Multi-Criteria Decision Analysis (MCDA) is a family of methods for systematically organizing evaluation criteria, the characteristics of alternatives, and decision-makers’ preferences to support decisions involving multiple, often conflicting objectives. Developed primarily within operations research (OR) and decision theory, MCDA has been applied to business strategy, R&D project selection, capital investment, public policy, environmental assessment, healthcare, product development, and many other fields.
Real-world decisions rarely depend on a single criterion. When selecting a research and development project, for example, an organization may need to consider technical feasibility, market potential, novelty, development costs, and alignment with business strategy. However, alternatives seldom perform equally well across all criteria. A project with strong market potential may involve substantial technical risks, while a technically feasible project may offer limited novelty.
A common approach is to assign numerical scores to each criterion and calculate an overall weighted score. However, the alternative with the highest score is not necessarily the most appropriate choice. The result may depend heavily on how criteria are defined, how scores and weights are assigned, whether criteria interact, and whether certain requirements must be satisfied regardless of other advantages.
MCDA provides mathematical and systematic approaches to examining these problems. Over the years, numerous methods have been developed, ranging from simple weighted scoring and pairwise comparisons to nonlinear utility models, outranking methods, and approaches that explicitly account for uncertainty.
The purpose of MCDA is not to identify a single objectively correct decision through mathematics. It is to make decision assumptions explicit, compare alternatives and value judgments, and support more informed and critically examined decisions.
Types of Decision Problems Addressed by MCDA
MCDA is closely related to Multiple-Criteria Decision-Making (MCDM). The two terms are frequently used interchangeably, although MCDA often emphasizes the broader analytical process, including problem structuring, preference elicitation, model construction, and interpretation of results.
Decision problems can be classified into several categories according to their objectives.
Choice involves selecting the most appropriate alternative or a subset of acceptable alternatives from a set of candidates. Examples include choosing equipment, selecting R&D projects, or deciding among business investments.
Ranking involves arranging alternatives according to their relative priorities. Examples include establishing R&D project priorities or determining the implementation sequence of technology development initiatives.
Sorting involves assigning alternatives to predefined categories. For example, R&D projects may be classified as “High-Priority Investment,” “Continue Evaluation,” “On Hold,” or “Discontinue.”
Portfolio Selection involves selecting a combination of alternatives under constraints such as budgets, personnel, or other resources. This is particularly important in R&D and capital investment, where ranking individual projects does not necessarily produce the best overall portfolio.
Different MCDA methods are appropriate for different decision objectives. It is not always necessary—or even desirable—to produce a complete ranking of all alternatives.
The Basic MCDA Process
An MCDA process generally begins by defining the decision problem, identifying alternatives, and establishing evaluation criteria. Evaluation data and preference information are then collected, and an appropriate mathematical model is used to compare the alternatives.
A typical process consists of the following steps.
- Define the decision objective: Clarify what must be decided, who the decision serves, and which constraints must be considered.
- Identify alternatives: List the available options and, where necessary, generate additional alternatives.
- Establish evaluation criteria: Identify the relevant dimensions of evaluation and define how each criterion will be measured.
- Collect evaluation data: Gather measurements, forecasts, expert judgments, and other information about the alternatives.
- Specify the preference model: Define weights, value functions, mandatory constraints, outranking relationships, or other preference information.
- Analyze and compare alternatives: Apply the selected MCDA method to calculate rankings, classifications, or preference relationships.
- Perform sensitivity and robustness analysis: Examine how the results change when assumptions, parameters, or evaluation data are modified.
- Make and review the decision: Interpret the results and revise the problem definition or evaluation model when necessary.
This process is not necessarily linear. If an analysis reveals a missing criterion or a previously overlooked alternative, the decision-maker may return to an earlier stage and reconsider the model.
Pareto Dominance and Trade-Offs
One of the fundamental concepts in MCDA is Pareto Dominance.
Alternative A is said to Pareto-dominate alternative B if A performs at least as well as B under every evaluation criterion and strictly better under at least one criterion, assuming all criteria have been expressed in a direction where higher performance is preferable.
