THE GOLD STANDARD OF JUDGEMENT · SPECIALIST STUDY
Aggregate ordered evidence without pretending its gaps are measurable distances
Many decisions use ordered ratings—poor, acceptable, good, excellent—where the order is meaningful but the numerical distance between levels is not. The Sugeno integral combines such ordinal evaluations with a capacity that can represent the importance of groups of criteria. It uses minimum and maximum operations rather than addition. That makes it useful when thresholds and coalitions matter, but it also makes the method non-compensatory: an excellent score may not cancel a critical weakness.
This guide belongs to the Judgement subject guide, the Gold Standard of Education, and the How X Works learning library. Choose a reading route, work the examples, then use the transfer checks to test whether the method survives a changed context.
Choose a reading route
Build the foundations · chapters 1–2
Apply the core method · chapters 3–5
Test evidence and alternatives · chapters 6–7
Handle limits and disagreement · chapters 8–9
Transfer and govern · chapters 10
The whole problem at a glance
| Element | Role | Question to ask |
|---|---|---|
| Ordinal scale | Orders evaluation levels | Are differences meaningful, or only rank? |
| Capacity | Assigns importance to subsets of criteria | Which coalitions are sufficient or influential? |
| Min operation | Limits a coalition by its weakest relevant evaluation or capacity level | What bottleneck remains? |
| Max operation | Chooses the strongest supported threshold | Which level has the best combined support? |
Do not import arithmetic meaning into ordinal labels. The method relies on order, min and max, plus a monotone capacity.
CHAPTER 1 OF 10 · Build the foundations
1. Start with an ordinal decision scale
Core idea. An ordinal scale tells us that one level is higher than another, not how much higher. Labels such as unacceptable, acceptable, good and excellent support order comparisons but do not justify averaging as if adjacent gaps were equal. Sugeno aggregation is attractive because it operates with order, minimum and maximum.
Method. Define a finite or bounded ordered set, explain every level with observable anchors, and test whether assessors can distinguish adjacent levels. Map all criterion evaluations and capacity values onto the same ordered scale or a defensible compatible lattice.
Worked example. A school-safety review rates evacuation clarity as poor, acceptable, good or excellent. Coding those levels 1–4 helps software sort them, but the difference from poor to acceptable is not proven equal to the difference from good to excellent. A Sugeno model uses their order without relying on that false distance.
Changed-context transfer. For medical triage, categories may be ordered by urgency but carry special rules and incomparable dimensions. A simple chain may be insufficient; the designer may need a richer lattice or an explicit rule system.
Limits. Ordinal labels can still be vague, culturally contested or too coarse. The Sugeno integral does not repair weak rubrics or biased observation.
Misconception to avoid. Numbers assigned to ranks become cardinal measurements. Numeric codes remain labels unless the measurement process supports meaningful differences and arithmetic.
Practical evidence check
- Ask three assessors to rate boundary cases and explain their choices; revise anchors before fitting any aggregation model.
- Return to chapter contents
CHAPTER 2 OF 10 · Build the foundations
2. Represent coalition importance with a capacity
Core idea. A capacity is a monotone set function: the empty set receives the bottom value, the full set the top value, and adding criteria cannot reduce capacity. Unlike additive weights, it can assign importance directly to groups, allowing interaction, redundancy and sufficiency to be represented.
Method. List all criteria and the coalitions that have decision meaning. Elicit or learn capacity values subject to boundary and monotonicity constraints. Explain why each important subset matters, and avoid estimating an unrestricted capacity when evidence is too thin.
Worked example. Safety and accessibility may each be important, but the coalition containing both can receive a higher level because a shelter is not operationally acceptable unless it protects and admits users. Capacity expresses the group requirement without pretending the two ratings add linearly.
Changed-context transfer. In supplier selection, quality and traceability may be partly redundant because both reflect process control. The coalition value need not equal the sum of individual weights; it reflects joint decision significance.
Limits. A general capacity has one value for every subset, so parameters grow exponentially. Elicitation can overwhelm decision-makers and create overfitting.
Misconception to avoid. A capacity is just a set of ordinary weights. Additive weights are a special case; general capacities describe subsets and interactions.
