Operations Research, Analytics and Decision-Making

Introduction and Context

Industrial managers make decisions about product mix, machine loading, workforce allocation, inventory, transportation, maintenance, quality, and investment. These decisions are difficult because resources are limited, demand is uncertain, processes interact, and alternative actions have different costs and benefits. Operations research (OR) provides systematic quantitative methods for representing such problems and identifying satisfactory or optimal decisions. Business and industrial analytics extend OR by using data to describe what has happened, predict what is likely to happen, and prescribe what should be done.

For a machine-technology enterprise, OR is not a substitute for engineering judgment. It is a disciplined decision-support approach. A mathematical model can reveal trade-offs, bottlenecks, and sensitivity to assumptions, while managers contribute knowledge about safety, maintainability, labor relations, quality, and strategic priorities.

Learning Objectives

After studying this chapter, the learner should be able to:

1. Explain the role, scope, and limitations of operations research and analytics.
2. Formulate linear programming, transportation, assignment, queuing, simulation, and decision-tree models.
3. Solve or interpret small quantitative models and evaluate their assumptions.
4. Distinguish descriptive, predictive, and prescriptive analytics.
5. Apply expected-value, regret, and decision-tree methods under certainty, risk, and uncertainty.
6. Interpret sensitivity, risk, ethics, and implementation issues in managerial decisions.

1. Operations Research as a Decision Discipline

Operations research is the application of mathematical models, data, and analytical reasoning to decisions involving scarce resources. A typical OR study follows these stages:

1. Define the decision problem and the decision-maker’s objective.
2. Identify controllable decision variables, parameters, constraints, and outcomes.
3. Construct a model that simplifies the real system without omitting essential relationships.
4. Obtain and validate data.
5. Solve the model using an appropriate method or software.
6. Test sensitivity and robustness.
7. Implement, monitor, and revise the decision.

A model is a purposeful representation, not a complete copy of reality. A linear programming model may represent labor hours and material usage accurately while omitting worker fatigue or setup disruption. Such omissions must be recognized when interpreting results.

Decision environments

A decision under certainty has known consequences for each alternative. Under risk, possible states of nature and their probabilities are available. Under uncertainty, states may be known but reliable probabilities are not. These categories affect the appropriate decision method.

The main components of a decision model are:

2. Linear Programming

Linear programming (LP) determines the best values of decision variables when the objective and constraints are linear. It is appropriate when proportionality, additivity, divisibility, and reasonably certain coefficients are acceptable assumptions.

General formulation

Here, xⱼ is the quantity of activity j, cⱼ is its unit contribution to the objective, aᵢⱼ is the amount of resource i used by one unit of activity j, and bᵢ is the available amount or requirement for constraint i. The objective value is Z. Units must be consistent: if cⱼ is dollars per unit and xⱼ is units, then Z is dollars.

The non-negativity condition is appropriate for production quantities, labor allocation, and shipments. If a variable can take negative values, it must be modeled differently, often by representing it as the difference of two non-negative variables.

Worked example: product mix

A workshop produces precision brackets (B) and housings (H). Each bracket requires 2 machine hours and 1 labor hour and contributes $30. Each housing requires 3 machine hours and 2 labor hours and contributes $50. Weekly capacity is 120 machine hours and 80 labor hours.

Let x = brackets and y = housings. The model is:

The corner points are (0, 0), (60, 0), (0, 40), and the intersection of the two binding constraints. Solving

gives y = 40 and x = 0. Thus, the feasible corner points include (0, 0), (60, 0), and (0, 40). Their contributions are $0, $1,800, and $2,000 respectively. The LP recommends 40 housings and no brackets, with both machine and labor constraints binding. In practice, management should check whether producing no brackets is commercially acceptable and whether integer quantities, customer commitments, or minimum product requirements have been omitted.

Interpretation and sensitivity

A binding constraint is fully used at the optimum. A slack constraint has unused capacity. The shadow price of a resource estimates the improvement in the objective from one additional unit of that resource, within the allowable range for which the current solution structure remains valid. Shadow prices are marginal values, not universal prices; they may change when capacity changes substantially or when another constraint becomes binding.

