Multiobjective optimization helps you choose when improving one outcome can worsen another: a design may become cheaper but slower, or a model more accurate but more expensive to run. The gold standard is to identify feasible alternatives, remove genuinely dominated choices, expose the remaining trade-offs, and state the preferences or limits used to select an action. A Pareto frontier describes efficient compromises; it does not establish a universally best decision.
This guide builds the subject from small calculations you can check by hand, then extends it to continuous models, unsupported alternatives, uncertain measurements and numerical search. All numerical cases are original teaching examples with deliberately simplified assumptions. They illustrate reasoning rather than provide estimates for a particular organisation, machine or public service.
For wider reading, use the Gold Standard master, the Education guide to learning that transfers and the How X Works hub. Follow the chapters sequentially for the mathematics, or choose the route closest to the decision you need to make.
Choose your multiobjective-optimization route
Chapters 1–3: Define decisions and trade-offs
Chapters 4–6: Use scores and recognise their limits
Chapters 7–9: Set thresholds and solve continuous models
Chapters 10–12: Compute, test uncertainty and repair
Chapter index
Define decisions and trade-offs · Chapters 1–3
1. Define the decision before combining its outcomes
Use scores and recognise their limits · Chapters 4–6
4. Derive a weighted score and test what its weights mean
5. Repair scale effects before interpreting preferences
6. Find the efficient alternative that every weighted sum misses
Set thresholds and solve continuous models · Chapters 7–9
7. Replace a preference weight with a meaningful threshold
8. Solve a continuous frontier from first principles
9. Compare compromise rules without disguising their differences
Compute, test uncertainty and repair · Chapters 10–12
10. Choose numerical methods according to the feasible decisions
11. Keep uncertainty separate from preference
12. Diagnose misleading results and repair the actual failure
Transfer and continue · Chapters 13–14
CHAPTER 1 OF 14 · Define decisions and trade-offs
1. Define the decision before combining its outcomes
Suppose a team must choose one of four designs for a small delivery system. Each design has an estimated weekly operating cost and average delay. For this example, cost is measured in hundreds of dollars and delay in minutes. Lower is better for both. Design A has outcomes (2, 10), B has (5, 6), C has (9, 3), and D has (6, 8). Assume each is technically feasible and the estimates describe the same workload and evaluation period.
The decision variable is the design label. The two objective functions return its cost and delay. These are different objects: a design is something the team can implement, while its objective vector describes what the model predicts. Keeping that distinction prevents a common mistake: treating a desirable pair of numbers as though it were an available design.
For instance, the cheapest cost is 2 and the shortest delay is 3, but no listed design achieves (2, 3). Combining the separate minima invents an alternative. A useful optimisation model must retain the connection between each decision and all its consequences.
Feasibility comes first
Now suppose a design cannot operate with the available power supply. It must be excluded before comparing its cost and delay. An attractive score does not make an impossible design implementable. Similarly, if the delivery system has a firm maximum permitted delay, that limit belongs in the feasible set rather than being silently exchanged for savings.
A practical model therefore records three things separately: what can be chosen, what must hold, and what should improve. Write the units, direction of improvement, evaluation horizon and measurement method beside each objective. If one quantity is a maximum rather than an average, say so explicitly. A frontier built from inconsistent definitions can look mathematically precise while answering no coherent question.
CHAPTER 2 OF 14 · Define decisions and trade-offs
2. Use dominance to remove avoidable losses
For minimisation, design P dominates design Q when P is no worse in every objective and strictly better in at least one. In the delivery example, B dominates D: its cost is 5 instead of 6 and its delay is 6 instead of 8. If these two outcomes capture everything relevant and both estimates are dependable, choosing D sacrifices both outcomes without compensation.
A, B and C do not dominate one another. Moving from A to B reduces delay by 4 minutes but adds 3 cost units. Moving from B to C reduces delay by another 3 minutes but adds 4 cost units. These three designs form the nondominated set among the four candidates. Their outcome vectors form the corresponding frontier in objective space.
The qualification “among the four candidates” matters. A fifth feasible design E with outcomes (4, 5) would dominate B and D. An option can be nondominated in a search result and still be dominated by an undiscovered feasible alternative. A finite list is a record of what was compared, not automatically a proof of global efficiency.