Practical evidence check
- Check every proposed capacity for boundary values and monotonicity, then ask whether its largest jumps correspond to intelligible coalitions.
- Return to chapter contents
CHAPTER 3 OF 10 · Apply the core method
3. Compute the Sugeno integral with threshold sets
Core idea. For each evaluation level, consider the criteria meeting or exceeding that threshold. Combine the threshold level with the capacity of that coalition using minimum, then take the maximum across thresholds. The result is the highest level jointly supported by performance and coalition importance.
Method. Sort criterion evaluations, form upper-level sets, read their capacity values, compute the minimum for each candidate threshold and select the maximum. Retain the threshold and coalition that attain the result so the score remains explainable.
Worked example. A candidate has evaluations acceptable, good and excellent. At the good threshold, two criteria remain and their coalition capacity is excellent, so the candidate support is min(good, excellent)=good. At the excellent threshold, one criterion remains with capacity acceptable, so support falls to acceptable. The maximum supported result is good.
Changed-context transfer. Transfer to a rule-like dashboard. Each threshold set can be displayed as “criteria clearing this bar,” making the aggregation easier to audit than a black-box total.
Limits. Several thresholds can tie, and a coarse scale can give many alternatives the same result. The integral supplies an order-consistent summary, not guaranteed fine ranking.
Misconception to avoid. The highest individual rating becomes the Sugeno result. A high criterion matters only through the capacity of the coalition that supports its threshold.
Practical evidence check
- Calculate every threshold row by hand for one alternative and identify all coalitions that tie for the maximum.
- Return to chapter contents
CHAPTER 4 OF 10 · Apply the core method
4. Understand max-min logic and idempotence
Core idea. Sugeno integrals are built from lattice operations. If every criterion has the same level, the aggregate returns that level, a property called idempotence. The result stays between the minimum and maximum evaluations, and monotonicity means improving a criterion cannot lower the aggregate under a fixed capacity.
Method. Use these properties as validation tests. Feed constant vectors, one-step improvements and extreme profiles. Check that the implementation respects order and that transformations preserving the ordinal scale also preserve conclusions.
Worked example. If every shelter criterion is good, the result must be good. If accessibility improves from acceptable to excellent while everything else stays fixed, the result cannot decline. A failed test signals a coding, scale or capacity error.
Changed-context transfer. For text labels, a strictly increasing relabelling from bronze/silver/gold to low/medium/high should preserve the order-based result. The method does not require distances between codes.
Limits. Monotonicity can preserve a tie even after a meaningful improvement because the decisive threshold or coalition did not change. That can be appropriate or frustrating, and should be visible.
Misconception to avoid. Any criterion improvement must raise the score. Monotonicity promises no decrease, not a strict increase.
Practical evidence check
- Run constant, dominance, boundary and relabelling tests before interpreting real alternatives.
- Return to chapter contents
CHAPTER 5 OF 10 · Apply the core method
5. Use non-compensation and bottlenecks deliberately
Core idea. Because aggregation uses min and max rather than addition, high performance on one criterion may not compensate for a low critical criterion. This suits decisions with thresholds, veto-like concerns or qualitative sufficiency, but only when those decision attitudes are intended and transparent.
Method. Identify which weaknesses should remain visible, which coalitions are sufficient and whether any criterion requires an explicit veto outside the integral. Compare the Sugeno result with a compensatory model and explain disagreements through the decisive threshold.
Worked example. A venue has excellent cost and location but unacceptable structural safety. A weighted average might still look moderate if safety receives too little weight. A Sugeno capacity that makes safety-containing coalitions decisive prevents unrelated excellence from washing out the bottleneck.
Changed-context transfer. In creative portfolio review, strict non-compensation may be inappropriate because exceptional originality can legitimately offset polish. The method choice should follow the decision philosophy, not mathematical fashion.
Limits. Sugeno aggregation does not automatically implement every veto or legal minimum. A hard constraint may need explicit pre-screening.
Misconception to avoid. Non-compensatory means fair. It can protect essential thresholds, but it can also entrench arbitrary gates or reduce opportunities for strengths to matter.
Practical evidence check
- Name every criterion whose poor performance should block or cap the decision and justify that rule to affected people.
