Optimization Foundations and Decision Modeling

Introduction

Mathematical optimization provides a disciplined way to convert managerial and operational choices into models that can be analyzed computationally. In business analytics, optimization is used to allocate scarce resources, schedule activities, design supply chains, set prices, select investments, manage inventories, and configure digital services.

An optimization model does not automatically produce a good decision. It produces the best decision according to a specified objective, subject to stated assumptions and restrictions. Consequently, the quality of an optimization study depends as much on model formulation and interpretation as on the solution algorithm.

This chapter develops the vocabulary and logic required to formulate optimization models. It distinguishes decisions from data, objectives from constraints, and feasible solutions from optimal solutions. It also introduces problem classifications and explains how to diagnose infeasibility, unboundedness, and misleading assumptions.

Learning Objectives

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

1. Define decision variables, parameters, objective functions, constraints, and feasible regions.
2. Translate a business situation into a mathematical optimization model.
3. Classify optimization problems according to structure and variable type.
4. Interpret feasible, optimal, infeasible, and unbounded models.
5. Identify assumptions, data requirements, and governance risks in decision models.
6. Communicate optimization results responsibly to managerial decision-makers.

1. The Logic of a Decision Model

A decision model represents a system in a simplified, purposeful form. It contains five central components.

Decision variables

A decision variable represents a choice controlled by the decision-maker. If a firm decides how many units of products A and B to manufacture, the variables may be x_A and x_B, measured in units per planning period.

Variables must be defined precisely. “Production” is inadequate unless the model specifies product, location, time period, and unit of measurement. A variable may be continuous, integer, binary, or otherwise restricted.

Parameters

Parameters are known or estimated inputs. Examples include selling price, unit cost, machine hours, demand, capacity, probability, and emissions per unit. Parameters are not decisions within the model, although they may be uncertain or subject to scenario analysis.

A useful distinction is:

Objective function

The objective function quantifies the performance criterion. A maximization objective seeks the largest value, whereas a minimization objective seeks the smallest value. Common objectives include profit, revenue, service level, cost, risk, delivery time, energy use, and emissions.

For a linear profit model with products A and B:

Here, Z is total profit, p_A and p_B are profit per unit, and x_A and x_B are production quantities. The equation is valid when unit profits are constant and total profit is additive.

Constraints

Constraints express limitations, requirements, or logical conditions. A machine-hour constraint might be:

where a_A and a_B are machine hours per unit and H is available machine time. Units are consistent: units produced × hours per unit = hours, so the left side can be compared with hours available.

Constraints may be:

• Capacity constraints: resources cannot exceed availability.
• Demand constraints: sales or production cannot exceed demand.
• Balance constraints: inflow plus production equals outflow plus inventory.
• Policy constraints: minimum staffing, contractual requirements, or service standards.
• Logical constraints: a facility must open before it can serve a market.
• Domain constraints: variables must be nonnegative, integer, or binary.

Feasible region

The feasible region is the set of all assignments to the decision variables that satisfy every constraint and domain restriction. A feasible solution is any point in this set. An optimal solution is a feasible solution with the best objective value.

Optimization is therefore not merely the search for a high objective value. A decision that violates a legal, resource, or operational restriction is not a candidate solution, regardless of its apparent profitability.

2. A General Mathematical Form

A broad optimization model can be written as:

The vector x contains decision variables, θ contains parameters, f is the objective function, g_i are inequality constraints, h_j are equality constraints, and X specifies domain restrictions such as nonnegativity or integrality. This formulation accommodates maximization as well as minimization; maximizing f is equivalent to minimizing −f.

A common linear programming form is:

The vector c contains objective coefficients, A is the constraint coefficient matrix, and b contains resource limits. The notation cᵀx means the sum of coefficient–variable products. Linear programming requires linear objective and constraint expressions, continuous variables unless otherwise specified, and parameters treated as fixed for the solve.

3. Translating Business Language into Models

Model formulation is a translation activity. Business statements must be converted into variables, coefficients, equations, inequalities, and domains.

A reliable formulation process is:

1. Define the decision context, time horizon, scope, and objective.
2. Identify controllable decisions and define variables with units and indices.
3. Collect parameters and document their sources.
4. Express the objective mathematically.
5. Express physical, commercial, policy, and logical restrictions.
6. Add domain restrictions.
7. Test the model with simple cases and extreme values.
8. Review assumptions with subject-matter experts.

Worked example: product mix

A bakery produces loaves and pastries. Each loaf earns £4 contribution and requires 2 minutes of oven time. Each pastry earns £3 contribution and requires 1 minute. The bakery has 240 oven minutes available. Demand is limited to 80 loaves and 120 pastries. Let x_L be loaves and x_P be pastries produced per day.

The model is:

The objective has units of pounds per day. The capacity constraint has units of minutes. Demand constraints prevent production from exceeding market demand.