Weak efficiency is a different condition
A weakly Pareto-efficient point has no feasible competitor that improves every objective strictly. This is a weaker requirement. Compare P = (1, 2) and Q = (1, 3). P dominates Q under ordinary Pareto dominance, but it does not improve the first coordinate strictly. If these are the only alternatives, Q is weakly efficient yet not Pareto-efficient.
This distinction becomes useful when an optimisation method ignores an objective or stops at a weak guarantee. Always specify which notion is being claimed. Also retain distinct designs with identical objective vectors when their implementation differs: identical measured outcomes do not imply identical maintainability, ownership or suitability under changed conditions.
CHAPTER 3 OF 14 · Define decisions and trade-offs
3. Read the frontier as an exchange, not a ranking
The three efficient delivery designs answer different priorities. A suits a decision that gives little value to reducing delay; C suits one that gives delay reduction considerable value. B occupies the middle of the available trade-off. Calling B “balanced” adds a preference judgement that dominance alone does not supply.
The incremental cost of moving from A to B is 3/4 = 0.75 cost units per minute of average delay reduction. From B to C it is 4/3, approximately 1.33 units per minute. These ratios describe the specific moves between these designs. They are not universal prices of time and should not be transferred to a different workload without checking the model.
If the decision-maker will pay up to one cost unit for each minute saved, B is preferable to A: the four minutes saved are worth four units under that rule, exceeding the additional cost of three. Moving from B to C fails the same test because three minutes are worth three units but cost four. The choice follows an explicit exchange rate.
Do not infer missing designs by drawing lines
A straight line between A and C can be useful visually, but points along it need not be feasible. You cannot necessarily build half of each indivisible design. Randomly choosing A in some weeks and C in others might create an average outcome, but it also creates variability, and the service requirement may apply every week rather than on average.
Before interpolating, establish what physical or operational mixing means. The same issue arises in schedules, machine configurations and experimental designs. The existing guide to cross-criterion relaxation frontiers develops the related problem of comparing changes that have different kinds of cost.
CHAPTER 4 OF 14 · Use scores and recognise their limits
4. Derive a weighted score and test what its weights mean
Consider the score S = cost + λ × delay, where λ is nonnegative and measured in cost units per minute. For A, B and C the scores are 2 + 10λ, 5 + 6λ and 9 + 3λ. At λ = 1, these are 12, 11 and 12, so B is selected.
To find the switch between A and B, solve 2 + 10λ = 5 + 6λ. This gives λ = 3/4. For B and C, solve 5 + 6λ = 9 + 3λ, giving λ = 4/3. Comparing all three confirms that A wins below 3/4, B wins between 3/4 and 4/3, and C wins above 4/3. At either boundary the adjacent designs tie.
Notice that A and C also tie at λ = 1, but neither wins there because B has the smaller score. Pairwise intersections alone do not identify the optimum. You must compare the lowest score across all feasible alternatives, including options that were not part of the pair used to calculate a threshold.
A useful guarantee, with conditions
With strictly positive weights on every objective, a global minimiser of the weighted sum is Pareto-efficient. If another feasible point dominated it, the weighted sum would be strictly smaller, contradicting global optimality. Zero weights require extra care because an ignored outcome may worsen without changing the score. This scalarisation result and its convex-case qualifications are treated in Boyd and Vandenberghe’s Convex Optimization, section 4.7.
The guarantee does not say that the weights are ethically appropriate, that the model includes every consequence, or that a numerical solver found the global minimum. Those are separate questions. A score is most useful when its units and interpretation are understandable enough for someone to challenge the exchange it makes.
CHAPTER 5 OF 14 · Use scores and recognise their limits
5. Repair scale effects before interpreting preferences
Suppose we rewrite the delivery cost in dollars rather than hundreds of dollars. B’s cost becomes 500. If we keep the numeric coefficient λ = 1 beside delay, we have changed the decision rule from valuing one minute at 100 dollars to valuing it at one dollar. The data describe the same designs, but the unchanged coefficient now expresses a different preference.
