Optimization and Operations Decision Models

Introduction

Machine-technology organizations continuously make decisions about product design, material usage, machine capacity, production quantities, maintenance timing, workforce allocation, inventory, and delivery schedules. These decisions are interconnected: improving one measure may worsen another. Optimization provides a systematic way to identify the best feasible decision according to a stated objective, while operations decision models represent the structure, constraints, uncertainty, and trade-offs of real systems.

In this chapter, an optimization problem is understood as the selection of decision variables that minimizes or maximizes an objective function subject to constraints. The objective may be cost, production time, energy consumption, profit, reliability, or a weighted combination of several criteria. The model is useful only when its assumptions, data, and implementation are credible.

Learning Objectives

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

1. Formulate objective functions, decision variables, and constraints.
2. Distinguish unconstrained, constrained, linear, nonlinear, integer, and multi-objective models.
3. Solve basic optimization problems analytically and interpret their solutions.
4. Apply linear programming to resource-allocation decisions.
5. Explain scheduling, inventory, and maintenance decision models.
6. Evaluate sensitivity, feasibility, limitations, and ethical implications of optimization results.

1. Fundamental Structure of an Optimization Model

A general mathematical optimization model can be written as:

Here, x is the vector of decision variables; f(x) is the objective function; g_i(x) are inequality constraints; h_j(x) are equality constraints; and X specifies domain restrictions such as non-negativity, integrality, or upper bounds. Every term must be dimensionally meaningful. For example, a cost objective should be expressed in a common currency, and a constraint comparing production quantities must use consistent units.

A good formulation answers four questions:

• What can the decision-maker control?
• What outcome is to be improved?
• What physical, technical, financial, or policy limits apply?
• Which assumptions convert the real system into a manageable model?

The feasible region is the set of all values satisfying every constraint. A feasible solution belongs to this region; an optimal solution is a feasible solution with the best objective value. A model may be infeasible when requirements conflict, unbounded when the objective can improve indefinitely, or have multiple optimal solutions.

Types of Variables

Continuous variables may take any value in an interval, such as cutting speed. Integer variables take whole-number values, such as the number of machines purchased. Binary variables take values 0 or 1, often representing whether a machine is selected, a job is accepted, or a maintenance action is performed. Mixed-integer models combine these types.

Objective Functions

A single-objective production model may minimize total cost:

The summation extends over all products or activities. Costs must be measured over the same planning period. In a profit model, profit equals revenue minus cost. If demand, prices, or processing times are uncertain, the objective may represent expected value, worst-case performance, or a risk-adjusted measure rather than a guaranteed outcome.

2. Unconstrained Optimization

In unconstrained optimization, a differentiable function is optimized without explicit restrictions on the variables. For one variable, a necessary condition for an interior optimum is:

The point x is a candidate, not automatically an optimum. For a twice-differentiable function, f″(x) > 0 indicates a local minimum, while f″(x) < 0 indicates a local maximum. If f″(x) = 0, the second-derivative test is inconclusive and further analysis is required.

Worked Example: Economic Cutting Speed

Suppose the estimated cost per component, in currency units, is

where v is cutting speed in metres per minute, and the model is considered valid for 10 ≤ v ≤ 40. Differentiating gives:

Setting the derivative to zero gives v = 25 m/min. Since C″(v) = 0.16 > 0, this is a local minimum. Checking the interval boundaries is still necessary:

• C(10) = 120 + 8 − 40 = 88
• C(25) = 120 + 50 − 100 = 70
• C(40) = 120 + 128 − 160 = 88

Thus the minimum within the stated operating range occurs at 25 m/min, with estimated cost 70 currency units per component. The result is valid only if quality, tool life, power, and safety requirements are already represented or separately verified.

Multivariable Optimization

For a differentiable function f(x₁, x₂, …, x_n), an interior stationary point satisfies:

where ∇f is the gradient vector of first partial derivatives. The Hessian matrix of second partial derivatives can help classify the point. If the Hessian is positive definite at x*, the point is a strict local minimum; if negative definite, it is a strict local maximum. These conclusions concern local behavior and do not necessarily establish a global optimum.

