Quantitative Decision Models and Optimization

Introduction

Human resource managers make decisions under constraints. They allocate employees to shifts, determine training budgets, select recruitment channels, schedule interviews, design compensation plans, and decide which projects should receive scarce resources. Quantitative decision models convert these managerial problems into structured representations that support transparent, rational, and often optimal choices.

Optimization is the process of selecting the best feasible alternative according to a stated objective. The objective may be to minimize cost, maximize productivity, maximize employee coverage, reduce hiring time, or balance several competing goals. Business mathematics and statistics provide the language for expressing objectives, constraints, uncertainty, and evidence.

A mathematical model does not replace managerial judgment. It clarifies assumptions, exposes trade-offs, and allows alternative policies to be compared. Its usefulness depends on the quality of the data, the realism of the assumptions, and the organization’s willingness to consider ethical and human consequences.

Learning Objectives

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

1. Explain the purpose and components of quantitative decision models.
2. Formulate linear programming models from business situations.
3. Solve simple optimization problems graphically and interpret their solutions.
4. Explain the logic of the simplex, transportation, assignment, and integer programming models.
5. Distinguish certainty, risk, and uncertainty in decision-making.
6. Use expected monetary value, regret, and decision trees appropriately.
7. Interpret sensitivity analysis and recognize model limitations.
8. Apply optimization to an HR management problem while considering fairness, privacy, and organizational context.

1. Foundations of Quantitative Decision Models

A decision model represents a real situation using variables, relationships, constraints, and an objective. Most operational models contain the following elements:

A general optimization model can be written as:

Here, x_j are decision variables, Z is the objective value, g_i represents a constraint relationship, and b_i is the available resource or required level. Non-negativity is appropriate when variables represent quantities such as workers, hours, or expenditure. It may be replaced by other bounds when negative or restricted values are meaningful.

Model-building process

A reliable modelling process involves:

1. Define the decision problem. State what must be decided and by whom.
2. Identify controllable variables. Avoid treating fixed or unknowable quantities as decisions.
3. Specify the objective. Define what “best” means and how it will be measured.
4. Collect parameters. Use relevant, reliable, and sufficiently current data.
5. Formulate constraints. Include operational, financial, legal, contractual, and ethical restrictions.
6. Solve the model. Use graphical methods, algorithms, spreadsheets, or optimization software.
7. Validate the model. Compare outputs with managerial experience and historical patterns.
8. Conduct sensitivity analysis. Test the effect of changes in assumptions.
9. Implement and monitor. Track actual outcomes and revise the model when conditions change.

2. Linear Programming

Linear programming (LP) is used when the objective and constraints are linear functions of decision variables. Linearity means that each unit contributes a constant amount and that effects are additive. For example, if one training session costs $500, two sessions cost $1,000, assuming no volume discount or capacity effect.

2.1 A simple HR formulation

A company offers two training programs. Program A produces 6 productivity points and requires 3 instructor-hours. Program B produces 4 productivity points and requires 2 instructor-hours. The organization has 30 instructor-hours and wants to maximize productivity. It must deliver at least 2 units of Program A.

Let:

• x = number of units of Program A;
• y = number of units of Program B;
• 6x + 4y = productivity points;
• 3x + 2y = instructor-hours used.

The model is:

The objective and constraints use consistent units: productivity points and instructor-hours. The model assumes constant productivity per program, divisible program quantities, and no interaction between programs.

2.2 Graphical solution

For two variables, each constraint can be plotted on a coordinate graph. The feasible region is the set of points satisfying every constraint. A linear objective reaches its optimum at a corner point of a bounded feasible region, although multiple optimal points may occur when the objective line is parallel to a binding boundary.

For the example, the instructor-hour boundary is:

Its intercepts are (10,0) and (0,15). Because x ≥ 2, the relevant corner points are (2,0), (2,12), and (10,0). Evaluate the objective:

The model has multiple optimal solutions: every feasible point on the segment between (2, 12) and (10, 0) gives 60 productivity points. This occurs because both programs produce the same productivity per instructor-hour: Program A produces 6 ÷ 3 = 2 points per hour, and Program B produces 4 ÷ 2 = 2 points per hour.

2.3 Slack, surplus, and binding constraints

A constraint is binding when it holds with equality at the optimum. A nonbinding resource constraint has unused capacity, called slack. For a “less than or equal to” resource constraint:

For a minimum requirement, the excess above the minimum is often called surplus. Binding constraints identify scarce resources. In HR planning, these may include available interview hours, budget, qualified supervisors, or legally permitted working time.

2.4 Assumptions and limitations of LP

Linear programming commonly assumes:

• Proportionality: contribution changes in direct proportion to activity.
• Additivity: total contribution is the sum of separate contributions.
• Divisibility: variables may take fractional values unless restricted.
• Certainty: coefficients are treated as known.
• Non-negativity: activity levels cannot be negative.