Sensitivity analysis asks questions such as:

• What happens if demand, contribution, or capacity changes?
• How much can a coefficient change before the solution changes?
• Is an investment in an additional machine justified?

LP has important limitations. It assumes linear relationships, often permits fractional quantities, and may not represent setup times, sequence-dependent changeovers, nonlinear costs, or uncertain data. Integer programming is needed when decisions involve indivisible machines, shifts, vehicles, or yes/no investments.

3. Transportation and Assignment Models

Transportation model

The transportation model allocates shipments from several sources to several destinations at minimum cost. Let xᵢⱼ be the quantity shipped from source i to destination j, cᵢⱼ the unit shipping cost, sᵢ source supply, and dⱼ destination demand.

For a balanced problem, total supply equals total demand and the inequalities are usually equations. If supply exceeds demand, a dummy destination represents unused supply; if demand exceeds supply, a dummy source represents shortage. The basic model assumes unit costs are known and linear, routes are feasible, and shipment quantities can be divided.

Initial feasible solutions can be generated by the northwest-corner, least-cost, or Vogel approximation methods. Optimality may be tested with the stepping-stone or MODI method, or by a general LP solver. Transportation models are useful for distributing raw materials, finished goods, spare parts, and maintenance teams.

Assignment model

The assignment model is a special transportation model in which each worker, machine, or job is assigned one-to-one. Let xᵢⱼ equal 1 if agent i performs task j, and 0 otherwise. If cᵢⱼ is cost or time:

The model requires equal numbers of agents and tasks; dummy rows or columns can balance an unbalanced problem. The Hungarian algorithm solves the classical assignment problem efficiently. In practice, skill qualifications, shift restrictions, fatigue, and fairness may require additional constraints or a more general integer model.

4. Queuing Models and Service Capacity

Queues occur when demand for a service temporarily exceeds available service capacity. Examples include machines awaiting repair, jobs waiting for inspection, vehicles at loading docks, and customers waiting for technical support. Queue performance involves a trade-off: more capacity reduces waiting but increases operating cost.

A standard single-server model is M/M/1. It assumes Poisson arrivals at rate λ, exponentially distributed service times with rate μ, one server, an unlimited calling population, first-come-first-served discipline, and stable operation. Stability requires:

where ρ is utilization. If λ ≥ μ, the theoretical steady-state queue grows without bound.

For M/M/1, the expected number in the system and expected time in the system are:

By Little’s Law:

Here, L includes waiting and service, and W is total time in the system. For the queue alone:

Rates must use the same time unit. If arrivals are 6 jobs per hour and service is 8 jobs per hour, then ρ = 6/8 = 0.75, W = 1/(8−6) = 0.5 hour, and L = 6(0.5) = 3 jobs. These are long-run expected values, not guarantees for a particular hour.

Real systems may have multiple servers, finite populations, priority classes, non-exponential service times, balking, reneging, or machine-dependent failures. In those cases, formulas for M/M/1 may be misleading. Simulation or specialized queueing models are preferable.

5. Simulation as an Experimental Method

Simulation imitates the operation of a system over time. It is especially useful when analytical formulas are unavailable or when interactions among variability, priorities, failures, and capacity are complex.

A discrete-event simulation progresses from one event to the next. Typical events include arrival, service completion, machine failure, repair completion, and shift change. A basic procedure is:

1. Define entities, resources, states, events, and performance measures.
2. Specify probability distributions from reliable data.
3. Generate random variates using a controlled random-number stream.
4. Run the model for a warm-up period if the system begins artificially empty.
5. Collect observations over multiple replications.
6. Compare alternatives using confidence intervals and practical significance.
7. Validate the model against observed system behavior.

For a sample mean ȳ based on n independent observations with sample standard deviation s, an approximate 95% confidence interval is:

The t value is the Student’s t critical value with n−1 degrees of freedom. Independence and approximate normality of the mean are assumptions; correlated time-series observations require suitable batch-means or other methods. Simulation does not produce certainty merely because a computer is used. Results depend on model logic, input distributions, random variation, and experimental design.