Under ordinary monotonic preferences, there is no reason to select B over A based solely on those criteria.
However, real-world decisions frequently involve trade-offs, so neither alternative dominates the other.
For example, R&D Project A may have greater market potential, while Project B has higher technical feasibility. The preferred project depends on how the decision-maker values these characteristics.
The set of alternatives that are not dominated by any other alternative is called the Pareto-efficient set, non-dominated set, or, in an appropriate objective space, the Pareto frontier.
MCDA is particularly relevant when alternatives are non-dominated and a decision requires additional preference information to resolve trade-offs.
Major MCDA Methods
MCDA includes numerous mathematical methods that differ in how they represent preferences and combine evaluation criteria.
| Method | Basic Principle | Main Characteristics |
|---|---|---|
| Weighted Sum Model (WSM) | Sum of weighted criterion scores | Simple and intuitive |
| MAVT / MAUT | Modeling multi-attribute values and utilities | Supports value functions and uncertainty |
| AHP | Deriving priorities through pairwise comparisons | Structures subjective judgments |
| ANP | Representing dependencies through a network | Accounts for feedback and interdependencies |
| TOPSIS | Comparing distances from ideal and anti-ideal solutions | Relative evaluation using multiple performance indicators |
| ELECTRE | Evaluating concordance and discordance between alternatives | Accommodates veto conditions and incomparability |
| PROMETHEE | Aggregating pairwise preference relationships | Uses preference functions and thresholds |
| Choquet Integral | Non-additive aggregation of criteria | Represents synergy and redundancy |
| SMAA | Analyzing uncertainty in weights and evaluations | Examines rank acceptability |
These methods should not be interpreted as successive levels of sophistication along a single development path. They emerged from different assumptions about preferences, evaluation, and decision-making.
Value Functions and Utility Functions
A central challenge in MCDA is converting different evaluation criteria into comparable measures of value.
When evaluating a new business opportunity, for example, market size might be expressed in monetary terms, development time in months, manufacturing costs in dollars, and competitive advantage through qualitative ratings.
These quantities cannot be meaningfully added without appropriate transformations.
A value function converts the performance of an alternative under a particular criterion into a measure of its value to the decision-maker.
Multi-Attribute Value Theory (MAVT) provides a framework for combining such values.
A common additive value model is expressed as:
V(a) = Σ wᵢ × vᵢ(xᵢ(a))
where:
- V(a) is the overall value of alternative a.
- xᵢ(a) is the performance of alternative a under criterion i.
- vᵢ is the value function for criterion i.
- wᵢ is the weight associated with criterion i.
Value functions do not need to be linear.
For example, an improvement in product performance may be extremely valuable until a minimum acceptable level is reached, while further improvements beyond that level produce diminishing benefits.
A nonlinear value function can represent such preferences.
Multi-Attribute Utility Theory (MAUT) extends the analysis to preferences involving uncertain outcomes and attitudes toward risk.
Although MAVT and MAUT are closely related, a useful distinction is that MAVT generally addresses the value of outcomes under certainty, whereas MAUT can represent preferences over uncertain consequences through expected utility.
However, specifying separate value functions for each criterion does not automatically justify an additive aggregation model. Additional assumptions, including appropriate forms of preferential independence, are required.
What Do Criterion Weights Actually Mean?
Multi-criteria evaluation is often explained as assigning larger weights to more important criteria.
However, the meaning of a weight depends on the mathematical model being used.
In an additive value model, weights are not simply abstract measures of importance. They are scaling constants that depend on the value ranges associated with the criteria and express trade-offs among improvements.
For example, a weight of 0.4 for one criterion and 0.2 for another does not necessarily mean that the first criterion is universally twice as important.
The interpretation depends on the ranges of performance considered and the associated value functions.
Consider market size as an evaluation criterion for R&D projects.
The value of improving market potential from $100 million to $1 billion may be very different from improving it from $100 million to $110 million.
The importance of the criterion must therefore be understood in relation to the range of possible improvements.
One method for addressing this issue is Swing Weighting.
