Nonlinear, Goal and Multi-Criteria Optimization

Introduction

Operations decisions frequently involve relationships that are not linear, objectives that cannot be reduced to a single measure, and stakeholders who assign different priorities to cost, speed, quality, resilience, sustainability and risk. Linear programming remains powerful, but its assumptions may be too restrictive when economies of scale, congestion, diminishing returns, nonlinear risk or service-level penalties are present. This chapter develops three connected extensions of operations research: nonlinear optimization, goal programming and multi-objective optimization.

Nonlinear optimization studies problems in which the objective function or at least one constraint is nonlinear. Goal programming treats several desired targets through weighted or lexicographically ordered deviations. Multi-objective optimization explicitly represents conflicting objectives and examines efficient, or Pareto-optimal, solutions rather than assuming that one universal optimum exists.

The central managerial issue is not merely how to solve a model. It is how to ensure that the model represents the decision context, that its assumptions are defensible, and that the resulting trade-offs are transparent.

Learning objectives

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

1. Recognize nonlinear structures in operations models.
2. Explain convex sets, convex functions and their importance for global optimality.
3. Apply first- and second-order optimality conditions, including the Karush–Kuhn–Tucker conditions.
4. Formulate weighted and pre-emptive goal-programming models.
5. Distinguish Pareto efficiency, dominance, weighting methods and ε-constraint methods.
6. Evaluate trade-offs among cost, service, sustainability, risk and strategic priorities.
7. Interpret solver output critically and communicate limitations to decision makers.

1. Nonlinear structures in operations management

A general nonlinear programming (NLP) model can be written as

Here, x is the vector of decision variables, f(x) is the objective, g_i(x) are inequality constraints, h_j(x) are equality constraints, and X describes bounds or domain restrictions. The formulation assumes that the functions are defined over the feasible domain and that the decision variables have meaningful units.

A model is nonlinear if any relevant function is nonlinear. Common operational examples include:

• Economies or diseconomies of scale: production cost may contain a term such as c(q) = aq + bq², where q is quantity and b captures increasing or decreasing marginal cost depending on its sign.
• Congestion: waiting time can rise sharply as utilization approaches capacity. A queueing approximation may have the form W = 1/(μ − λ), under restrictive assumptions, where λ is arrival rate and μ is service rate, with 0 ≤ λ < μ.
• Inventory risk: shortage or obsolescence costs may be nonlinear in inventory levels.
• Portfolio or supply risk: a quadratic variance term can represent interaction among uncertain suppliers.
• Service-level penalties: late deliveries may incur step, piecewise-linear or nonlinear penalties.
• Sustainability: emissions may vary nonlinearly with load, distance, technology and utilization.

A nonlinear expression should not be introduced simply because it is mathematically sophisticated. The functional form should reflect empirical evidence, engineering relationships, accounting rules or defensible managerial assumptions.

Continuous, discrete and mixed models

If all variables are continuous, the model is a continuous NLP. If some variables are integer or binary, it is a mixed-integer nonlinear program (MINLP), generally more difficult because it combines nonconvexity or nonlinear calculation with combinatorial search. For example, a facility-opening binary variable multiplied by a nonlinear production expression can create an MINLP.

A model may also be nonconvex. Nonconvexity arises from fixed-charge decisions, bilinear products such as xy, economies of scale with inappropriate functional forms, or discrete alternatives. Nonconvex models can contain multiple local optima, so a locally optimal solution need not be globally optimal.

2. Convexity and why it matters

Convexity provides the most important bridge between model structure and solution reliability.

A set S is convex if, for every x,y ∈ S and every θ ∈ [0,1]],

The line segment joining any two feasible points therefore remains feasible. A function f is convex on a convex domain if

for all feasible x,y and θ ∈ [0,1]. A differentiable function is convex if its gradient does not decrease too rapidly; if twice differentiable, a sufficient condition is that its Hessian matrix is positive semidefinite throughout the domain:

The Hessian ∇²f(x) contains second partial derivatives. Positive semidefinite means z^T∇²f(x)z ≥ 0 for every vector z. Strict positive definiteness, z^T∇²f(x)z > 0 for every nonzero z, implies strict convexity and usually uniqueness of the minimizer when the feasible set is convex.