Consider the plan (x_L, x_P) = (60, 120). Its oven use is:

and its contribution is:

It is feasible because all restrictions are satisfied. By contrast, (80, 120) would earn £680 but require 280 minutes, so it is infeasible. The example illustrates why the objective alone cannot determine an acceptable business decision.

4. Classifying Optimization Problems

Classification helps analysts choose appropriate algorithms and interpret results.

Linear and nonlinear optimization

In a linear model, variables appear only to the first power and are not multiplied by one another. A term such as 5x₁ − 2x₂ is linear. Terms such as x₁², x₁x₂, or log(x₁) are nonlinear.

Nonlinear models can represent economies of scale, congestion, utility, risk, and diminishing returns more realistically, but they may be harder to solve and may have multiple local optima.

Continuous, integer, and mixed-integer models

Continuous variables can take values such as 2.35. Integer variables must take whole-number values, which is important for trucks, employees, facilities, or products. Binary variables take values 0 or 1 and often represent yes/no decisions.

A mixed-integer linear program (MILP) contains both continuous and integer variables. For example, y_j = 1 may indicate whether facility j is opened, while x_ij represents a continuous shipment quantity.

Deterministic and stochastic optimization

A deterministic model treats parameters as known fixed values. A stochastic model represents uncertainty explicitly, often through scenarios, probability distributions, or random variables. Deterministic models are simpler and useful when uncertainty is modest or decisions are frequently revised. Stochastic models may better represent uncertain demand or prices but require probability estimates and additional computational structure.

Single-objective and multi-objective optimization

A single-objective model optimizes one criterion. Real decisions frequently involve trade-offs, such as cost versus service or profit versus emissions. A weighted objective can be written as:

where P, C, and E represent profit, cost, and emissions, while nonnegative weights α, β, and γ express priorities. This approach requires careful scaling because pounds, service points, and tonnes are not naturally comparable. Alternative approaches include goal programming, constraint-based targets, and Pareto analysis.

Network, dynamic, and simulation-optimization models

Network models represent flows through nodes and arcs, such as transportation or logistics systems. Dynamic models explicitly connect decisions across time. Simulation-optimization embeds an optimization search within a simulation when system behavior is too complex for exact algebraic representation. Each class creates different data, computational, and validation requirements.

5. Geometry and Interpretation of Solutions

For two continuous variables, each linear inequality defines a half-plane. Their intersection is the feasible region, usually a convex polygon or an unbounded convex set. In a linear program, if an optimum exists, at least one optimum occurs at an extreme point (corner) of the feasible region. This property underlies the simplex method, although modern solvers also use interior-point and hybrid methods.

The geometric interpretation is useful even for larger models. A binding constraint is satisfied as an equality at a solution; it limits further improvement in the relevant direction. A nonbinding constraint has unused capacity, often called slack. For a constraint aᵀx ≤ b, slack is:

The slack has the same units as b and must be nonnegative for feasibility.

A shadow price, or dual value, estimates the marginal change in the optimal objective associated with a one-unit change in a right-hand-side parameter, within the range where the current basis or solution structure remains valid. It is not automatically a causal market price and should not be interpreted outside its sensitivity range.

6. Diagnosing Model Pathologies

Infeasibility

A model is infeasible when no assignment satisfies all constraints. Typical causes include contradictory targets, incorrect units, excessive minimum requirements, or an overlooked operational option.

For example:

has no feasible value of x. In practice, analysts can use an irreducible infeasible subsystem, feasibility relaxation, or systematic constraint removal to identify the conflict. A relaxation should be treated as a diagnostic device unless management has explicitly approved the resulting violations and penalties.

Unboundedness

A maximization model is unbounded if the objective can increase indefinitely while all constraints remain satisfied. For example:

has no finite optimum. Unboundedness often indicates a missing demand limit, capacity restriction, upper bound, or sign condition. It may also reveal that a revenue term was included without its corresponding cost or physical limitation.

Multiple optima

More than one feasible solution may achieve the same optimal objective. This is not an error. It can provide managerial flexibility, although secondary criteria may be needed to select among alternatives.

Numerical and scaling problems

Poorly scaled coefficients can make computation less stable. A model combining values measured in millions with values measured in tiny fractions may benefit from sensible unit conversions. Rounding solver outputs can also create apparent constraint violations; feasibility should be checked using unrounded values and appropriate tolerances.

7. Assumptions, Validation, and Sensitivity

Every optimization model is conditional: “optimal” means optimal under the model’s structure, data, and assumptions. Important assumptions include:

• demand is represented accurately;
• coefficients are constant over the planning period;
• resources are divisible or indivisible as modeled;
• all relevant constraints have been included;
• objective weights reflect legitimate priorities;
• data are comparable in time, geography, and units;
• decisions can actually be implemented;
• uncertainty is either negligible, updated, or modeled explicitly.