These assumptions may be unrealistic. Employees cannot always be divided into fractions; productivity may decline with fatigue; and demand, absence, and hiring outcomes are uncertain. Integer programming, nonlinear programming, and simulation may therefore be more appropriate.

3. Optimization Models Used in Business and HR

3.1 Integer programming

Integer programming requires some or all decision variables to be integers:

Binary variables take only the values 0 or 1. They are useful for yes/no decisions, such as whether to adopt a recruitment platform. A fixed-cost model may use a binary variable y:

where c is variable cost per unit, F is the fixed cost of activation, x is activity level, and y = 1 if the option is selected. A linking constraint such as x ≤ My prevents activity unless the option is activated; M must be a valid upper bound, not an unnecessarily enormous number.

3.2 Assignment models

An assignment model allocates each worker to exactly one job, or each job to exactly one worker, while minimizing cost or maximizing suitability. If c_ij is the cost of assigning worker i to job j, and x_ij is 1 when that assignment occurs and 0 otherwise:

The model assumes a balanced one-to-one assignment unless dummy workers or jobs are added. In HR, “cost” might represent travel time, training need, mismatch, or risk—not merely monetary expenditure. Ethical review is essential when suitability scores are derived from historical performance data.

3.3 Transportation models

Transportation models determine quantities shipped from origins to destinations at minimum cost. An HR analogue is allocating recruiters or training resources from regional offices to business units. If x_ij is the amount sent from source i to destination j, and c_ij is unit cost:

Here, s_i is source capacity and d_j is destination demand. A balanced transportation problem has total supply equal to total demand; otherwise, a dummy source or destination can represent surplus or unmet demand.

3.4 Scheduling and workforce planning

Scheduling models assign employees to time periods while satisfying coverage and employment rules. A basic coverage constraint is:

where x_i indicates whether schedule pattern i is selected, a_it equals 1 if that pattern covers period t, and R_t is required staffing in period t. Additional constraints may cover maximum consecutive shifts, rest periods, contractual hours, skills, and employee preferences.

4. Decision-Making Under Certainty, Risk, and Uncertainty

Optimization often assumes known coefficients. Decision analysis explicitly considers the information available about future states.

• Certainty: outcomes are known for each alternative.
• Risk: several outcomes are possible and probabilities can reasonably be assigned.
• Uncertainty: possible outcomes are known or partly known, but reliable probabilities are unavailable.

4.1 Expected monetary value

When probabilities are available, expected monetary value (EMV) is:

where a is an alternative, s_j is a state of nature, p_j is the probability of that state, and V(a,s_j) is the payoff. Probabilities must satisfy 0 ≤ p_j ≤ 1 and ∑p_j = 1. EMV is a long-run average criterion; it does not guarantee the outcome of one decision.

Example. A recruitment campaign costs $2,000. If demand is high, it generates a benefit of $10,000 before campaign cost; if demand is low, the benefit is $1,000. Suppose the probability of high demand is 0.6.

Net payoffs are $8,000 and −$1,000. Therefore:

The positive EMV of $4,400 supports the campaign under a risk-neutral criterion. A risk-averse manager may still reject it if the possible loss is unacceptable.

4.2 Expected opportunity loss

Regret, or opportunity loss, is the difference between the best payoff in a state and the payoff of the chosen alternative:

Expected opportunity loss is the probability-weighted average regret. It is useful when managers want to minimize the expected cost of making the wrong choice. The alternative with the highest EMV also has the lowest expected opportunity loss when the same payoff table and probabilities are used.

4.3 Decision trees

A decision tree represents sequential decisions and uncertain events. Decision nodes are usually shown as squares; chance nodes as circles. The tree is evaluated from right to left: calculate expected values at chance nodes, then select the best alternative at decision nodes. Decision trees are useful for recruitment expansion, training investment, and phased workforce restructuring.

A tree must distinguish decisions that the organization controls from events it does not control. Probabilities should be justified by historical data, research, or clearly documented expert judgment.

4.4 Decisions without reliable probabilities

When probabilities cannot be defended, several criteria are possible:

• Maximax: choose the alternative with the best possible payoff; optimistic.
• Maximin: choose the alternative with the best worst-case payoff; conservative.
• Minimax regret: choose the alternative with the smallest maximum regret.
• Laplace criterion: treat states as equally likely when no better basis exists.

These are not universally correct answers. They express different attitudes toward uncertainty and should be reported transparently.

5. Sensitivity Analysis and Model Validation

An optimal solution is conditional on the model’s inputs. Sensitivity analysis asks how much the solution changes when a coefficient, resource limit, or assumption changes.