Swing Weighting asks decision-makers to compare the value of moving each criterion from its least desirable level to its most desirable level.
The resulting weights reflect the relative value of these improvements rather than only general statements about importance.
AHP, by contrast, derives relative priorities through pairwise comparisons. Weights produced by AHP and scaling constants used in an additive value model are not necessarily identical in meaning.
In MCDA, understanding what weights represent is as important as calculating them.
Problems with Overall Weighted Scores
The simplest form of MCDA uses a weighted sum of evaluation scores.
Consider three R&D projects evaluated using three criteria.
| Project | Market Potential | Technical Feasibility | Strategic Importance |
|---|---|---|---|
| A | 8 | 6 | 7 |
| B | 6 | 9 | 8 |
| C | 9 | 5 | 6 |
Assume that all scores range from 0 to 10 and represent comparable value scales.
Suppose the criterion weights are:
- Market Potential: 40%
- Technical Feasibility: 35%
- Strategic Importance: 25%
The overall scores become:
| Project | Overall Score |
|---|---|
| A | 7.05 |
| B | 7.55 |
| C | 6.85 |
Project B receives the highest score.
Now consider a different set of weights:
- Market Potential: 60%
- Technical Feasibility: 15%
- Strategic Importance: 25%
The results change to:
| Project | Modified Overall Score |
|---|---|
| A | 7.45 |
| B | 6.95 |
| C | 7.65 |
Project C now ranks first.
Both calculations are mathematically correct, but they produce different conclusions because the evaluation weights have changed.
Suppose an additional mandatory requirement is introduced: projects must receive a technical feasibility score greater than 5.
Project C would then be excluded, despite receiving the highest weighted score under the modified weights.
However, such a requirement might be inappropriate for fundamental research, where low technical feasibility can itself justify further investigation.
The mathematical correctness of an evaluation and the appropriateness of its underlying assumptions are different questions.
Compensatory and Non-Compensatory Evaluation
An important distinction in MCDA concerns whether poor performance under one criterion can be compensated for by strong performance under another.
In an additive weighted model, low scores under one criterion may be offset by high scores elsewhere. This is known as compensatory evaluation.
However, some requirements cannot reasonably be compensated for.
For example, manufacturing equipment that fails to satisfy mandatory safety regulations cannot be accepted simply because it offers high productivity and low costs.
In such situations, mandatory constraints or non-compensatory comparison methods may be more appropriate.
Outranking methods, such as ELECTRE, evaluate whether one alternative can be considered at least as good as another based on the pattern of supporting and opposing criteria.
Rather than reducing all evaluations to a single overall score, ELECTRE examines preference relationships between alternatives.
For example, Alternative A may outperform B on most criteria. However, if A performs unacceptably poorly under a critical criterion, the method may prevent A from being considered superior to B.
Depending on the particular ELECTRE method, this can involve concepts such as:
- Concordance: The extent to which criteria support the conclusion that one alternative outranks another.
- Discordance: The extent to which criteria oppose that conclusion.
- Veto Thresholds: Conditions under which a sufficiently serious disadvantage prevents an outranking relationship.
An important feature of many outranking methods is that alternatives need not always be placed into a complete ranking.
Two alternatives may remain incomparable when the available preference information does not justify a clear preference.
PROMETHEE is another family of outranking methods.
It uses criterion-specific preference functions and aggregates pairwise preferences into measures such as positive, negative, and net preference flows.
Depending on the variant, PROMETHEE can produce partial or complete rankings.
These methods differ from simple weighted sums in their underlying preference models, although their results still depend on how criteria, weights, and thresholds are specified.
Nonlinearity and Interactions Among Criteria
Evaluation criteria are often treated as independent contributors to overall value.
In reality, however, the value associated with one criterion may depend on another.
Consider evaluating the competitive strength of a new business based on technological capability and manufacturing capability.
A company may possess advanced technology but lack the manufacturing capacity required for commercialization.
Technological superiority alone may therefore generate limited business value.
Conversely, combining advanced technology with strong manufacturing capabilities may create a competitive advantage greater than the sum of their individual contributions.
