Multi-Criteria Decision-Making (MCDM)

Updated 17 April 2026

Multi-Criteria Decision-Making (MCDM) is a framework that breaks down complex decisions into alternatives, criteria, and weighted scores.
It employs aggregation, distance, outranking, and fuzzy methods to resolve conflicts between quantitative and qualitative criteria.
Researchers use MCDM for applications such as site selection and resource allocation by combining subjective expert judgments with data-driven approaches.

Multi-Criteria Decision-Making (MCDM) is a universally adopted paradigm in operations research and decision sciences for ranking, selecting, or classifying alternatives when evaluated against multiple, potentially conflicting criteria. MCDM frameworks decompose decision problems into formal structures, aggregate quantitative and qualitative judgments, and provide reproducible mechanisms for transparent, context-sensitive decision support across engineering, management, environmental, and artificial intelligence domains.

1. Formal Structure and Problem Decomposition

MCDM problems are defined over a finite set of alternatives $\{A_1, \dots, A_m\}$ and a vector of $n$ criteria $\{C_1, \dots, C_n\}$ , forming a decision or alternatives-criteria matrix $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ , where $x_{ij}$ quantifies the performance of $A_i$ under $C_j$ (Wang et al., 2024). Problem decomposition follows:

Alternatives and Criteria: Alternatives can be products, sites, process models, or even strategies (e.g., bank ranking (Hien et al., 9 Sep 2025), site selection (Ahmed et al., 5 Apr 2025), software frameworks (Basciani et al., 2023)). Criteria typically span heterogeneously measured objectives, e.g., cost, efficiency, risk, and subjective factors (stakeholder satisfaction).
Hierarchy and Levels: Decision structures may incorporate hierarchies, subcriteria, and expert groups (e.g., AHP (Bemthuis, 15 May 2025), hierarchical quality models (Basciani et al., 2023)).
Weights and Importance: Each criterion $C_j$ is assigned a weight $w_j \geq 0$ , $\sum_{j=1}^n w_j = 1$ , elicited subjectively (expert pairwise table [AHP], rank-based [SMARTER/SMARTS]), or objectively via data-driven schemes (e.g., entropy, CRITIC, Gini) (Wang et al., 2024, Bhatia et al., 2023).

Collectively, an MCDM instance may be abstracted as $n$ 0 (Wang et al., 8 Sep 2025, Najafi et al., 12 Feb 2025).

2. Algorithmic Families and Solution Paradigms

MCDM methods fall into several mathematical families:

2.1 Aggregation-Type (Weighted Sum/Product, Hierarchical)

Simple Additive Weighting (SAW):

$n$ 1

where $n$ 2 are normalized values (Wang et al., 8 Sep 2025). Full compensation is allowed.

Multiplicative Exponent Weighting (MEW):

$n$ 3

Useful in risk-averse settings (Wang et al., 8 Sep 2025).

Analytic Hierarchy Process (AHP):
- Pairwise comparisons generate an $n$ 4 reciprocal matrix $n$ 5, eigenvector solution gives weights, local alternative scores aggregated hierarchically (Bemthuis, 15 May 2025, Wang et al., 8 Sep 2025).
SMARTER/SMARTS:
- Ranks (ROC, RR, RS) or swing weights are mapped to $n$ 6 (Basciani et al., 2023).

2.2 Reference-Type (Distance to Internal Benchmarks)

TOPSIS:
- Ranks alternatives by Euclidean distance to the Positive Ideal Solution (PIS) and Negative Ideal Solution (NIS):
$n$ 7

$n$ 8 and $n$ 9 are distances to PIS/NIS, respectively (Wang et al., 22 Aug 2025, Nabavi et al., 2024, Bhatia et al., 2023).
VIKOR:
- Balances group utility ( $\{C_1, \dots, C_n\}$ 0) and individual regret ( $\{C_1, \dots, C_n\}$ 1) using compromise programming (Wang et al., 22 Aug 2025).
MABAC, CODAS, PROBID, MARCOS:
- Employ various reference constructs (geometric means, tiered ideals, relative measures) to score alternatives (Wang et al., 22 Aug 2025).