One remedy is dimensional consistency: scale λ along with the cost units. Another is to define dimensionless objectives using fixed reference values. For example, cost relative to 10 cost units and delay relative to 10 minutes gives c/10 and d/10. Equal weights then value an equal fraction of these chosen reference scales equally. The reference scales are part of the judgement; they are not neutral merely because the resulting numbers lack units.
Keep reference values stable during comparisons
Normalising each experimental run by its own observed minimum and maximum can change scores even when an existing design has not changed. Imagine adding an extremely expensive alternative to the candidate list. A range-based cost denominator grows, so cost differences among the old designs shrink after normalisation. Their ranking can change because of the reference range, not because their performance improved.
Use physically meaningful scales or a documented fixed comparison set when stability matters. When a range is zero, do not divide by it: the objective is constant in that sample and needs separate handling. A sample minimum is also only an estimate of the ideal coordinate, not proof that the true minimum has been found.
The official pymoo decision-making tutorial demonstrates normalisation and distinguishes computed approximations of ideal and nadir values. Its practical lesson is to examine scales before interpreting distances or compromise scores. Preserve original-unit outcomes alongside every normalised score so readers can recover the concrete consequences.
CHAPTER 6 OF 14 · Use scores and recognise their limits
6. Find the efficient alternative that every weighted sum misses
Weighted sums are useful, but they do not always recover every Pareto-efficient alternative. Consider a separate discrete example with exactly three feasible objective vectors: A = (0, 4), B = (2, 3), and C = (4, 0). Both coordinates are to be minimised. Each alternative is nondominated because improving one coordinate requires worsening the other.
Use the weighted score αf1 + (1 − α)f2 for 0 ≤ α ≤ 1. The scores are 4 − 4α for A, 3 − α for B, and 4α for C. For B to beat or tie A, we need 3 − α ≤ 4 − 4α, which simplifies to α ≤ 1/3. For B to beat or tie C, we need 3 − α ≤ 4α, so α ≥ 3/5.
No α satisfies both inequalities. Therefore B never minimises this weighted sum, even though it is Pareto-efficient. Sweeping a million weights cannot fix the problem. The failure comes from the geometry of the feasible outcomes, not from an insufficiently fine weight grid.
Why the middle point is unsupported
The midpoint of A and C is (2, 2), which would dominate B if it were feasible. But it is not one of our three allowed decisions. A weighted linear score behaves as though the straight-line support of the attainable set matters; it can pass over an efficient point lying above that connecting segment.
The calculation provides a diagnostic. If discrete or nonconvex choices are involved, do not interpret a weighted-sum sweep as a complete frontier without an additional argument. Try a different scalarisation, constrained formulation or exact enumeration when possible. Changing the optimisation method may reveal an option that was hidden by the scoring method rather than absent from the real decision set.
CHAPTER 7 OF 14 · Set thresholds and solve continuous models
7. Replace a preference weight with a meaningful threshold
The epsilon-constraint approach keeps one objective as the quantity to minimise and places limits on the others. In the three-point example from chapter 6, minimise f1 subject to f2 ≤ 3. A fails the limit because f2 = 4. B and C satisfy it, and B has the smaller first objective, 2 rather than 4. The method therefore recovers the efficient alternative that weighted sums missed.
Vary the threshold to see the full response. If the limit is at least 4, A is selected. If it is at least 3 but below 4, B is selected. If it is at least 0 but below 3, C is the only feasible choice. A negative limit leaves no feasible choice. These discontinuous changes are a property of the discrete feasible set.
Resolve ties before calling a result efficient
Minimising f1 under thresholds can return an alternative tied on f1 but worse on another objective. A second optimisation can minimise f2 among the first-stage minimisers. In a two-objective problem, this lexicographic tie-breaking removes that avoidable loss when both stages are solved globally. For more objectives, further checks or tie-breaking stages may be needed.
A threshold is especially useful when a decision-maker can state an acceptable outcome more clearly than a willingness-to-trade coefficient. “Average delay must not exceed six minutes” is a different instruction from “one minute is worth one cost unit.” The first rejects violations; the second can accept them if other gains compensate.
The related specialist article on frontier selection under loss functions, thresholds and regret examines how such rules turn comparisons into actions. Use it when the difficult part is justifying the choice rule rather than calculating the feasible trade-offs.