Validation should occur at several levels. Conceptual validation asks whether the model represents the real decision. Data validation checks definitions, units, missing values, outliers, and provenance. Mathematical validation checks equations, signs, domains, and limiting cases. Operational validation compares recommendations with expert judgment and historical outcomes. Stress testing evaluates the solution under adverse demand, capacity, or cost scenarios.

Sensitivity analysis changes selected parameters and observes effects on the solution and objective. It helps distinguish robust recommendations from decisions that depend on a narrow estimate. Scenario analysis is especially valuable when uncertainty is structural, such as a new regulation or supplier failure.

8. Applied Case: Regional Distribution Design

The following is an illustrative case, not a report of a specific company.

Background and problem

A retailer serves three regions from two potential distribution centres. Management wants to minimize weekly transportation and fixed operating cost while meeting regional demand. Opening a centre incurs a fixed cost; shipments incur variable costs. Each centre has a capacity.

Let:

• y_j = 1 if centre j is opened, and 0 otherwise;
• x_ij = units shipped from centre j to region i;
• d_i = demand in region i;
• K_j = capacity of centre j;
• c_ij = variable shipping cost per unit;
• F_j = fixed cost of opening centre j.

Model

The capacity constraint links the binary opening decision to shipments. If y_j = 0, capacity available from that centre is zero. If y_j = 1, shipments may use up to K_j.

Analysis and implementation

The analyst first verifies that total available capacity can meet total demand. If not, the model is infeasible unless unmet demand, outsourcing, or demand reduction is represented. Next, shipping costs and fixed costs are checked for consistent currency and time units. The model is solved as a MILP, and the selected centres and shipment quantities are reviewed with logistics managers.

Suppose the solution opens centre 1 and assigns shipments so that its capacity is nearly full, while centre 2 remains closed. This result is not accepted solely because a solver produced it. The team checks road access, labour availability, service-time feasibility, disruption risk, and whether demand forecasts are credible. If a single-centre solution is highly sensitive to a modest demand increase, a two-centre plan may be preferred despite a slightly higher expected cost.

Outcome and learning points

The main outcome is a transparent trade-off between fixed facility cost, variable transport cost, and capacity. The case demonstrates the value of binary variables, linking constraints, unit discipline, and scenario analysis. It also shows that mathematical optimality is one input to a governance process, not a substitute for operational judgment.

9. Governance, Ethics, and Contemporary Practice

Optimization systems can distribute opportunities and burdens. A staffing model may affect employee well-being; a pricing model may affect access; a routing model may shift pollution between communities. Analysts should identify affected stakeholders, document objectives and exclusions, and test for disparate impacts where decisions concern people or protected groups.

Governance practices include version-controlled model files, data lineage, documented assumptions, approval of objective weights, audit logs, reproducible solver settings, and clear separation between recommendation and final authority. Privacy-preserving data practices are required when individual-level information informs demand, credit, workforce, or customer decisions.

Modern solvers increasingly support cloud execution, automated model generation, explainable sensitivity reports, and integration with forecasting and machine learning. Machine learning may estimate parameters such as demand, while optimization chooses actions. This separation is important: predictive accuracy does not guarantee a feasible, fair, or strategically appropriate decision. Human review remains necessary, especially when the model affects rights, safety, essential services, or legally regulated outcomes.

Best-Practice Checklist

Before approving an optimization model, ask:

1. Are variables, units, time horizon, and scope explicit?
2. Does every objective term represent a legitimate organizational priority?
3. Are all material constraints and logical conditions included?
4. Are data sources, estimation methods, and uncertainty documented?
5. Has feasibility been tested under normal and extreme scenarios?
6. Have infeasibility, unboundedness, and numerical warnings been investigated?
7. Can the recommended decisions be implemented operationally?
8. Are affected stakeholders, fairness, privacy, and accountability addressed?
9. Is sensitivity information communicated with the solution?
10. Is there a monitoring and model-update process?

Chapter Summary

Optimization models formalize decisions by combining decision variables, parameters, objectives, constraints, and domain restrictions. The feasible region contains all allowable decisions; the optimal solution is the feasible decision with the best modeled objective. Linear, nonlinear, integer, mixed-integer, deterministic, stochastic, multi-objective, network, and dynamic models differ in structure and computational requirements.

Successful formulation requires translating business language into units, equations, inequalities, and logical relationships. A model may be infeasible when restrictions conflict, unbounded when the objective can improve indefinitely, or misleading when assumptions omit important operational or ethical realities. Validation, sensitivity analysis, scenario testing, and stakeholder review are therefore essential. Optimization software supplies computational power, but professional judgment determines whether a model is appropriate and whether its recommendation should be implemented.