- Return to chapter contents
Worked emergency-shelter comparison
| Criterion | Shelter A | Shelter B | Interpretation |
|---|---|---|---|
| Structural safety | Excellent | Good | A leads on a critical condition |
| Accessibility | Acceptable | Excellent | B leads for inclusive use |
| Travel time | Good | Good | Tie |
| Capacity coalition | Safety plus accessibility has high capacity | Same capacity model | Both must clear a joint threshold |
The final score depends on the declared ordinal scale and capacity. It should be accompanied by the criterion vector, decisive coalition and sensitivity checks rather than reported as a mysterious rank.
CHAPTER 6 OF 10 · Test evidence and alternatives
6. Interpret capacities without overclaiming causality
Core idea. Capacity values express decision importance for subsets under the model. They do not by themselves prove causal influence, moral worth or statistical interaction. Interpretation should focus on how the aggregation responds to coalitions and thresholds.
Method. Use sensitivity, marginal contribution descriptions and decision examples. Compare capacities of nested and related subsets. Where specialized interaction indices are used, explain their definition and scale rather than labelling a positive value “synergy” without context.
Worked example. If the capacity of safety plus accessibility is much higher than either alone, the model treats their coalition as influential. That can represent a policy of joint sufficiency; it does not prove that accessibility causes safety or vice versa.
Changed-context transfer. In sensor fusion, correlated measurements may motivate a lower joint capacity to avoid double influence. Statistical dependence should still be assessed directly rather than inferred from the capacity alone.
Limits. Several capacities can reproduce the same observed choices, especially on a coarse scale. Interpretation may be underdetermined.
Misconception to avoid. The learned capacity reveals the true structure of the world. It reveals one fitted decision model under data, constraints and assumptions.
Practical evidence check
- Give two different substantive stories compatible with the same capacity, then identify the extra evidence needed to choose between them.
- Return to chapter contents
CHAPTER 7 OF 10 · Test evidence and alternatives
7. Elicit and learn a capacity with constraints
Core idea. A capacity can be elicited from expert judgements, example rankings, threshold decisions or data. Learning must preserve monotonicity and avoid fitting every idiosyncratic case. Restricted families such as k-maxitive capacities can reduce complexity and improve explanation.
Method. Collect representative alternatives and preference statements, separate training and validation cases, impose boundary and monotonicity constraints, regularize complexity and retain feasible ranges when data do not identify a unique capacity. Invite experts to review decisive coalitions.
Worked example. A team provides pairwise preferences for twenty shelter profiles. Many capacities fit those examples. A constrained learner chooses a simpler two-additive or k-maxitive structure, then tests it on new profiles and presents cases where several parameter sets disagree.
Changed-context transfer. In classroom assessment, expert examples can seed a model, but the result should support teacher judgement rather than convert contested educational values into an opaque automatic grade.
Limits. Small, biased or strategically labelled samples can produce unstable capacities. Optimization success does not prove the model is substantively valid.
Misconception to avoid. A unique numerical solution means the preferences were objectively measured. Uniqueness can come from arbitrary regularization or software tie-breaking.
Practical evidence check
- Bootstrap or perturb the elicitation data and report which rankings and decisive coalitions remain stable.
- Return to chapter contents
CHAPTER 8 OF 10 · Handle limits and disagreement
8. Compare Sugeno with Choquet and other methods
Core idea. The Choquet and Sugeno integrals both use capacities, but their operations match different measurement assumptions. Choquet aggregation uses addition and supports cardinal trade-offs; Sugeno aggregation is ordinal and max-min. Outranking, rules and robust ordinal regression may better fit other structures.
Method. Start from measurement, compensation, interaction and output needs. Fit at least one credible alternative method, compare rankings and explanations, and treat disagreement as evidence about assumptions rather than a contest to find the most favorable result.
Worked example. Two shelters tie under a coarse Sugeno scale but separate under a Choquet model using defensible cardinal utility. If the decision truly supports magnitude comparisons, Choquet may add information. If the numbers are only labels, the apparent precision is misleading.
Changed-context transfer. An outranking method can include explicit vetoes and produce incomparability, which may better represent a case where evidence is too conflicting for a total ranking.