This represents an interaction between criteria.
In regression analysis, such relationships can be represented using interaction terms.
In MCDA, non-additive evaluation models can represent situations in which the combined value of several criteria differs from the sum of their separate contributions.
One representative approach is the Choquet Integral.
The Choquet integral uses a non-additive set function that can assign importance not only to individual criteria but also to combinations of criteria.
This makes it possible to represent synergy or redundancy.
For example, a model may account for the additional value generated when technological capability and manufacturing capability are both strong.
It may also represent partial redundancy between evaluations of market potential and profitability.
However, this flexibility comes at a cost.
A general Choquet integral may require a number of parameters that grows exponentially with the number of evaluation criteria.
A mathematically sophisticated model does not necessarily improve decision quality if users cannot understand or appropriately specify its parameters.
The practical challenge is therefore to determine which interactions are important enough to justify additional model complexity.
Uncertainty and Robustness Analysis
Evaluation criteria, weights, and scores are not always known precisely.
This is particularly important in R&D and new business evaluation, where market size, development costs, technical feasibility, and the likelihood of success may all involve substantial uncertainty.
Stakeholders may also disagree about the relative importance of the evaluation criteria.
Under such circumstances, calculating a single ranking from one set of weights and scores may conceal important information.
MCDA therefore includes approaches such as sensitivity analysis, scenario analysis, and robustness analysis.
Sensitivity Analysis examines how rankings or evaluations change when weights, scores, or other parameters are modified.
Scenario Analysis evaluates alternatives under different possible future conditions, such as changes in market demand or technological development.
Robustness Analysis investigates whether conclusions remain valid across a range of plausible assumptions or preference models.
More specialized methods have also been developed to address uncertain or incomplete preference information.
Robust Ordinal Regression (ROR)
Robust Ordinal Regression considers multiple value models that are compatible with the preference information provided by the decision-maker.
Rather than selecting a single set of parameters immediately, it examines the conclusions that remain valid across the set of compatible models.
For example, it can distinguish between a preference for Alternative A over B that holds under every compatible model and one that holds only under some models.
This helps identify which conclusions are robust and which depend on unresolved preference assumptions.
Stochastic Multicriteria Acceptability Analysis (SMAA)
SMAA examines decision problems in which weights or evaluation values are uncertain.
By considering distributions or feasible ranges of these quantities, SMAA can estimate how frequently different alternatives achieve particular rankings under the assumed model of uncertainty.
For example, it can investigate questions such as:
- Under what proportion of admissible weight combinations does Project A rank first?
- Which criteria must receive relatively high weights for Project B to become the preferred alternative?
- How sensitive are rankings to uncertain evaluation data?
However, the rank acceptability indices produced by SMAA depend on the distributions and assumptions used in the analysis.
They should not be confused with real-world probabilities that a project will succeed.
These methods are valuable because they shift attention away from obtaining a single optimal ranking and toward understanding the uncertainty surrounding the decision itself.
Decision-Making Involving Multiple Stakeholders
Corporate and public-sector decisions often involve stakeholders with different priorities.
For example, an R&D department may emphasize technological novelty, a sales department may emphasize market potential, and corporate management may emphasize return on investment.
Simply averaging their weights is not necessarily the most appropriate approach.
Differences in weighting may reflect legitimate differences in organizational responsibilities, strategic perspectives, and underlying objectives.
MCDA can support both the aggregation of multiple stakeholders’ preferences and the separate analysis of their differing value judgments.
For example, if the technical department ranks Project A first while the sales department prefers Project B, the analysis can identify which evaluation criteria are responsible for the disagreement.
This can lead to a discussion of possible compromises, revised strategies, or combinations of alternatives.
The objective need not be to force all stakeholders to adopt identical weights.
Instead, MCDA can help reveal the structure of disagreements and identify conditions under which acceptable decisions or agreements become possible.
In this sense, mathematical decision analysis can also serve as a structured framework for organizational discussion.
Relationship Between MCDA and Bayesian Networks
MCDA primarily addresses the evaluation and comparison of alternatives based on preferences and values.