CHAPTER 8 OF 14 · Set thresholds and solve continuous models
8. Solve a continuous frontier from first principles
Let a decision x vary continuously between 0 and 4. Define f1(x) = x² and f2(x) = (4 − x)². Treat these as dimensionless loss measures. Increasing x raises the first loss and lowers the second. Thus every x in the interval is Pareto-efficient: any different x improves one loss and worsens the other.
The separate minima are f1 = 0 at x = 0 and f2 = 0 at x = 4. The ideal vector is (0, 0), but it is unattainable. At x = 2 the losses are (4, 4); at x = 1 they are (1, 9); at x = 3 they are (9, 1). These values make the trade-off visible without relying on a plotting package.
Calculate the weighted optimum
Minimise J(x) = w x² + (1 − w)(4 − x)², where 0 ≤ w ≤ 1. Differentiation gives J′(x) = 2x − 8(1 − w). Its zero is x = 4(1 − w), which lies in the allowed interval. Because J″(x) = 2, this is the unique global minimiser.
At w = 1/4, x = 3 and the objective vector is (9, 1). At w = 1/2, x = 2 and the vector is (4, 4). At w = 3/4, x = 1 and the vector is (1, 9). Increasing the weight on the first loss reduces x, as it should. This directional check catches a surprisingly common mistake: reversing which objective a software weight favours.
Equal increments of w produce equal increments of x here, but not equal spacing along the curved frontier in objective space. In other models even the decision-space spacing can be highly uneven. A tidy weight grid is not evidence of uniformly good frontier coverage. Inspect the outcomes themselves, especially near a threshold that matters operationally.
CHAPTER 9 OF 14 · Set thresholds and solve continuous models
9. Compare compromise rules without disguising their differences
For the continuous example, suppose we want to minimise the worse of the two normalised losses. Use T(x) = max{x²/16, (4 − x)²/16}. Below x = 2, the second term is larger and falls as x increases. Above x = 2, the first term is larger and rises. Therefore T is minimised at x = 2, where both terms equal 1/4.
This maximum-deviation rule produces the same decision as equal weighted sums in this symmetric example. That agreement is a consequence of the model’s symmetry, not proof that the two methods are interchangeable. If scales, target values or weights change, their preferences can differ substantially.
Lexicographic priorities produce another answer
If the instruction is to minimise f1 first and consider f2 only among ties, the unique choice is x = 0, with losses (0, 16). Giving f1 priority is more demanding than merely assigning it a larger finite weight. It allows no deterioration in the first-stage optimum to improve the second loss.
If instead the decision-maker permits f1 to be at most one unit above its optimum, the second stage becomes: minimise (4 − x)² subject to x² ≤ 1 and 0 ≤ x ≤ 4. The feasible interval is [0, 1], so x = 1 is selected and the losses are (1, 9). The tolerance has a clear operational meaning and a visible cost.
Pymoo’s decomposition documentation distinguishes weighted sums, Tchebycheff functions and achievement scalarisations. When translating a mathematical rule into software, check the exact formula: a parameter named “weight” may multiply a deviation in one method and divide it in another. Verify a small hand-solvable case before interpreting the output of a large model.
CHAPTER 10 OF 14 · Compute, test uncertainty and repair
10. Choose numerical methods according to the feasible decisions
For a small finite set, enumerate every feasible alternative and compare objective vectors directly. This is often the clearest baseline. For a continuous model with suitable convex structure, repeated scalar optimisation can provide solutions with strong optimality evidence. For integer or nonconvex models, a global method may be needed if a complete frontier or certified objective bounds matter.
When evaluations are expensive or the model is available only as a simulation, population-based search can offer a practical approximation. NSGA-II is a prominent example. Deb and colleagues’ original 2002 paper describes an elitist method that combines parent and offspring candidates and uses nondominated sorting and spread information in selection. Its computational experiments are evidence about tested problems, not a general certificate for every future application.
The official pymoo NSGA-II implementation guide explains frontwise survival and crowding-distance selection when a front must be split. Crowding encourages diversity in objective space; it does not measure whether a design is desirable to the eventual decision-maker.