2.3 Outranking and Noncompensatory

ELECTRE, PROMETHEE:
- Define pairwise concordance/discordance or preference flows, enabling partial rankings with veto logic (Wang et al., 2024, Najafi et al., 12 Feb 2025).

2.4 Specialized and Robustness-Enhancing Methods

Thermodynamic Approach:
- Ranks by “exergy,” penalizing ratings with high dispersion among experts, distinguishing between quantity (energy) and quality (exergy) (Verma et al., 2015).
Multivariate Quantiles and Set Optimization:
- Provides [0,1]-valued rankings via infimum over judge-weighted projections, with “good” and “bad” quantile sets as multidimensional generalizations of classic quantiles (Kostner, 2018).
ML-Integrated MCDM:
- Utilizes (e.g., Random Forest) feature importances as weights in the decision matrix, enabling objective, high-dimensional optimization without subjective weighting (Ahmed et al., 5 Apr 2025).

3. Weight Determination and Sensitivity

Weight determination is pivotal:

Subjective: Pairwise (AHP), rank-based (SMARTER/SMARTS), direct assignment, or group-elicited, with consistency checks (CR) in eigenvector methods (Bemthuis, 15 May 2025, Basciani et al., 2023).
Objective: Entropy, CRITIC, StDev, Gini, and data-driven ML importances (Bhatia et al., 2023, Nabavi et al., 2024, Ahmed et al., 5 Apr 2025).
Hybrid: Probabilistic Bayesian aggregation (Mixture of Dirichlet, logistic-Normal) supports uncertainty and subgroup discovery (Mohammadi, 2022).

Sensitivity analysis is critical for robustness:

Entropy weights are demonstrably sensitive to linear and reciprocal objective transformations, while CRITIC and StDev exhibit greater robustness (Nabavi et al., 2024).
Methods relying on Euclidean/taxicab distances (CODAS, MABAC, SAW, TOPSIS) combined with StDev or CRITIC weighting tend to yield stable rankings under typical domain perturbations.

4. Practical Workflow, Software, and Implementations

The canonical workflow (Wang et al., 2024, Najafi et al., 12 Feb 2025):

Define alternatives and criteria
Populate $\{C_1, \dots, C_n\}$ 2: Quantitative/qualitative measures per alternative and criterion
Normalize $\{C_1, \dots, C_n\}$ 3: Vector, min–max, or sum normalization; treat benefit/cost criteria symmetrically
Determine $\{C_1, \dots, C_n\}$ 4: Elicit or calculate weights as appropriate
Aggregate or compute distances: As per method (see above)
Rank alternatives: By score, closeness, or preference net-flow

Tooling:

RMCDA (R package) implements AHP, TOPSIS, PROMETHEE, VIKOR, SMCDM, SBWM, with extensive visualization and auditability (Najafi et al., 12 Feb 2025).
PyMCDM (Python), EMCDM (ExcelVBA) provide parallel suites for computational experiments (Wang et al., 2024).

Example Table: Common MCDM Methods and Core Features

Method	Score/Mechanism	Core Equation
SAW	Additive	$\{C_1, \dots, C_n\}$ 5
TOPSIS	Dist. to PIS/NIS	$\{C_1, \dots, C_n\}$ 6
AHP	Hierarchical	$\{C_1, \dots, C_n\}$ 7
VIKOR	Compromise	$\{C_1, \dots, C_n\}$ 8 via $\{C_1, \dots, C_n\}$ 9 (utility), $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 0 (regret), $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 1
PROMETHEE	Net flows	$X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 2
CODAS	Euclidean, $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 3	$X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 4

Extensive worked examples in renewable energy (Bhatia et al., 2023), process mining (Bemthuis, 15 May 2025), COVID-19 ML model selection (Chowdhury et al., 2021), and bank supervision (Hien et al., 9 Sep 2025) illustrate method adaptation across domains.