For a minimization problem, a convex objective and convex feasible region are highly desirable. In a convex optimization problem, every local optimum is a global optimum. If the objective is differentiable and strictly convex, there can be at most one optimal decision vector, although multiple equivalent descriptions or redundant variables may still occur.

A differentiable function is concave when its negative is convex. Maximizing a concave objective over a convex feasible set is also a convex optimization problem in the standard optimization sense. However, maximizing a convex function or minimizing a concave function is generally nonconvex.

Convex constraints

A constraint g(x) ≤ 0 defines a convex feasible region when g is convex. Affine equalities and inequalities are both convex. Thus, a model with a convex objective, convex inequality constraints, affine equalities and convex variable domains is a convex program.

Convexity is not only a mathematical property. It affects:

• whether a local solver’s answer can be trusted as globally optimal;
• the usefulness of dual prices and sensitivity analysis;
• the stability of numerical algorithms;
• the interpretation of trade-offs;
• the need for global optimization or multiple-start procedures.

3. Optimality conditions

Unconstrained optimization

For a differentiable unconstrained minimization problem, an interior local optimum x^* must satisfy the first-order condition

This condition is necessary, not sufficient. In one dimension, a second derivative greater than zero at a stationary point indicates a strict local minimum; a second derivative less than zero indicates a local maximum. In several dimensions, the corresponding second-order test uses the Hessian. If ∇²f(x^*) is positive definite, x^* is a strict local minimum.

For example, let

where q is production volume in units and f(q) is a cost index over a relevant bounded operating range. The derivative is f'(q) = 20 + 0.1q − 0.0003q². Solving f'(q)=0 gives stationary points, but only points within the feasible production range should be considered, and the second derivative f''(q)=0.1−0.0006q must be examined. This example also demonstrates why an empirically fitted nonlinear function should not be extrapolated beyond the range in which it was estimated.

Constrained optimization and KKT conditions

For the problem

the Lagrangian is

where λ_i are multipliers for inequality constraints and ν_j are unrestricted multipliers for equality constraints. Under a suitable constraint qualification, a local optimum must satisfy the Karush–Kuhn–Tucker (KKT) conditions:

The last condition is complementary slackness. It says that an inequality constraint may have a positive shadow value only if it is active. KKT conditions are necessary under regularity conditions such as linear independence of active constraint gradients or an appropriate constraint qualification. For convex problems, with convex inequality functions and affine equality constraints, KKT conditions are also sufficient under standard conditions such as feasibility and, commonly, Slater’s condition.

The multipliers can have economic meaning. In a minimization model, the multiplier associated with a right-hand-side resource constraint often approximates the marginal change in the optimal objective when that resource is relaxed, provided regularity and local sensitivity assumptions hold. It should not automatically be interpreted as a causal effect or a valid value outside the local range.

Numerical solution approaches

Common approaches include sequential quadratic programming, interior-point methods, reduced-gradient methods, active-set methods and derivative-free methods. A solver may terminate because it has found a stationary point, because progress is small, because feasibility is inadequate, or because numerical limits have been reached. Professional practice therefore requires checking:

1. primal feasibility and constraint violations;
2. stationarity or optimality residuals;
3. variable bounds and integrality, if applicable;
4. sensitivity to starting values;
5. alternative local optima;
6. plausibility of units and operational interpretation.

For nonconvex models, multistart local optimization, spatial branch-and-bound, convex relaxations, global optimization software or carefully designed heuristics may be required. No solver label alone proves global optimality.

4. Goal programming

Many operational plans have several target levels rather than a single natural objective. Examples include total cost below a budget, emissions below a threshold, delivery time below a target and employment above a target. Goal programming converts deviations from these targets into an optimization model.