Important questions include:

1. How much can a resource limit change before the optimal basis changes?
2. How much can an objective coefficient change before another activity becomes preferable?
3. Which constraints are binding?
4. Which data uncertainties could reverse the decision?

The shadow price of a resource is the marginal change in the optimal objective value caused by one additional unit of that resource, within an applicable range. For example, a shadow price of $35 per additional recruiter-hour means that, locally and assuming other conditions remain unchanged, one extra hour could improve the objective by $35. It is not automatically the market price of the resource and is not valid beyond the sensitivity range.

Validation should include:

• Face validity: do knowledgeable managers find the structure reasonable?
• Historical validation: does the model reproduce known outcomes?
• Extreme-case testing: does it behave sensibly under unusually high or low inputs?
• Out-of-sample testing: does it perform acceptably on data not used to build it?
• Implementation review: can the recommendation actually be executed?

6. Illustrative Case: Regional Recruitment Allocation

Background

A fictional organization must allocate 120 recruiter-hours across two recruitment channels: university partnerships and online advertising. University partnerships require 4 hours per campaign unit and yield 18 expected qualified applicants. Online advertising requires 2 hours per campaign unit and yields 7 qualified applicants. At least 10 university campaign units are required to maintain strategic campus relationships. The organization wants to maximize expected qualified applicants.

Formulation

Let x be university campaign units and y be online campaign units.

The resource productivity is 18 ÷ 4 = 4.5 applicants per hour for university partnerships and 7 ÷ 2 = 3.5 applicants per hour for online advertising. Because university partnerships are more productive per hour and there is no upper bound on them, the mathematical optimum is x = 30, y = 0, producing Q = 540 expected qualified applicants.

Managerial interpretation

The result does not mean online advertising is worthless. It means that, under the stated assumptions, every available hour should be allocated to the higher-yield channel after satisfying the minimum requirement. In reality, the organization may need channel diversity, may face a limited number of university events, or may value geographic and demographic reach not represented by the single objective.

A more realistic model could add:

• a maximum of 20 university campaign units;
• a minimum of 15 online campaign units;
• separate diversity or inclusion targets;
• uncertainty in applicant yields;
• costs and budget limits;
• quality-adjusted applicant measures rather than raw counts.

Learning points

1. The objective function determines what “best” means.
2. Unit productivity can reveal the economic logic of an optimum.
3. A mathematically optimal answer may be operationally incomplete.
4. Constraints can represent strategic, legal, and ethical requirements, not only physical resources.
5. Model refinement should be guided by material weaknesses, not by unnecessary complexity.

7. Technology, AI, and Governance

Spreadsheet solvers, specialized optimization packages, workforce-management systems, and cloud analytics platforms can solve large models rapidly. Artificial intelligence may forecast demand, estimate attrition risk, or generate candidate suitability scores that become inputs to an optimization model. However, predictive accuracy does not guarantee a fair or lawful decision.

HR professionals should maintain:

• documented definitions of variables and objectives;
• data-quality checks and version control;
• access controls for personal and sensitive data;
• human review for high-impact decisions;
• bias and disparate-impact testing where relevant;
• explainable decision rules;
• audit trails showing data, model version, and approval;
• mechanisms for employee or candidate appeal;
• periodic monitoring for model drift.

Optimization can produce unfair outcomes if historical inequalities are treated as desirable targets. For example, minimizing recruitment cost may systematically reduce outreach to underrepresented groups. A responsible model may therefore include explicit fairness constraints, but such constraints should be designed with legal, organizational, and stakeholder expertise.

8. Best Practices and Common Errors

Best practices

• Define the managerial purpose before selecting an algorithm.
• Use the simplest model that captures material relationships.
• State all assumptions and units.
• Separate factual inputs from subjective judgments.
• Test multiple scenarios rather than relying on one forecast.
• Include infeasibility checks and practical implementation limits.
• Present trade-offs, not merely the recommended solution.
• Combine quantitative evidence with professional and employee perspectives.

Common errors

• Maximizing a measurable quantity that is only a weak proxy for organizational value.
• Omitting a legally required or ethically important constraint.
• Treating a fractional employee as implementable without rounding and rechecking feasibility.
• Using probabilities that do not sum to one.
• Interpreting a shadow price outside its valid range.
• Confusing correlation-based forecasts with causal effects.
• Assuming software output is correct because it is numerically precise.
• Failing to distinguish model infeasibility from poor software configuration.

Chapter Summary

Quantitative decision models represent managerial problems through decision variables, objectives, parameters, and constraints. Linear programming is appropriate when relationships are linear and assumptions such as proportionality and divisibility are acceptable. Integer, assignment, transportation, and scheduling models extend optimization to discrete and allocation decisions. Decision analysis distinguishes certainty, risk, and uncertainty, using expected monetary value, regret, decision trees, and alternative criteria when probabilities are unavailable. Sensitivity analysis reveals the importance of scarce resources and the stability of a recommendation. In HR, optimization can improve staffing, recruitment, training, and workforce allocation, but responsible use requires valid data, human oversight, fairness safeguards, privacy protection, and continual monitoring.