Preserve what the algorithm actually established
A returned set may be nondominated among evaluated candidates without being globally Pareto-optimal. Record evaluation budgets, feasibility tolerances, random seeds, stopping conditions and the model version. Rerun stochastic searches when variability could alter the decision, and compare against hand calculations, known special cases or independently obtained bounds.
Use variables that represent implementable choices. Rounding a continuous design into an integer schedule after optimisation can violate constraints or destroy its apparent advantage. Evaluate the rounded or repaired design again. If repair changes the decision, its repaired outcomes are the relevant ones. The frontier should describe the design you can actually execute, not an intermediate representation that exists only inside the solver.
Measure coverage with a fixed reference
A frontier-quality measure can help compare numerical runs, provided its definition stays fixed. For a two-objective minimisation example, hypervolume is the area dominated by the returned outcomes and bounded by a reference point that is worse in both coordinates. The pymoo performance-indicator documentation explains this construction and other quality measures.
For our delivery frontier A = (2, 10), B = (5, 6), C = (9, 3), choose reference (10, 12). Split the dominated region into non-overlapping vertical strips. Between cost 2 and 5, the area is (5 − 2)(12 − 10) = 6. Between 5 and 9, it is (9 − 5)(12 − 6) = 24. Between 9 and 10, it is (10 − 9)(12 − 3) = 9. The total is 39 square units in this objective coordinate system.
Returning only B and C gives area 33, so A contributes six additional units of coverage under this reference. That does not mean A contributes six units of social value. Hypervolume measures a geometric property of the set; it does not replace the decision-maker’s preferences. Changing units or the reference changes the calculation, so hold them fixed when comparing runs.
CHAPTER 11 OF 14 · Compute, test uncertainty and repair
11. Keep uncertainty separate from preference
Return to the delivery designs. B was better than D on both stated outcomes. Now suppose B’s delay estimate is 6 ± 1 minutes and D’s is 8 ± 1 minutes, interpreted as specified deterministic uncertainty intervals. Their ranges are [5, 7] and [7, 9]. If costs remain exactly 5 and 6, respectively, B is no worse in delay for every independent pairing of values within those intervals and strictly cheaper. This is a strong interval-based dominance conclusion under the stated assumptions.
If B’s interval expands to [5, 9], that conclusion fails: B could have delay 9 while D has delay 7. B’s nominal delay remains smaller, but nominal dominance no longer establishes dominance across the full uncertainty ranges. It may still be preferable under an explicit probabilistic or robustness rule; the rule must be supplied.
Do not turn confidence intervals into guarantees
Statistical intervals require their own interpretation. Separate intervals may not provide simultaneous coverage for every objective and alternative. Outcomes measured under the same workload may also be correlated, making arbitrary independent combinations implausible. A joint scenario model can preserve these dependencies better than a box of unrelated ranges.
Changing weights addresses preferences. Changing uncertainty sets or probability distributions addresses beliefs about possible outcomes. A broad weight sweep does not repair inaccurate measurements, and a large simulation sample does not decide how much one outcome should matter. Keep these two sources of variation visible in separate sensitivity analyses.
Continue to Robust Optimization when feasibility must hold over a specified uncertainty set, or Stochastic Programming when probabilities and decisions made after observing outcomes are central. Multiobjective modelling can be combined with either, but the resulting model must state both its trade-off rule and its treatment of uncertainty.
CHAPTER 12 OF 14 · Compute, test uncertainty and repair
12. Diagnose misleading results and repair the actual failure
If nearly every candidate is nondominated, first inspect the number and relevance of objectives. Adding more objectives makes domination harder because a competitor must improve or match more coordinates. Some objectives may duplicate one another, conflict only through measurement noise, or measure consequences outside the decision’s intended scope. Do not delete important outcomes merely to simplify the picture; investigate why the comparison has become uninformative.
If the frontier looks unusually favourable, audit feasibility and evaluation consistency. Were all designs tested under the same demand, time horizon and failure assumptions? Did one simulation silently omit a cost? Are maximisation objectives converted consistently when the solver expects minimisation? A sign error can create a smooth, attractive frontier pointing in exactly the wrong direction.
Check the proposed improvement in original units
Suppose a report says design B improves the composite score by 8%. Ask which individual outcomes changed and by how much. A percentage of a score with an arbitrary offset can be meaningless. Dominance, threshold compliance and original-unit differences are often easier to interpret than a relative change in an artificial scalar.