Limits. No method is assumption-free. Multiple models can be equally defensible and produce different recommendations.
Misconception to avoid. Sugeno is always safer because it uses less arithmetic. It can still encode strong, opaque or discriminatory thresholds through its capacity and scale.
Practical evidence check
- For one decision, write the measurement claim that licenses each candidate method and reject any operation the data cannot support.
- Return to chapter contents
Model-choice comparison
| Method | Uses | Main caution |
|---|---|---|
| Weighted sum | Cardinal or interval scores with compensatory trade-offs | Scale and trade-off meaning must be defensible |
| Choquet integral | Cardinal values with non-additive criterion interactions | Still uses addition and magnitude |
| Sugeno integral | Ordinal levels, thresholds and max-min aggregation | Limited discrimination and non-compensation |
| Outranking or rule system | Pairwise vetoes, concordance or explicit rules | Can produce incomparability or rule complexity |
A familiar arithmetic score is not automatically better. The defensible method is the one whose operations are meaningful for the data and decision.
CHAPTER 9 OF 10 · Handle limits and disagreement
9. Test robustness, ties and strategic ratings
Core idea. Ordinal aggregation can be sensitive to category boundaries, capacity values and rating manipulation. Coarse scales create ties; assessors may move a rating just across a decisive threshold; and a single capacity may hide legitimate stakeholder disagreement.
Method. Vary adjacent ratings and capacity values, enumerate feasible capacities, test dominance and threshold crossings, and report rank regions or ties. Use independent review and evidence anchors for criteria with high decision leverage.
Worked example. A supplier moves from acceptable to good on traceability and jumps into a decisive coalition. Sensitivity analysis shows that this one boundary changes the winner. The team audits the underlying evidence rather than celebrating a precise final score.
Changed-context transfer. For participatory budgeting, different groups may provide different capacities. Present common robust choices and contested alternatives instead of averaging away value differences.
Limits. Large combinatorial models can be difficult to communicate, and robustness analysis can itself be computationally heavy. Focus on decision-relevant uncertainty.
Misconception to avoid. Ties mean the method failed. A tie can honestly show that the available ordinal information does not support a finer distinction.
Practical evidence check
- Perturb every rating by one adjacent level where evidence is uncertain and map which alternatives remain undominated or tied.
- Return to chapter contents
CHAPTER 10 OF 10 · Transfer and govern
10. Govern qualitative aggregation as a public judgement
Core idea. A Sugeno score is a structured judgement, not a neutral fact. Governance should make the scale, capacity, evidence, decisive coalitions, alternatives, uncertainty, rights constraints and review route visible to those affected.
Method. Publish the model card and worked examples, separate eligibility rules from aggregation, preserve assessor evidence, enable correction and appeal, monitor group effects and schedule review after policy or population change. Never let a single score erase material criterion details.
Worked example. A scholarship committee uses an ordinal capacity model for shortlisting. Applicants receive the criteria, anchors and review route. The committee audits whether one coalition systematically excludes students with interrupted schooling and changes the policy when the effect lacks justification.
Changed-context transfer. Transfer to personal decision-making: a family can use the method to expose non-negotiables in choosing a home while keeping the underlying ratings and disagreements visible.
Limits. Full transparency can expose private data or invite gaming. Publish decision logic and safeguards while protecting sensitive individual evidence.
Misconception to avoid. Once a model is approved, applying it consistently is automatically just. Consistency can reproduce an unjust scale or capacity; outcome and process review remain necessary.
Practical evidence check
- Trace one recommendation from source evidence through rating, threshold, coalition and result, then show how an affected person can challenge every step.
- Return to chapter contents
Build an ordinal scale that can survive challenge
Define each level with observable conditions and boundary examples. Avoid adjectives that invite assessors to substitute personal taste.
Check whether all criteria truly share the same ordered scale. If not, document a defensible mapping or use a method that permits different structures.
Test inter-rater agreement and disagreement reasons. Revise the rubric where experts use the same label for different evidence.
Preserve the original observations beside the category. A rating is a judgement about evidence, not the evidence itself.