Let f_k(x) denote the achievement level of goal k, and let T_k be its target. Introduce nonnegative underachievement and overachievement variables d_k^- and d_k^+:

The deviation variables have the same units as the corresponding goal. For a “minimize cost” target, overachievement may be undesirable, whereas for a “maximize service” target, underachievement is usually undesirable. The model should penalize only the deviation that has managerial meaning.

A weighted goal-programming model is

Weights must be nonnegative and should account for both priority and measurement scale. If one deviation is measured in dollars and another in tonnes, raw weights are not comparable without normalization. A normalized deviation can be defined as (f_k(x)−T_k)/s_k, where s_k is a meaningful scale, such as the target, tolerance or monetary conversion. The choice of scale is a value judgment and should be documented.

Pre-emptive or lexicographic goal programming

In pre-emptive goal programming, priorities are ranked. All deviations at priority level 1 are minimized before any deterioration at level 2 is accepted. This approach is appropriate when a requirement is genuinely non-negotiable, such as regulatory compliance or a safety limit. It is inappropriate when priorities are only slightly different, because a strict hierarchy can hide substantial trade-offs.

Hard constraints should remain hard constraints. Treating a legal emission limit as merely a soft goal can produce an infeasible or unlawful recommendation. Goal programming is not a substitute for compliance analysis.

Illustration

Suppose a distribution plan chooses quantities x₁ and x₂, with capacity constraint x₁+x₂≤100. Management wants cost of at most 5,000 and emissions of at most 800 kg. If cost is C(x) and emissions are E(x), the soft-goal equations are

If only exceeding cost and emissions is undesirable, minimize w_Cd_C⁺ + w_Ed_E⁺. The units differ, so the weights must incorporate either monetary valuation, normalization or a stated managerial preference. A solution that exceeds the emissions goal by 100 kg to save $10,000 is not “optimal” independently of how those consequences are valued.

5. Multi-objective optimization

A multi-objective problem has a vector objective:

Usually, objectives conflict. Reducing cost may worsen delivery reliability; increasing inventory may improve service but increase working capital and obsolescence; using expedited transport may reduce lead time but increase emissions and cost.

A feasible solution x dominates y in a minimization problem if it is no worse in every objective and strictly better in at least one. A solution is Pareto efficient if no feasible solution dominates it. The set of efficient objective vectors is the Pareto frontier.

Efficiency does not mean that every efficient solution is acceptable. It means that improvement in one objective requires deterioration in at least one other objective. Decision makers must select among efficient alternatives using preferences, constraints, risk attitudes and strategic judgments.

Scalarization methods

The weighted-sum method solves

The objectives must be normalized or converted to commensurable units. In a convex problem, positive weighted sums can recover supported efficient solutions. In a nonconvex objective region, some Pareto-efficient points may not be found by any positive weighted sum.

The ε-constraint method optimizes one primary objective while imposing limits on the others:

By varying ε_k, analysts can generate a broader view of trade-offs, including unsupported efficient points. The method is especially useful when one objective, such as cost, is primary and service or emissions have explicit acceptable limits.

Other approaches include lexicographic optimization, reference-point methods, achievement scalarizing functions and interactive decision support. Interactive methods allow decision makers to learn about trade-offs rather than specifying all preferences before seeing the feasible consequences.

Marginal trade-offs

On a smooth efficient frontier, a local trade-off between objectives f₁ and f₂ can be examined through a marginal rate of substitution. Along a parameterized frontier, the slope is

where defined. Its interpretation depends on objective direction and units. A slope is not automatically a policy value; it describes a local exchange rate generated by the model. Nonlinear, discontinuous or discrete frontiers may not possess a meaningful derivative.

6. Integrated decision criteria: cost, service, sustainability and risk

Advanced operations models often combine several dimensions:

Risk can be represented through expected loss, variance, chance constraints, scenario-based measures or coherent risk measures such as conditional value at risk (CVaR). For a random loss L, the value-at-risk at confidence level α is a quantile of the loss distribution. CVaR is the expected loss in the worst (1−α) tail, subject to the chosen definition and distributional assumptions. Risk measures should not be combined with cost without specifying whether risk is monetized, constrained or treated as a separate objective.