5. High-Dimensional, Uncertainty, and Modern Developments

High-dimensional, mixed-data problems expose scalability and error-proneness in traditional MCDM. Key advances:

LLM-Based MCDM: Reformulates MCDM as a textual reasoning task for LLMs, employing chain-of-thought, few-shot prompting, and LoRA-based fine-tuning to reach $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 5 classification accuracy (vs. $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 6 for vanilla zero-shot LLMs) and human-expert-level solutions (Wang et al., 17 Feb 2025).
Fuzzy/Uncertain MCDM: Frameworks handle linguistic, interval, and probabilistic inputs. Membership functions (triangular/trapezoidal), aggregation by centroid (COA), and specialized operators (DEMATEL, ISM, FCM) model vagueness/conflict in group and multi-stage settings (Fujita et al., 16 Mar 2026).
Probabilistic Rankings: Bayesian mixtures (e.g., Unified Bayesian Framework), with group consensus/subgroup discovery and credal ordering of both criteria and alternatives, furnish full posterior distributions, not just point estimates (Mohammadi, 2022).
ML-MCDM Hybrids: Machine learning supplies both criteria weights (feature importances) and unbiased aggregation, with demonstrated predictive improvements in real-world site selection and other spatially complex domains (Ahmed et al., 5 Apr 2025).

6. Limitations, Robustness, and Method Selection

Critical limitations and risks:

Rank Reversal: All reference-type MCDM methods (especially TOPSIS, SAW) are susceptible upon addition/removal of alternatives (Wang et al., 22 Aug 2025, Wang et al., 2024).
Subjectivity and Human Error: Weight selection, criteria definition, normalization scaling, and template sensitivity (for LLMs) may unintentionally introduce bias, especially as dimensionality increases (Wang et al., 17 Feb 2025, Nabavi et al., 2024).
Interpretability: Outranking and reference-based distances may obscure the causal path from input to ranking; LLM and black-box ML methods further complicate explanation.
Computational Complexity: Outranking and pairwise methods (PROMETHEE, CURLI) scale poorly in $X = [x_{ij}] \in \mathbb{R}^{m \times n}$ 7; LP-based approaches (Virtual Gap Analysis (Liu et al., 2024)) scale efficiently but require LP solvers.

Best-practices suggest:

Perform sensitivity analysis (vary normalization, weighting, rank across methods)
Use multiple MCDM methods and ensemble/aggregate rankings for robustness (Najafi et al., 12 Feb 2025)
Match method to data (ordinal/ranked: FUCA, CURLI; mixed units: TOPSIS, AHP; uncertain/fuzzy: Fuzzy MCDM)
Use transparent software tools for traceable, reproducible computation and result dissemination (Wang et al., 2024, Najafi et al., 12 Feb 2025)

7. Emerging Challenges and Future Trends

Research directions include:

Explainable MCDM: Developing methods to extract compact, contrastive explanations and highlight key criteria responsible for decision differences, even in hierarchical or AI-driven workflows (Erwig et al., 2022).
Data-Driven and Adaptive Weighting: Integration of online learning, user feedback, and active learning to refine weights dynamically (Ahmed et al., 5 Apr 2025, Fujita et al., 16 Mar 2026).
Scalability and Dynamic Contexts: Efficient algorithms for very large alternative/criteria matrices, real-time process monitoring, adaptive decision support (Wang et al., 2024, Wang et al., 17 Feb 2025).
Hybrid, Uncertainty-Robust Pipelines: Frameworks synthesizing classical, fuzzy, probabilistic, Bayesian, and ML components into modular, robust workflows for group, sequential, and multi-agent decision environments (Fujita et al., 16 Mar 2026, Mohammadi, 2022).
Alignment with Human and Machine Reasoning: Aligning LLM chain-of-thought outputs and expert MCDM procedures, automating prompt construction, and enhancing the explainability and auditability of AI-generated decisions (Wang et al., 17 Feb 2025).

In summary, MCDM provides rigorous, extensible, and software-supported mechanisms for rational decision support in complex, multi-attribute environments, with active research shifting toward scalable, explainable, and uncertainty-aware methods bridging expert judgment, automation, and intelligent reasoning across application domains.