6. Analytics for Industrial Decisions

Analytics is commonly divided into three connected forms:

Descriptive analytics should establish data definitions and quality before predictive modeling. Predictive models estimate an outcome, but prediction is not automatically causation. Prescriptive models combine predictions with constraints, objectives, and values; an inaccurate forecast can therefore produce a confident but poor recommendation.

Basic forecasting and predictive measures

For observed demand yₜ and forecast Fₜ, the error is:

Common measures are:

MAE is in the original units, MSE is in squared units, and RMSE returns to the original units while penalizing large errors more heavily than MAE. Percentage measures require care when actual values are zero or close to zero.

A simple linear regression model is:

where y is the response, x a predictor, β₀ the intercept, β₁ the expected change in y per unit change in x, and ε an unexplained error term. Interpretation requires attention to sampling, omitted variables, nonlinearity, outliers, and whether the data support causal claims.

Industrial analytics may use sensors, enterprise-resource-planning data, maintenance records, and digital twins. However, data governance is essential: timestamps, units, missing values, sensor calibration, access controls, and versioned models should be documented.

7. Decision Analysis under Certainty, Risk, and Uncertainty

A decision table lists alternatives, states of nature, probabilities, and payoffs. Under risk, the expected monetary value (EMV) of alternative a is:

where pⱼ is the probability of state j, V(a,j) is the payoff, and ∑ⱼpⱼ = 1. EMV is appropriate when probabilities are meaningful and the decision-maker is evaluating repeated or sufficiently comparable decisions. It does not mean that the expected payoff will actually occur.

For a loss or cost table, expected cost is minimized rather than payoff maximized. Expected opportunity loss uses regret:

The minimax-regret rule chooses the alternative with the smallest maximum regret. Under uncertainty, other rules include:

• Maximax: choose the alternative with the highest possible payoff; optimistic.
• Maximin: choose the alternative with the highest worst-case payoff; conservative.
• Laplace: treat states as equally likely when no probabilities are available.
• Hurwicz: use a weighted combination of best and worst outcomes, with optimism coefficient α, where 0 ≤ α ≤ 1.

The expected value of perfect information (EVPI) is:

EVPI is an upper bound on what information is worth; it is not automatically the amount a firm should pay for a forecast or study.

Decision trees

A decision tree represents sequential decisions and uncertain events. Squares denote decision nodes, circles denote chance nodes, and terminal branches show payoffs. The tree is evaluated by rollback: calculate expected values at chance nodes and select the best feasible branch at decision nodes.

Figure: Generic decision-tree logic, with decisions followed by uncertain outcomes.

8. Integrated Illustrative Case: Maintenance-Capacity Decision

Background

The following is a fictional illustrative case. A machining plant experiences failures in a critical computer numerical control (CNC) cell. Management must decide whether to retain its current maintenance arrangement or install a condition-monitoring system.

Problem

The current arrangement costs $18,000 per year. A monitoring system costs $30,000 initially and $8,000 per year thereafter. If failures are high, the current arrangement causes an additional expected disruption cost of $42,000 per year. With monitoring, disruption cost in a high-failure year is expected to fall to $12,000. In a low-failure year, disruption cost is $8,000 under the current arrangement and $6,000 with monitoring. Engineering estimates the probability of a high-failure year as 0.40.

Analysis

Ignoring the common time-value-of-money issue for this one-year comparison, annual costs are:

Expected cost of current arrangement:

Expected cost of monitoring:

On these assumptions, the current arrangement has lower expected annual cost by $6,800. However, the result is sensitive to the probability of high failures and to whether the monitoring system reduces other losses, such as safety exposure, emergency overtime, or customer penalties. If the probability of a high-failure year is p, monitoring is preferred when:

which simplifies to p > 0.60. Thus, monitoring becomes economically attractive when the high-failure probability exceeds 60%, before considering discounting or strategic benefits.

Approach and outcome

The plant should not implement the recommendation mechanically. It should validate failure probabilities, distinguish avoidable from unavoidable downtime, pilot the sensors on the critical cell, and measure false alarms, detection lead time, and maintenance response. A multi-year net-present-value analysis is required if the equipment lasts several years.

Learning points

1. A decision model makes assumptions visible.
2. Expected value supports risk-based comparison but does not capture every managerial value.
3. Sensitivity analysis identifies the probability threshold that changes the recommendation.
4. Pilot testing can improve both data quality and implementation confidence.
5. Safety and reliability consequences should not be reduced to cost alone.