A robust workflow is:

1. separate legal and safety constraints from preferences;
2. identify measurable objectives and their units;
3. define scenarios, uncertainty and data provenance;
4. normalize or monetize only with explicit justification;
5. generate efficient alternatives;
6. conduct sensitivity and stress tests;
7. present consequences, not only a single score.

7. Worked application: regional distribution redesign

The following is an illustrative case. It uses simplified data and is not a claim about a real company.

Background and problem

A retailer serves a region from two candidate distribution centres. It must choose annual shipment allocation x₁ and x₂ in thousands of orders. Capacity is 80 and 70 thousand orders respectively, and demand is 120 thousand orders. Management wants to minimize cost while maintaining service and reducing emissions.

Assume the simplified functions, in thousands of currency units and tonnes respectively, are

The quadratic terms represent increasing marginal handling and transport cost in this illustrative range. The model includes

The equality ensures that all demand is served. Substituting x₂=120−x₁ gives

and the feasible interval is 50≤x₁≤80. Differentiating,

The unconstrained stationary point is x₁=53.33, which lies in the feasible interval. Hence the cost-minimizing allocation is approximately x₁=53.33 and x₂=66.67. The cost is approximately 1,746.67 thousand currency units. Emissions are approximately E=0.9(53.33)+0.5(66.67)=81.0 tonnes.

Suppose management imposes an emissions target of 70 tonnes. Since

meeting the target requires 60+0.4x₁≤70, or x₁≤25. That is impossible because capacity at centre 2 limits x₁ to at least 50. Therefore the target is infeasible under the current network. This is a crucial managerial finding: optimization cannot manufacture feasibility.

The minimum-emission feasible allocation is x₁=50, x₂=70, producing 80 tonnes. Its cost is

thousand currency units. The arithmetic illustrates a substantial cost increase relative to the cost-minimizing allocation, while emissions fall by only one tonne from the cost solution. Management could instead evaluate capacity expansion, cleaner transport, demand redesign or a less stringent target. The efficient frontier, rather than either isolated solution, supports that discussion.

Analysis and outcome

The cost solution is not automatically the best strategic decision, and the minimum-emission solution is not automatically justified. The analyst should present:

• the two extreme feasible policies;
• intermediate efficient allocations;
• marginal cost per tonne of emissions reduction;
• sensitivity to cost coefficients and capacity;
• service implications and implementation constraints;
• the investment required to make the 70-tonne target feasible.

Learning points

First, a nonlinear cost function can make the best allocation interior rather than at a capacity boundary. Second, a target may be infeasible because of structural constraints. Third, multi-objective analysis reveals a trade-off that a single weighted score can conceal. Fourth, units and data interpretation are as important as solver mechanics.

8. Governance, ethics and contemporary practice

Optimization models can distribute opportunities and burdens. A routing model may systematically assign poorer service to particular communities; a workforce model may encode biased historical performance data; a sustainability objective may shift emissions upstream or to regions with weaker regulation. Governance should therefore include:

• documented data sources and version control;
• explicit objective definitions and normalization choices;
• audit trails for weights, targets and constraints;
• fairness and distributional-impact analysis;
• human review of exceptional recommendations;
• privacy and security controls for operational data;
• scenario testing against regulation and disruption;
• clear communication of uncertainty and model limitations.

AI and machine learning can estimate nonlinear demand, travel time, failure probability or emissions. They do not eliminate the need for optimization governance. Predictive accuracy does not establish causal validity, and a prediction model may reproduce historical inequities. A sound architecture separates prediction, optimization and decision review, while monitoring drift and recalibrating when operating conditions change.

Contemporary research increasingly combines data-driven surrogate models, robust and distributionally robust optimization, digital twins, explainable decision support and human-in-the-loop multi-objective optimization. These developments are valuable only when uncertainty, computational error and managerial accountability are made visible.