It does not necessarily focus on calculating the probabilities of uncertain events.
A Bayesian Network, by contrast, represents probabilistic dependencies among variables and updates probability distributions when evidence becomes available.
For example, a Bayesian Network could be used to estimate the probability that an R&D project will achieve technical success, that market demand will increase, or that commercialization will become feasible.
However, the project with the highest probability of success is not necessarily the most valuable project.
A project with a lower success probability may offer much greater returns or strategic benefits if successful.
This suggests the possibility of combining probabilistic models with decision models.
One established framework for doing so is the Influence Diagram.
An Influence Diagram incorporates chance nodes, decision nodes, and utility nodes to represent uncertainty, available actions, and the value of possible outcomes.
It can then be used to compare decisions based on expected utility under uncertainty.
MCDA criteria and value functions can also be incorporated into appropriate probabilistic decision models, allowing strategies to be evaluated under uncertain technological and market conditions.
However, it is essential to distinguish between three different concepts:
- The relative importance assigned to evaluation criteria.
- The probabilities of uncertain events.
- The values or utilities of possible outcomes.
These quantities have different meanings and should not be combined without a mathematically appropriate model.
Limitations and Practical Considerations
Although MCDA provides systematic approaches to complex decisions, mathematical modeling does not guarantee that the resulting judgment is correct.
Several issues require particular attention.
Evaluation criteria may be incomplete, overlapping, or poorly defined. Measurement scales may be inappropriate. Weights may lack a clear basis. Important interactions or constraints may be omitted.
There is also the risk of introducing unnecessary model complexity.
Different MCDA methods may produce different rankings even when applied to the same evaluation data.
For example, a weighted sum model, TOPSIS, and ELECTRE may reach different conclusions because they embody different assumptions about preferences, aggregation, and compensation.
Such differences do not necessarily indicate computational errors.
Instead, they reveal that decision-making depends partly on how preferences are represented mathematically.
The important question is whether the selected method accurately reflects the structure of the actual decision problem.
Complex models may require numerous parameters. If these parameters are assigned without adequate evidence or understanding, the analysis may create an impression of precision without improving the quality of the decision.
In MCDA, selecting a model appropriate to the problem and examining its assumptions is more important than choosing the most mathematically elaborate method.
MCDA in the Age of AI
The development of Large Language Models (LLMs) creates new opportunities to make MCDA more accessible and useful.
Traditionally, MCDA required substantial preparatory work, including problem structuring, criterion definition, alternative identification, data collection, and selection of an appropriate mathematical model.
LLMs can assist with many of these activities.
Users can describe a decision problem in natural language and ask an LLM to suggest evaluation criteria, identify alternatives, and organize relevant information.
LLMs can also retrieve evidence from technical literature, market research, corporate documents, and other sources.
Furthermore, they can support the critical examination of decision models.
For example, an LLM might raise questions such as:
- Are market potential and profitability being evaluated using overlapping information?
- Do any alternatives fail to satisfy mandatory requirements?
- Are interactions among evaluation criteria important?
- What uncertainties have the greatest influence on the decision?
- Would a different MCDA method be more appropriate?
- Does the preferred alternative remain favorable under different assumptions?
Such questions can help decision-makers identify weaknesses in their models.
However, mathematical calculations should generally be performed by dedicated analytical engines rather than relying solely on LLM-generated numerical results.
This provides greater reliability, reproducibility, and transparency.
It is also important not to treat LLM-generated scores or weights as automatically representing the user’s values.
The importance assigned to evaluation criteria depends on the decision-maker’s objectives, preferences, and strategic priorities.
LLMs can provide information and identify issues, but they do not need to determine the final value judgments.
Another promising direction is Preference Learning.
By examining previous decisions or pairwise comparisons, computational models may learn aspects of a decision-maker’s preferences and propose corresponding evaluation models.
This connects MCDA with research in preference learning, interactive decision support, and methods such as Robust Ordinal Regression.