3. Constrained Optimization

Real engineering decisions almost always have constraints. For equality constraints, the method of Lagrange multipliers forms:

and solves ∇_xL = 0 together with h(x) = 0. The multiplier λ measures the marginal change in the optimized objective associated with a small relaxation of the constraint, subject to sign and unit conventions.

For inequality constraints, the Karush–Kuhn–Tucker conditions provide necessary conditions under suitable regularity assumptions. They include feasibility, nonnegative multipliers for correctly oriented inequalities, complementary slackness, and stationarity. In practice, software solvers are normally used, but the conditions explain why active constraints often determine the solution.

A constraint is active when it holds as an equality at the solution. For example, if machine time is limited by T and the schedule uses exactly T hours, the capacity constraint is active. Slack is the unused amount of a constraint.

4. Linear Programming

A linear programming (LP) model has a linear objective and linear constraints:

The coefficient c_j is contribution or cost per unit of activity j; a_ij is resource i used per unit of activity j; b_i is available resource i; and x_j is the activity level. Linearity assumes proportionality, additivity, divisibility, and known coefficients. These assumptions may be unrealistic when setup times, quantity discounts, learning effects, or congestion exist.

Graphical Interpretation

For two variables, each inequality defines a half-plane. Their intersection is the feasible polygon. If the feasible region is nonempty and bounded, an LP optimum occurs at an extreme point, although an entire edge may be optimal when multiple solutions exist.

Worked Example: Product Mix

A workshop makes shafts (S) and brackets (B). Each shaft contributes 50 currency units and uses 2 machining hours and 1 finishing hour. Each bracket contributes 40 currency units and uses 1 machining hour and 2 finishing hours. Available capacity is 100 machining hours and 80 finishing hours. The model is:

At the intersection of the two capacity boundaries:

From the first equation B = 100 − 2S. Substitution gives S + 2(100 − 2S) = 80, so −3S = −120 and S = 40. Therefore B = 20. The contribution is Z = 50(40) + 40(20) = 2,800 currency units.

The axis vertices are (0,0), (50,0), and (0,40), producing contributions 0, 2,500, and 1,600. Hence (40,20) is optimal for this model. Both capacity constraints are binding. The solution is meaningful only if fractional production is acceptable; if products must be scheduled in batches, an integer model may be required.

Simplex, Solver, and Sensitivity Analysis

The simplex method moves between extreme points while improving the objective. Interior-point methods approach an optimum through the interior of the feasible region. Spreadsheet, algebraic, and specialized optimization software can solve larger models, but the analyst must check units, bounds, infeasibility, and whether the solver status is optimal rather than merely feasible.

Sensitivity analysis examines the effect of changing coefficients or right-hand-side resources. A shadow price estimates the marginal improvement in the objective from one additional unit of a scarce resource, within the allowable range for which the current basis remains valid. It is not automatically a market price and should not be extrapolated beyond that range.

5. Nonlinear and Integer Programming

Nonlinear programming is required when the objective or constraints contain nonlinear terms. Examples include energy proportional to speed cubed, nonlinear reliability, economies of scale, and product quality models. Nonlinear problems may have several local optima. Convexity is especially important: for a minimization problem with a convex objective and convex feasible region, every local optimum is global.

Integer programming is appropriate when decisions are discrete. A binary machine-selection model may use y_k ∈ {0,1}, where y_k = 1 means machine k is selected. A logical constraint such as “at least one of machines A and B” can be represented by y_A + y_B ≥ 1. Big-M formulations can express conditional relationships but should use the smallest valid M because excessively large values can cause numerical instability and weak relaxations.

6. Scheduling Models

Scheduling assigns jobs to machines and time periods. Important elements include processing time, release time, due date, setup time, precedence, machine eligibility, and objective measure. Common objectives are makespan, total completion time, tardiness, flow time, and machine utilization.

For a single-machine sequence, completion time C_j is the time at which job j finishes. Tardiness is:

where d_j is the due date. Minimizing maximum completion time is minimizing makespan. Job-shop scheduling is difficult because machine order and job precedence interact; exact optimization may become computationally expensive as problem size grows.

