greater emphasis on tail risk.