9. Implementation, Ethics, and Governance

A technically optimal recommendation can fail during implementation. Managers should assign ownership, define decision rights, document assumptions, and monitor actual outcomes against predicted outcomes. Key governance practices include:

• Data lineage: record where data originated, how it was cleaned, and who changed it.
• Model validation: compare outputs with historical and operational evidence.
• Version control: preserve the model, data snapshot, and assumptions used for each decision.
• Human oversight: require review when recommendations affect safety, employment, or regulatory compliance.
• Explainability: communicate important drivers, constraints, and uncertainty.
• Privacy and security: restrict access to worker, customer, and production data.
• Fairness: test whether assignments, schedules, or performance scores disadvantage protected groups.
• Cybersecurity: secure connected machines, sensors, interfaces, and analytical platforms.

AI and machine learning can improve predictive maintenance, visual inspection, demand forecasting, and scheduling. They do not eliminate the need for OR. A prediction of failure must be connected to a prescriptive maintenance action subject to technician capacity, spare parts, production priorities, and safety rules. Automated optimization should include constraints that prevent unsafe or infeasible recommendations.

10. Best Practices and Limitations

Effective OR and analytics projects follow several principles:

1. Begin with a decision, not with a fashionable algorithm.
2. Define the objective operationally and distinguish cost, profit, service, risk, and safety measures.
3. Involve operators, engineers, finance, and affected stakeholders in model design.
4. Use the simplest valid model before adding complexity.
5. Validate both data and model logic.
6. Perform sensitivity, scenario, and stress analysis.
7. Separate statistical significance from practical significance.
8. Report uncertainty rather than presenting false precision.
9. Design implementation and feedback mechanisms from the beginning.
10. Reassess the model when technology, demand, regulations, or process conditions change.

OR models can be wrong because of poor data, unrealistic assumptions, omitted constraints, unstable environments, or objective functions that reward undesirable behavior. Optimization may exploit a loophole in the model—for example, minimizing labor cost while increasing unsafe workload. The quality of a decision therefore depends on model validity, organizational judgment, and responsible implementation, not only on mathematical optimality.

Chapter Summary

Operations research provides structured methods for allocating scarce industrial resources and evaluating alternatives. Linear programming models continuous allocation problems with linear objectives and constraints; transportation models route quantities between sources and destinations; assignment models allocate agents to tasks one-to-one. Queuing theory evaluates congestion and service capacity under stated probabilistic assumptions, while simulation experiments with complex systems that may not have tractable formulas.

Analytics progresses from descriptive understanding to predictive estimation and prescriptive action. Forecast errors can be summarized with MAE, MSE, and RMSE, but predictive accuracy does not establish causation. Decision analysis distinguishes certainty, risk, and uncertainty. Expected monetary value is useful when probabilities are credible, while minimax regret and related criteria support decisions when probabilities are unavailable. Decision trees represent sequential choices and uncertain outcomes.

Professional practice requires sensitivity analysis, data governance, ethical oversight, human judgment, and implementation monitoring. A model is valuable when it improves the quality, transparency, and robustness of a real decision—not merely when it produces a mathematically elegant answer.

Key Terms

Operations research: Quantitative study of decisions involving scarce resources.

Decision variable: A controllable quantity or choice in a model.

Objective function: Mathematical expression representing what is maximized or minimized.

Constraint: A limitation or required relationship in a model.

Linear programming: Optimization with linear objective and constraint relationships.

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

Transportation model: Model allocating shipments from sources to destinations.

Assignment model: One-to-one allocation of agents, machines, or jobs.

Queue: Entities waiting for service or processing.

Utilization: Proportion of service capacity being used, commonly ρ = λ/μ.

Little’s Law: Relationship L = λW between average number, arrival rate, and time.

Simulation: Computer-based imitation of system behavior.

Descriptive analytics: Analysis of what has happened.

Predictive analytics: Estimation of what is likely to happen.

Prescriptive analytics: Analysis of what action should be taken.

Expected monetary value: Probability-weighted average payoff.