Practical heuristics include earliest due date, shortest processing time, critical-ratio priority, and dispatching rules. Heuristics can produce good schedules quickly but do not generally prove optimality. A responsible implementation compares the rule with lower bounds, historical performance, or an optimization benchmark.

7. Inventory Models

Inventory decisions balance ordering or setup cost against holding cost and stockout risk. In the basic economic order quantity (EOQ) model, annual demand D is constant, replenishment is instantaneous, ordering cost per order is S, and annual holding cost per unit is H. Shortages are not permitted and unit price is independent of order quantity. Under these assumptions:

where Q* is the order quantity in units. Average cycle inventory is Q/2, so annual ordering plus holding cost is:

excluding purchase cost when it is constant. The EOQ minimizes this expression for Q > 0.

Worked Example

If annual demand is 12,000 units, ordering cost is 60 currency units per order, and annual holding cost is 3 currency units per unit, then:

A practical order quantity is approximately 693 units, subject to supplier pack sizes and capacity. The number of orders per year is D/Q ≈ 17.3. EOQ does not determine the reorder point when demand or lead time is uncertain.

With constant demand rate d and lead time L, the basic reorder point is R = dL. If demand during lead time has standard deviation σ_L and a service factor z is selected, a common safety-stock model is:

This assumes a specified probability model, typically approximately normal demand during lead time. The service factor must correspond to the intended cycle-service probability, and uncertainty in lead time should be included in σ_L when relevant.

8. Maintenance and Replacement Decisions

Maintenance optimization determines when to inspect, repair, replace, or operate equipment. Preventive maintenance may reduce failure probability but consumes planned downtime and resources. Corrective maintenance may be cheaper under low failure rates but can create disruption and safety risk.

A replacement model compares the expected annualized cost of keeping equipment for n periods with replacing it. A generic discounted-cost expression is:

where r is the discount rate per period. Costs, revenues, inflation treatment, and salvage values must be expressed consistently. For uncertain failures, expected cost should weight each possible failure state by its probability. Reliability, safety, environmental impact, and regulatory compliance may impose constraints that cannot be traded away solely for financial savings.

9. Multi-Criteria Decision Models

Many machine-technology decisions have conflicting criteria: cost, precision, throughput, energy, maintainability, noise, and safety. A weighted-sum model normalizes criterion scores and forms:

where a is an alternative, s_i(a) is its normalized score on criterion i, and weights satisfy w_i ≥ 0 and ∑w_i = 1. Benefit criteria are scored so larger values are better; cost criteria must be transformed so smaller costs receive higher scores. The normalization method must be stated because it can alter rankings.

For example, a decision-maker may assign weights of 0.40 to cost, 0.30 to accuracy, 0.20 to reliability, and 0.10 to energy use. The resulting ranking is conditional on these judgments. Sensitivity analysis should vary the weights and identify whether the preferred alternative changes. Safety and legal compliance are often treated as minimum requirements rather than compensable low scores.

Other methods include goal programming, which minimizes weighted deviations from target values, and Pareto analysis. An alternative is Pareto-dominated if another alternative is at least as good on every criterion and strictly better on at least one. The Pareto frontier presents efficient trade-offs without forcing one arbitrary weighting system.

10. Illustrative Case: CNC Cell Capacity and Maintenance

Background

A small CNC cell produces two components, A and B. Management must select a weekly production plan while reserving time for preventive maintenance. Component A contributes 70 currency units and requires 3 machining hours and 1 inspection hour. Component B contributes 50 currency units and requires 2 machining hours and 2 inspection hours. The cell has 120 machining hours and 70 inspection hours per week. At least 10 hours must be reserved for maintenance, leaving 110 machining hours for production.

Problem and Formulation

Let A and B be weekly quantities. The model is:

The intersection satisfies 3A + 2B = 110 and A + 2B = 70. Subtracting gives 2A = 40, so A = 20 and B = 25. Contribution is 70(20) + 50(25) = 2,650 currency units. The machining and inspection constraints are both active.