Challenges and limitations

Nonlinear and multi-objective models have several limitations:

1. Parameter uncertainty: small changes in coefficients can change the selected policy.
2. Functional-form risk: a convenient quadratic or exponential function may not represent actual operations.
3. Local optima: nonconvex algorithms may report a solution dependent on starting values.
4. Scale dependence: weighted objectives are sensitive to units and normalization.
5. Preference ambiguity: managers may not know their trade-offs until they see alternatives.
6. Data gaps: emissions, risk and service measures may be incomplete or inconsistent.
7. Computational burden: MINLP and robust multi-objective models can be difficult at scale.
8. Implementation gap: an apparently efficient plan may be infeasible because of labour, contracts, changeover or information constraints.

Best practice is to begin with the simplest defensible model, validate it with domain experts, compare alternative formulations, inspect the efficient frontier, and stress-test recommendations before implementation.

Chapter Summary

Nonlinear optimization represents operational relationships involving economies of scale, congestion, risk, capacity effects and nonlinear service or sustainability outcomes. Convexity is central because, under appropriate assumptions, every local optimum of a convex minimization problem is global. First-order, second-order and KKT conditions provide necessary or sufficient optimality tests depending on differentiability, constraint qualifications and convexity.

Goal programming represents target deviations using underachievement and overachievement variables. Weighted models require careful scaling, while pre-emptive models impose strict priority levels. Multi-objective optimization does not assume one naturally correct objective; it identifies efficient alternatives and supports preference-based selection. Weighted sums, ε-constraint methods, lexicographic approaches and interactive procedures each have distinct strengths and limitations.

Professional practice requires separating hard legal or safety constraints from soft preferences, validating nonlinear assumptions, checking feasibility, interpreting multipliers cautiously, and communicating trade-offs among cost, service, sustainability, risk and strategic priorities. Optimization is most valuable when it improves the quality and transparency of managerial judgment rather than replacing accountability.

Key Terms

Nonlinear programming: Optimization in which the objective or at least one constraint is nonlinear.

Convex set: A set containing every line segment joining any two of its points.

Convex function: A function whose value along a line segment does not exceed the corresponding weighted average of endpoint values.

Hessian: The matrix of second partial derivatives of a scalar function.

Positive semidefinite: A matrix satisfying z^THz≥0 for every vector z.

Nonconvex problem: An optimization problem lacking the convexity structure needed to guarantee that local optima are global.

Stationary point: A point at which the gradient of a differentiable unconstrained objective is zero.

KKT conditions: First-order constrained optimality conditions involving stationarity, feasibility, multiplier signs and complementary slackness.

Lagrangian: A function combining an objective with weighted constraints.

Shadow price: A marginal value associated with relaxing a constraint, subject to sensitivity assumptions.

Goal programming: A method that minimizes deviations from specified targets.

Deviation variable: A nonnegative variable representing underachievement or overachievement of a goal.

Lexicographic priority: An ordered hierarchy in which higher-priority goals are optimized before lower-priority goals.

Multi-objective optimization: Optimization involving two or more objectives simultaneously.

Dominance: A relation in which one feasible solution is no worse in every objective and better in at least one.

Pareto-efficient solution: A feasible solution that is not dominated by another feasible solution.

Pareto frontier: The set of objective outcomes associated with efficient solutions.

Scalarization: Conversion of multiple objectives into a single optimization criterion or constrained form.

ε-constraint method: Optimization of one objective while bounding the remaining objectives.

CVaR: Conditional value at risk, commonly interpreted as expected loss in a specified worst-case tail.

Review Questions

5 Short Answer Questions

1. What distinguishes a nonlinear program from a linear program?
2. Why does convexity matter for interpreting local optima?
3. What is the purpose of underachievement and overachievement variables in goal programming?
4. Define Pareto dominance.
5. Why should objective functions be normalized before applying weighted sums?