If a search reports no feasible alternatives, distinguish genuine incompatibility from numerical failure. Test a known feasible design, inspect each violated constraint and check scaling. Relaxing a real requirement is a substantive decision; loosening a numerical tolerance is a computational one. They can have different consequences even when both make a solver return a result.
Finally, ask whether the chosen objectives include the people and consequences that matter. A design can be Pareto-efficient for cost and average delay while producing unacceptable delays for a small group. Add or constrain the relevant distributional outcome explicitly. Mathematical efficiency is always relative to the feasible set and objectives that were modelled; it is not a certificate of completeness or fairness.
CHAPTER 13 OF 14 · Transfer and continue
13. Transfer the reasoning when the system changes
Try this changed-context exercise before reading the answer. The continuous model now has f1(x) = x² and f2(x) = (6 − x)², with 0 ≤ x ≤ 4. Find the equal-weight optimum. Then impose a firm limit f1 ≤ 4 and minimise f2. Finally, explain why copying the previous model’s x = 2 recommendation would be unjustified without recalculation.
For equal weights, J(x) = 0.5x² + 0.5(6 − x)². Its derivative is 2x − 6, so x = 3 is the unique minimiser on [0, 4]. The objective vector is (9, 9). At the old choice x = 2, the vector is (4, 16). Neither dominates the other: the changed choice reduces the second loss while increasing the first.
Under the firm limit x² ≤ 4, the feasible interval becomes [0, 2]. The second loss decreases throughout that interval, so the constrained optimum is x = 2, with losses (4, 16). The old number reappears, but for a different reason. It is now determined by a binding constraint rather than by the equal-weight symmetry of the earlier model.
Test an assumption change, not just a coefficient change
Suppose x must also be an integer. In this particular exercise both answers remain feasible integers, so they remain valid. That outcome should be checked, not presumed. In a nearby model with an unconstrained optimum of 2.6, an integer restriction would require comparing feasible integer choices and recalculating their outcomes.
A learner who understands the subject can explain the decision variable, identify the changed relationship, derive the new feasible set and justify the selected option. Merely remembering “choose the midpoint” fails as soon as the model or priorities change. This is the practical connection to the Education guide: transfer is demonstrated by preserving the reasoning while revising the assumptions and answer.
CHAPTER 14 OF 14 · Transfer and continue
14. Answer common questions and build a useful next project
Does a Pareto frontier choose the best option?
No. It removes alternatives that are inferior on the specified objectives, subject to the comparison’s scope. Selecting among efficient choices still needs preferences, thresholds, priorities or another explicit decision rule. If stakeholders disagree about that rule, the disagreement remains real even after the optimisation is correct.
Is the middle of the frontier the fairest compromise?
Not necessarily. “Middle” changes with units, plotting scales and the definition of distance. Fairness may require modelling who receives benefits and who bears costs. A geometric midpoint has no automatic claim to represent those considerations.
Can I use equal weights when preferences are unknown?
You can explore equal weights as a stated scenario, but they are still a preference assumption. Report the objective scales and compare nearby weights or meaningful thresholds. When a small change reverses the choice, present that instability rather than concealing it behind a single recommended point.
Is multiobjective optimisation the same as bilevel optimisation?
No. Multiple objectives describe several outcomes to improve. Bilevel optimisation imposes a hierarchy in which one decision depends on another decision-maker’s optimisation response. A problem can contain both structures. The Bilevel Optimization guide develops the response constraint through a separate worked example.
What should a beginner implement first?
Start with the four delivery designs in chapter 1. Store each label, cost and delay. Write a pairwise dominance check, retain A, B and C, and reproduce the score switches at 3/4 and 4/3. Then add the unsupported three-point example and show that an epsilon constraint selects B even though a weighted-sum sweep never does.
For a real project, prepare a compact decision record: feasible choices, objective definitions, original-unit results, measurement assumptions, rejected dominated options, the selection rule and the conditions that would trigger reconsideration. Have someone else reproduce one calculation and challenge one assumption. The useful outcome is a decision whose consequences and reasoning remain visible when the environment changes.