Interpretation and Implementation

The result schedules 20 units of A and 25 units of B, while preserving the maintenance reservation. Before implementation, the supervisor should confirm demand, batch sizes, setup time, tooling availability, inspection staffing, and whether maintenance can actually be completed in the reserved period. If unexpected failures occur, a rolling re-optimization model can update available capacity.

A sensitivity review should ask whether a lost inspection hour or an additional machining hour changes the plan. The model should also include minimum customer commitments if those exist. A solution that maximizes contribution but violates delivery agreements is operationally unacceptable.

Learning Points

The case demonstrates that maintenance is not merely a cost: reserving capacity can protect reliability and future availability. It also shows why optimization results require managerial validation. The mathematical optimum is conditional on the data and model structure.

11. Technology, Governance, and Limitations

Modern manufacturing execution systems, enterprise-resource-planning platforms, digital twins, and cloud solvers can provide real-time data for optimization. Artificial intelligence can forecast demand, estimate failure probabilities, or generate heuristic schedules. However, prediction and optimization are different tasks: a forecast estimates what may happen, whereas optimization selects an action under a model of consequences.

Automated decisions require traceability. Organizations should retain the data version, model formulation, solver status, parameter values, and approval record. Bias may arise when historical data reflect unequal access to maintenance, inconsistent quality inspection, or unsafe work practices. Optimization should not conceal risks by assigning arbitrary low penalties to safety or environmental outcomes. Human review is essential for exceptional conditions, uncertain data, and decisions affecting workers.

Common modeling errors include omitted constraints, double-counted resources, incorrect units, unrealistic linearity, overconfident demand estimates, and treating a local nonlinear solution as globally optimal. Validation should use historical back-testing, pilot trials, constraint audits, stress tests, and comparison with expert decisions. Robust or stochastic optimization may be appropriate when uncertainty is material, but these methods require credible probability distributions or uncertainty sets.

Best Practices for Professional Use

1. Define the decision horizon, units, and scope before writing equations.
2. Separate controllable decisions from forecasts and fixed parameters.
3. Document every assumption and test its operational plausibility.
4. Check feasibility independently of the solver.
5. Use integer variables when fractional decisions are impossible.
6. Examine sensitivity to demand, capacity, cost, and weight changes.
7. Treat safety, legal, and environmental requirements as binding where appropriate.
8. Compare the optimized plan with a practical baseline.
9. Monitor actual performance and re-estimate parameters.
10. Preserve human accountability for consequential decisions.

Chapter Summary

Optimization models select the best feasible decisions according to explicit objectives and constraints. Unconstrained calculus identifies stationary points, while constrained methods account for resource and policy limits. Linear programming is effective when relationships are proportional and additive; nonlinear and integer programming represent curvature and discrete decisions. Scheduling models assign jobs under precedence and capacity limitations. Inventory models balance ordering, holding, and stockout considerations, while maintenance models compare planned intervention with failure and replacement consequences. Multi-criteria models make trade-offs among cost, quality, reliability, energy, and safety visible. In every case, sensitivity analysis, validation, ethical governance, and operational judgment are necessary because a mathematically optimal solution is only as reliable as its data and assumptions.

Key Terms

Optimization: Selection of the best feasible decision according to an objective.

Decision variable: A controllable quantity determined by the model.

Objective function: Mathematical measure to maximize or minimize.

Constraint: A mathematical representation of a limit or requirement.

Feasible region: Set of decisions satisfying all constraints.

Active constraint: Constraint satisfied exactly at a solution.

Slack: Unused amount of a resource or inequality limit.

Linear programming: Optimization with linear objective and constraints.

Nonlinear programming: Optimization containing nonlinear objective or constraints.

Integer programming: Optimization with whole-number decision variables.

Binary variable: Variable restricted to 0 or 1.

Shadow price: Marginal value of relaxing a resource constraint.

Scheduling: Assignment and sequencing of jobs over time.

Makespan: Completion time of the final scheduled job.

Tardiness: Positive amount by which completion exceeds due date.