5 Long Answer Questions

1. Explain convex sets, convex functions and Hessian-based tests, and relate them to operations decisions.
2. State and interpret the KKT conditions, including complementary slackness and the role of constraint qualifications.
3. Compare weighted, lexicographic and ε-constraint approaches to goal and multi-objective optimization.
4. Discuss how cost, service, sustainability and risk can be integrated without obscuring units or managerial priorities.
5. Evaluate the strengths and limitations of nonlinear optimization in strategic supply-chain planning.

3 Analytical / Application Questions

1. Construct a goal-programming model for a hospital that seeks to limit cost and emissions while achieving a minimum service target. Identify which deviations should be penalized and justify the scales or weights.
2. A firm has two suppliers with different costs, lead times and emissions. Formulate a multi-objective model and explain how an ε-constraint analysis could reveal trade-offs that a weighted sum might miss.
3. Using the illustrative distribution case, examine how a change in centre-2 capacity would affect feasibility of the emissions target. State what additional data would be required before recommending network expansion.

Multiple Choice Questions with Answers

Question 1

Which condition is sufficient for a twice-differentiable function to be convex on a domain?

A. Its Hessian is positive semidefinite throughout the domain.

B. Its gradient is zero at every feasible point.

C. Its Hessian is negative definite throughout the domain.

D. Every constraint is nonlinear.

Correct Answer: A. Its Hessian is positive semidefinite throughout the domain.

Question 2

Under standard constraint qualifications, which KKT condition states that an inequality multiplier and its constraint cannot both be nonzero in the required directions?

A. Primal feasibility

B. Dual feasibility

C. Complementary slackness

D. Objective normalization

Correct Answer: C. Complementary slackness

Question 3

In goal programming, what does a positive overachievement variable usually represent?

A. Achievement above the specified target

B. A violation of every hard constraint

C. The objective-function gradient

D. A probability of disruption

Correct Answer: A. Achievement above the specified target

Question 4

A feasible solution is Pareto efficient when:

A. It minimizes every objective independently.

B. No feasible solution improves at least one objective without worsening another.

C. All objective weights are equal.

D. It is always the cheapest solution.

Correct Answer: B. No feasible solution improves at least one objective without worsening another.

Question 5

What is a principal limitation of the weighted-sum method in a nonconvex multi-objective problem?

A. It cannot represent any objective numerically.

B. It may fail to recover unsupported Pareto-efficient points.

C. It requires all variables to be binary.

D. It guarantees a unique solution.

Correct Answer: B. It may fail to recover unsupported Pareto-efficient points.

Question 6

In the illustrative distribution case, demand is 120, centre 1 has capacity 80 and centre 2 has capacity 70. What is the minimum feasible value of x₁?

A. 0

B. 25

C. 50

D. 70

Correct Answer: C. 50

Question 7

In the illustrative case, emissions are E = 60 + 0.4x₁. What emissions result at x₁ = 50?

A. 70 tonnes

B. 75 tonnes

C. 80 tonnes

D. 90 tonnes

Correct Answer: C. 80 tonnes

Question 8

Which objective direction is conventionally compatible with a convex optimization problem?

A. Minimizing a convex function over a convex feasible set

B. Minimizing a concave function over a convex feasible set

C. Maximizing a convex function over a convex feasible set

D. Maximizing an arbitrary function without constraints

Correct Answer: A. Minimizing a convex function over a convex feasible set

Question 9

Why should a legal emissions limit normally be represented as a hard constraint rather than only as a soft goal?

A. Hard constraints require no data.

B. Soft goals always produce infeasible models.

C. Compliance cannot be traded away through arbitrary deviation weights.

D. Legal limits are necessarily nonlinear.

Correct Answer: C. Compliance cannot be traded away through arbitrary deviation weights.

Question 10

What does an ε-constraint method do?

A. Removes all secondary objectives.

B. Optimizes one objective while imposing bounds on other objectives.

C. Replaces every constraint with an equality.

D. Uses only integer variables.

Correct Answer: B. Optimizes one objective while imposing bounds on other objectives.