Combining MCDA with LLMs may therefore do more than automate traditional analysis. It could create new environments in which users can examine and refine the assumptions, preferences, and reasoning processes underlying their decisions.
MCDA in ThinkNavi
ThinkNavi regards MCDA as an important family of mathematical thinking methods for expressing human value judgments and examining the validity of decisions.
ThinkNavi Free is planned to provide analytical tools that allow users to enter evaluation criteria, alternatives, scores, and weights manually and examine the resulting calculations.
By combining basic weighted evaluation through Decision Matrices, AHP-based weighting, mandatory constraints, nonlinear value functions, and sensitivity analysis, users will be able to examine limitations of conventional overall-score approaches.
The integration of different methods will also allow results from one analysis to be transferred into another.
For example, users could begin with SWOT analysis to structure the business environment, use TOWS to develop strategic alternatives, compare those alternatives through a Decision Matrix, apply AHP to examine relative importance, and perform sensitivity analysis to assess the stability of their conclusions.
More advanced approaches—including ANP, Choquet integrals, outranking methods, and SMAA—may also be considered.
However, implementing every available method is not the objective. Development should proceed by examining the practical value each method provides for the types of problems users need to solve.
In an LLM-assisted version, AI could support criterion identification, alternative generation, information retrieval, evaluation model construction, and critical examination of assumptions.
The broader ConceptMiner product family could also incorporate MCDA tools to evaluate and prioritize R&D themes, business hypotheses, and other opportunities generated through concept exploration.
Nevertheless, the essential value of MCDA extends well beyond ranking alternatives according to their overall scores.
What criteria are being used to make the decision? Are those criteria appropriate? Are relationships among them represented correctly? Would the conclusion remain valid if the assumptions changed?
Providing systematic ways to examine these questions is one of the most important contributions of mathematical decision analysis.
ThinkNavi aims to make established decision-making methods more accessible through computers and AI, while developing a thinking-support environment that helps people understand and improve the decision process itself—not merely calculate its results.
Related Thinking Methods:
- Decision Matrix | Comparing Alternatives Across Multiple Criteria and Evaluating Decision Validity
- Analytic Hierarchy Process (AHP) | Quantifying Evaluation Criteria Through Pairwise Comparisons
- Analytic Network Process (ANP) | Decision-Making with Interdependencies and Feedback
- Multi-Attribute Utility Theory (MAUT),
- Nonlinear Value Functions,
- Choquet Integral | Modeling Synergy and Redundancy Among Evaluation Criteria
- ELECTRE,
- PROMETHEE,
- TOPSIS,
- Sensitivity Analysis | Examining How Changes in Assumptions Affect Model Results
- Scenario Analysis | Exploring Multiple Futures and Developing Strategies Under Uncertainty
- Bayesian Networks | Structuring Uncertainty and Reasoning with Probabilistic Models
- Influence Diagrams
References and Further Reading
- Ralph L. Keeney and Howard Raiffa, Decisions with Multiple Objectives: Preferences and Value Tradeoffs, Wiley, 1976.
- Valerie Belton and Theodor J. Stewart, Multiple Criteria Decision Analysis: An Integrated Approach, Kluwer Academic Publishers, 2002.
- José Figueira, Salvatore Greco, and Matthias Ehrgott (eds.), Multiple Criteria Decision Analysis: State of the Art Surveys, Springer, 2005.
- Salvatore Greco, Roman Słowiński, and Jyrki Wallenius, “Fifty Years of Multiple Criteria Decision Analysis: From Classical Methods to Robust Ordinal Regression,” European Journal of Operational Research, Vol. 323, No. 2, pp. 351–377, 2025.
- Risto Lahdelma and Pekka Salminen, “SMAA-2: Stochastic Multicriteria Acceptability Analysis for Group Decision Making,” Operations Research, 49(3), pp. 444–454, 2001.
- UK Government — Multi-Criteria Analysis: A Manual
- UK Government — Use of MCDA in Options Appraisal
- Fifty Years of Multiple Criteria Decision Analysis — Research Review
- How to Support the Application of MCDA? A Comprehensive Taxonomy