Key Terms

Decision variable: A controllable quantity representing a decision.

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

Objective function: A mathematical expression measuring the criterion to maximize or minimize.

Constraint: A restriction imposed on decision variables.

Feasible solution: A decision satisfying every constraint and domain condition.

Feasible region: The set of all feasible solutions.

Optimal solution: A feasible solution with the best objective value.

Linear programming: Optimization with linear objective and constraint expressions.

Integer programming: Optimization requiring one or more integer variables.

Binary variable: A variable restricted to 0 or 1.

Mixed-integer linear program: A linear model containing continuous and integer variables.

Deterministic optimization: Optimization using fixed parameter values.

Stochastic optimization: Optimization representing uncertainty explicitly.

Binding constraint: A constraint satisfied as an equality at a solution.

Slack: Unused capacity in a ≤ constraint.

Shadow price: A marginal value associated with relaxing a constraint right-hand side.

Infeasibility: The absence of any solution satisfying all restrictions.

Unboundedness: The ability to improve the objective indefinitely without violating constraints.

Sensitivity analysis: Examination of solution changes when inputs vary.

Model validation: Evidence that a model is mathematically correct and appropriately represents its decision context.

Review Questions

5 Short Answer Questions

1. Distinguish a decision variable from a parameter and give one example of each.
2. What is the difference between a feasible solution and an optimal solution?
3. Why are units important when formulating constraints?
4. Define a binary variable and identify one business use.
5. What does a binding constraint indicate?

5 Long Answer Questions

1. Explain the principal components of an optimization model and their relationships.
2. Compare linear programming, nonlinear programming, and mixed-integer programming.
3. Describe a systematic process for translating a business problem into a mathematical model.
4. Explain the causes and diagnostic approaches for infeasibility and unboundedness.
5. Discuss why sensitivity analysis and stakeholder validation are necessary even when a solver reports an optimum.

3 Analytical / Application Questions

1. A firm maximizes profit from two products but obtains an unbounded solution. Identify likely modeling omissions and propose diagnostic checks.
2. Formulate a production model with three products, two resources, product-specific demand limits, and an integer requirement for one product. State all variables, parameters, objective, constraints, and assumptions.
3. A distribution model recommends one facility under the forecast but two facilities under a high-demand scenario. Analyze the cost, capacity, resilience, and governance considerations before implementation.

Multiple Choice Questions with Answers

Question 1

Which component of an optimization model represents a controllable choice?

A. Parameter

B. Decision variable

C. Constraint

D. Shadow price

Correct Answer: B. Decision variable

Question 2

For the constraint 3x + 2y ≤ 100, what are the units of the left-hand side if the coefficients are hours per unit and x and y are units produced?

A. Units

B. Hours

C. Pounds

D. Products per hour

Correct Answer: B. Hours

Question 3

A model containing both continuous shipment quantities and binary facility-opening decisions is a:

A. Linear regression model

B. Continuous nonlinear program

C. Mixed-integer linear program

D. Unconstrained optimization model

Correct Answer: C. Mixed-integer linear program

Question 4

What does it mean if an optimization model is infeasible?

A. The objective has multiple optima

B. No assignment satisfies all restrictions

C. The objective can increase indefinitely

D. All constraints are nonbinding

Correct Answer: B. No assignment satisfies all restrictions

Question 5

In a maximization model, unboundedness commonly indicates:

A. A missing upper bound or capacity restriction

B. Too many feasible solutions

C. A binding demand constraint

D. A binary variable

Correct Answer: A. A missing upper bound or capacity restriction

Question 6

If x = 40 and y = 30 satisfy 2x + y ≤ 120, what is the slack?

A. 10

B. 20

C. 40

D. 120

Correct Answer: B. 20

Question 7

Which expression is nonlinear?

A. 4x + 2y

B. 7x − 3

C. x² + y

D. 5x + 8y ≤ 50

Correct Answer: C. x² + y

Question 8

What is the primary purpose of sensitivity analysis?

A. To remove all constraints

B. To examine how changes in inputs affect results

C. To convert all variables to binary form

D. To guarantee a unique optimum

Correct Answer: B. To examine how changes in inputs affect results

Question 9

If a facility-opening variable y_j is binary and the linking constraint is Σ_i x_ij ≤ K_j y_j, what happens when y_j = 0?

A. The facility can operate at double capacity

B. Shipment from the facility is restricted to zero

C. Demand is automatically reduced

D. Variable shipping costs disappear

Correct Answer: B. Shipment from the facility is restricted to zero

Question 10

Which statement best describes an optimization solver's reported optimum?

A. It is automatically the best real-world decision under every circumstance

B. It is optimal under the model's data, structure, and assumptions

C. It proves that all parameters are accurate

D. It eliminates the need for managerial review

Correct Answer: B. It is optimal under the model's data, structure, and assumptions