Elicit a capacity through coalition questions
Begin with policy questions: which single criteria are sufficient, which pairs become decisive, which evidence is redundant and which weaknesses cannot be compensated?
Translate answers into boundary and monotonicity constraints before asking for exact values. Feasible ranges may be more honest than forced point estimates.
Use worked profiles to check whether the implied capacity produces intended outcomes. People often discover contradictions only when they see concrete cases.
Document dissent and consider multiple stakeholder capacities. A negotiated average can conceal incompatible values.
Audit the decisive threshold and coalition
For every result, retain the sorted evaluation vector, threshold sets, capacity values, min results and maximum-attaining coalition.
Explain the score in plain language: which level was supported, by which criteria and why a higher level failed.
Where several coalitions attain the same result, report all materially different explanations. Selecting one convenient story is misleading.
Compare the winning explanation across demographic or operational groups to detect whether the model relies on systematically unequal evidence.
Run a robustness and alternatives review
Vary uncertain ratings, capacity constraints and scale boundaries. Report stable choices, ties and reversals.
Compare a weighted, Choquet, Sugeno, outranking or explicit-rule model only when each is meaningful for the data. Method disagreement is a prompt to revisit assumptions.
Test strategic cases where one criterion crosses a decisive threshold. Strengthen evidence review where leverage is high.
Set a decision rule for unresolved ties: collect targeted evidence, choose a robust alternative, use deliberation or declare incomparability rather than inventing decimals.
Questions readers ask
What kind of data suits a Sugeno integral?
- Ordered evaluations where rank and thresholds are meaningful but numerical distances and addition are not.
What is a capacity?
- A monotone importance function defined on subsets of criteria, able to represent coalition effects beyond additive weights.
Why is the method called non-compensatory?
- Max-min aggregation can preserve bottlenecks, so excellence on one criterion need not offset a critical weakness elsewhere.
Can Sugeno scores produce many ties?
- Yes. Coarse ordinal scales and threshold logic can limit discrimination; ties may honestly reflect insufficient information.
Is a Sugeno integral the same as a Choquet integral?
- No. Both use capacities, but Choquet aggregation is cardinal and additive in form, while Sugeno aggregation is ordinal and lattice-based.
Can a learned capacity be treated as objective truth?
- No. It is a fitted decision model shaped by examples, constraints, regularization and stakeholder values.
Practise, verify and continue
Create a four-level ordinal scale and four criteria for selecting an emergency shelter. Define at least six coalition capacities, verify monotonicity, compute the Sugeno integral for three alternatives, identify every decisive threshold and test whether one-level rating changes reverse the choice.
Move the model to supplier selection, school projects or personal planning. Decide whether the new domain truly supports ordinal non-compensation; if it supports meaningful trade-off magnitudes instead, explain why another method is preferable.
Related eduKate reading
Choquet integrals, capacities and non-additive expectations
Possibility theory, necessity measures and maxitive uncertainty
Stochastic dominance, utility classes and partial preferences
Imprecise probabilities, credal sets and robust choice
Dempster–Shafer evidence, belief functions and conflict
Authoritative sources and scope
Michio Sugeno, Fuzzy Measure and Fuzzy Integral
Jean-Luc Marichal, An Axiomatic Approach of the Discrete Sugeno Integral as a Tool to Aggregate Interacting Criteria in a Qualitative Framework
Dieter Denneberg and Michel Grabisch, Measure and Integral with Purely Ordinal Scales
Brabant and colleagues, k-maxitive Sugeno integrals as aggregation models for ordinal preferences
Halaš, Mesiar and Pócs, Sugeno integrals on bounded distributive lattices
This article is an educational decision-analysis framework, not a recommendation for automated eligibility, medical, legal or financial decisions. Examples are original. A Sugeno model is defensible only when the scale is genuinely ordinal, capacity constraints are meaningful, evidence is reliable and non-compensation fits the decision. Preserve criterion-level evidence, report ties and sensitivity, compare credible alternative methods, assess accessibility and disparate effects, and keep a human review and correction route. A capacity describes the model’s treatment of coalitions; it does not prove causal interaction or moral importance. The gold standard uses lattice aggregation to make qualitative thresholds explicit without pretending that ranks are measurable distances.