Key Terms

Decision variable: A controllable quantity selected by the decision-maker.

Objective function: A mathematical expression representing the goal to maximize or minimize.

Constraint: A restriction that feasible decisions must satisfy.

Feasible solution: A solution satisfying all model constraints.

Optimal solution: The best feasible solution according to the objective.

Linear programming: Optimization with linear objective and constraint relationships.

Integer programming: Optimization in which specified variables must be integers.

Binary variable: A variable restricted to 0 or 1.

Slack: Unused amount of a resource in a less-than-or-equal constraint.

Surplus: Amount by which a minimum requirement is exceeded.

Binding constraint: A constraint satisfied exactly at a solution.

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

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

Expected monetary value: Probability-weighted average payoff of an alternative.

Regret: Loss relative to the best decision after a state occurs.

Risk: Decision context in which probabilities of outcomes can be assigned.

Uncertainty: Decision context in which reliable probabilities are unavailable.

Decision tree: A graphical representation of sequential decisions and chance events.

Model validation: Evaluation of whether a model is credible and useful for its intended purpose.

Review Questions

Short Answer Questions

1. What are the main components of a quantitative decision model?
2. Define a binding constraint and give an HR example.
3. What is the difference between risk and uncertainty?
4. Why are binary variables useful in HR planning?
5. What does a shadow price measure?

Long Answer Questions

1. Explain the process of formulating and validating a linear programming model for workforce planning.
2. Discuss the assumptions of linear programming and explain when they may be unrealistic in HR applications.
3. Compare assignment, transportation, and scheduling models.
4. Explain expected monetary value, expected opportunity loss, and minimax regret.
5. Discuss the ethical and governance requirements for using AI-supported optimization in HR.

Analytical / Application Questions

1. A company has 80 recruiter-hours. Channel A yields 12 qualified applicants per 3 hours, while Channel B yields 15 applicants per 5 hours. Formulate a model that includes a minimum of 5 units of Channel B and explain which channel has the higher applicant yield per hour.
2. Construct a payoff table for two recruitment strategies under high and low demand. Assign probabilities, calculate EMV, and explain why the EMV recommendation may not suit a highly risk-averse manager.
3. Design a workforce scheduling model for two shifts and three employee groups. Identify decision variables, coverage constraints, maximum-hour constraints, and one fairness constraint.

Multiple Choice Questions with Answers

Question 1

Which component of an optimization model states what is to be maximized or minimized?

A. Decision variable

B. Objective function

C. Slack variable

D. Parameter

Correct Answer: B. Objective function

Question 2

A feasible solution is one that:

A. Has the largest possible objective value

B. Uses every resource completely

C. Satisfies all stated constraints

D. Contains only integer values

Correct Answer: C. Satisfies all stated constraints

Question 3

If an activity produces 20 units of output using 5 hours, its output per hour is:

A. 2

B. 4

C. 15

D. 25

Correct Answer: B. 4

Question 4

Which model is most directly designed for one-to-one allocation of workers to jobs?

A. Assignment model

B. Inventory model

C. Regression model

D. Index-number model

Correct Answer: A. Assignment model

Question 5

If a resource constraint is nonbinding, it has:

A. Negative slack

B. No feasible solution

C. Unused capacity

D. An unavoidable shortage

Correct Answer: C. Unused capacity

Question 6

Expected monetary value is calculated by:

A. Adding all payoffs without probabilities

B. Multiplying each payoff by its probability and summing

C. Selecting the largest possible payoff

D. Subtracting the smallest payoff from the largest

Correct Answer: B. Multiplying each payoff by its probability and summing

Question 7

In a decision tree, a chance node represents:

A. A decision controlled by management

B. An uncertain event with possible outcomes

C. A binding constraint

D. A fixed cost

Correct Answer: B. An uncertain event with possible outcomes

Question 8

A binary decision variable can take which values?

A. Any real number

B. −1 and 1 only

C. 0 and 1 only

D. Any positive integer

Correct Answer: C. 0 and 1 only

Question 9

A shadow price is best interpreted as:

A. The total accounting price of all resources

B. The marginal value of relaxing a constraint within a valid range

C. The probability of an uncertain event

D. The cost of model software

Correct Answer: B. The marginal value of relaxing a constraint within a valid range

Question 10

Which is an important governance practice when using AI in HR optimization?

A. Removing all human review

B. Ignoring historical data quality

C. Maintaining audit trails and testing for bias

D. Maximizing speed regardless of fairness

Correct Answer: C. Maintaining audit trails and testing for bias