Regret: Opportunity loss from not selecting the best alternative for a realized state.

Decision tree: Graphical representation of sequential decisions and chance events.

Sensitivity analysis: Study of how results change when assumptions or inputs vary.

Data governance: Policies and controls for data quality, security, access, and accountability.

Review Questions

Short Answer Questions

1. What are the principal stages of an operations-research study?
2. Distinguish a binding constraint from a slack constraint in linear programming.
3. State the stability condition for an M/M/1 queue and explain its meaning.
4. Differentiate descriptive, predictive, and prescriptive analytics.
5. What is the purpose of sensitivity analysis?

Long Answer Questions

1. Explain the assumptions of linear programming and discuss situations in which integer programming would be more appropriate.
2. Describe how a transportation model is formulated, including the treatment of unbalanced supply and demand.
3. Explain the M/M/1 queueing model, define its principal measures, and discuss why its assumptions may fail in a real factory.
4. Compare expected monetary value, minimax regret, maximax, and maximin decision criteria.
5. Discuss the role of data governance, human oversight, and ethical review in industrial analytics.

Analytical / Application Questions

1. A plant receives 10 jobs per hour and one inspection station can process 12 jobs per hour. Using the M/M/1 assumptions, calculate utilization, expected time in the system, and expected number of jobs in the system. Interpret the results.
2. Formulate a linear programming model for two products requiring different machine and labor hours when capacities, contributions, and a minimum contractual quantity for one product are provided. Explain which constraints are likely to be binding and how you would test the result.
3. Construct a decision tree for choosing between an equipment upgrade and continued operation when demand may be high or low. Assign probabilities and payoffs from a realistic industrial scenario, calculate EMV, and conduct a sensitivity analysis on the probability of high demand.

Multiple Choice Questions with Answers

Question 1

Which condition is required for a stable M/M/1 queue?

A. λ = μ

B. λ > μ

C. λ < μ

D. λ = 0

Correct Answer: C. λ < μ

Question 2

In a linear programming model, what does a shadow price generally measure?

A. The total fixed cost of a product

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

C. The probability that a constraint will be violated

D. The average selling price of an activity

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

Question 3

A transportation model primarily determines how to:

A. Assign one worker to one task only

B. Forecast demand from time-series data

C. Route quantities from sources to destinations at minimum cost

D. Calculate the variance of service times

Correct Answer: C. Route quantities from sources to destinations at minimum cost

Question 4

If arrivals occur at 6 jobs per hour and service occurs at 8 jobs per hour in an M/M/1 system, utilization is:

A. 0.25

B. 0.50

C. 0.75

D. 1.33

Correct Answer: C. 0.75

Question 5

Which type of analytics answers the question, “What should be done?”

A. Descriptive analytics

B. Predictive analytics

C. Prescriptive analytics

D. Exploratory data entry

Correct Answer: C. Prescriptive analytics

Question 6

For an M/M/1 queue with λ = 6 per hour and μ = 8 per hour, the expected time in the system is:

A. 0.125 hour

B. 0.25 hour

C. 0.50 hour

D. 2.00 hours

Correct Answer: C. 0.50 hour

Question 7

Which decision criterion selects the alternative with the smallest maximum regret?

A. Maximax

B. Maximin

C. Laplace

D. Minimax regret

Correct Answer: D. Minimax regret

Question 8

In a simulation study, why are multiple replications commonly used?

A. To eliminate all randomness from the model

B. To estimate variability and improve confidence in results

C. To guarantee that the model is valid

D. To avoid defining probability distributions

Correct Answer: B. To estimate variability and improve confidence in results

Question 9

In an assignment model, the binary variable xᵢⱼ = 1 usually means that:

A. Agent i is assigned to task j

B. Task j is canceled

C. Agent i is unavailable

D. The cost of task j is zero

Correct Answer: A. Agent i is assigned to task j

Question 10

A decision tree is evaluated by rollback by:

A. Ignoring probabilities at chance nodes

B. Selecting the largest cost at every node

C. Calculating expected values at chance nodes and selecting the best feasible decision

D. Replacing all uncertain outcomes with the worst outcome

Correct Answer: C. Calculating expected values at chance nodes and selecting the best feasible decision