EOQ: Economic order quantity minimizing basic ordering and holding cost.

Safety stock: Inventory held to protect against uncertainty.

Pareto frontier: Set of non-dominated alternatives.

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

Robust optimization: Optimization designed to perform acceptably under uncertainty.

Review Questions

Short Answer Questions

1. What are the three principal components of an optimization model?
2. Why must a stationary point be checked before being called an optimum?
3. What does an active constraint indicate?
4. State the assumptions of the basic EOQ model.
5. Distinguish a continuous variable from a binary variable.

Long Answer Questions

1. Explain the formulation and interpretation of a linear programming product-mix model.
2. Compare linear, nonlinear, and integer programming in machine-technology applications.
3. Explain how scheduling objectives differ from dispatching rules.
4. Discuss the role of sensitivity analysis in responsible optimization.
5. Explain why safety and regulatory constraints should not always be treated as ordinary weighted criteria.

Analytical / Application Questions

1. A workshop earns 60 currency units per unit of product X and 45 per unit of product Y. X uses 2 machining hours and 1 inspection hour; Y uses 1 machining hour and 2 inspection hours. With 80 machining hours and 70 inspection hours available, formulate and solve the continuous product-mix LP.
2. Annual demand is 8,000 units, ordering cost is 40 currency units per order, and annual holding cost is 2.50 currency units per unit. Calculate EOQ and explain two practical reasons why the actual order quantity might differ.
3. A maintenance manager must choose among three machines using cost, accuracy, reliability, and energy criteria. Design a multi-criteria evaluation procedure and explain how you would test whether the ranking is sensitive to the assigned weights.

Multiple Choice Questions with Answers

Question 1

Which component specifies what the optimization model seeks to maximize or minimize?

A. Objective function

B. Feasible region

C. Decision variable

D. Shadow price

Correct Answer: A. Objective function

Question 2

For the function C(v) = 120 + 0.08v² − 4v, the stationary cutting speed is:

A. 10 m/min

B. 20 m/min

C. 25 m/min

D. 40 m/min

Correct Answer: C. 25 m/min

Question 3

In linear programming, a feasible solution must:

A. Make every constraint active

B. Satisfy all constraints and domain restrictions

C. Maximize every individual variable

D. Contain only integer variables

Correct Answer: B. Satisfy all constraints and domain restrictions

Question 4

In the product-mix example, the optimal production plan is:

A. 40 shafts and 20 brackets

B. 20 shafts and 40 brackets

C. 50 shafts and 0 brackets

D. 0 shafts and 40 brackets

Correct Answer: A. 40 shafts and 20 brackets

Question 5

Using EOQ with D = 12,000, S = 60, and H = 3 gives approximately:

A. 200 units

B. 490 units

C. 693 units

D. 1,200 units

Correct Answer: C. 693 units

Question 6

Tardiness for a job is calculated as:

A. dⱼ − Cⱼ

B. max(0, Cⱼ − dⱼ)

C. Cⱼ + dⱼ

D. Cⱼ/dⱼ

Correct Answer: B. max(0, Cⱼ − dⱼ)

Question 7

A binary variable is most appropriate for representing:

A. Continuous cutting speed

B. Annual demand

C. Whether a machine is selected

D. Average inventory level

Correct Answer: C. Whether a machine is selected

Question 8

A Pareto-dominated alternative is one for which another alternative is:

A. Worse on every criterion

B. Equal or better on every criterion and better on at least one

C. Cheaper but less reliable only

D. Identical in all criteria

Correct Answer: B. Equal or better on every criterion and better on at least one

Question 9

In the CNC case, the contribution at A = 20 and B = 25 is:

A. 2,150 currency units

B. 2,500 currency units

C. 2,650 currency units

D. 3,100 currency units

Correct Answer: C. 2,650 currency units

Question 10

A shadow price generally estimates:

A. The total fixed cost of a product

B. The marginal objective improvement from relaxing a resource constraint

C. The probability of machine failure

D. The number of jobs in a schedule

Correct Answer: B. The marginal objective improvement from relaxing a resource constraint