Fuzzy Cognitive Maps (FCM)
- Introduced by
- Bart Kosko
- Year
- 1986
- Type
- Fuzzy causal/feedback network for scenario analysis
- Family
- Cognitive mapping
What does it do?
A fuzzy cognitive map, introduced by Bart Kosko in 1986, represents a system as a network of concepts connected by signed, weighted causal links, and simulates how the concepts influence one another over time.
By combining the intuitive structure of a cognitive map with fuzzy weights and iterative activation, FCMs let experts encode causal knowledge and then run what-if scenarios — making them popular for policy analysis, strategic decision-making, and modelling complex socio-technical systems.
The intuition behind the method
Experts often understand a system as 'more of this causes more of that, but less of the other' — a web of causal arrows of varying strength and sign. An FCM turns that mental model into a directed graph: nodes are concepts (e.g., investment, demand, pollution), arrows carry weights between −1 and +1 for how strongly and in which direction one concept pushes another. To explore consequences, you set initial concept values and let activation propagate around the network, each concept updating from the weighted influences of the others and being squashed into a bounded range, until the system settles. The resulting steady state shows where the system's feedback loops drive it under a given scenario.
History
Bart Kosko introduced fuzzy cognitive maps in 1986, extending Axelrod's earlier (binary) cognitive maps with fuzzy weights and dynamic activation, uniting fuzzy logic with neural-network-style propagation. The method matured through the 1990s–2010s with learning algorithms (Hebbian and evolutionary weight learning), extensions (dynamic, grey, and intuitionistic FCMs), and a broad applications literature surveyed by Papageorgiou and Salmeron (2013). FCMs sit at the intersection of soft computing, system modelling, and decision support, complementing system dynamics and Bayesian networks.
When to use it
Use fuzzy cognitive maps to model and simulate systems dominated by causal feedback among qualitative or semi-quantitative concepts, especially when expert knowledge is the main data source and the goal is scenario and policy exploration rather than precise numerical prediction — common in environmental management, medical decision support, engineering control, and strategic planning. They are valuable for participatory modelling (stakeholders can build and inspect the map) and for combining several experts' maps. Limitations: the dynamics depend on the chosen squashing function and weights, FCMs capture relative causal influence rather than calibrated magnitudes, convergence behaviour can be sensitive, and they assume the causal structure is correct. When precise quantitative dynamics or rigorous probabilistic inference are required, system dynamics or Bayesian networks may be more appropriate.
Limitations and considerations
- Captures relative causal influence, not calibrated quantitative magnitudes.
- Outcomes depend on the squashing function, weights, and update scheme.
- Convergence can be sensitive and may yield limit cycles or chaos.
- Assumes the elicited causal structure is correct; missing links bias results.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Encodes expert causal knowledge in an intuitive, inspectable graph.
- Simulates feedback dynamics and supports rapid what-if scenario analysis.
- Handles qualitative/semi-quantitative concepts without precise data.
- Supports participatory modelling and merging of multiple experts' maps.
Applications
- Environmental and ecological management scenario analysis.
- Medical decision support and diagnosis from causal concept models.
- Strategic planning and policy what-if analysis with stakeholders.
- Engineering process control and fault analysis.
- Modelling socio-economic systems where feedback loops dominate.
How it works
How it works
Define the concepts C₁…Cₙ that matter in the system and the causal weight matrix W, where wᵢⱼ ∈ [−1, 1] encodes the strength and sign of concept i's influence on concept j. Weights come from expert judgement, from data, or from learning algorithms; positive means promoting, negative means inhibiting.
Formula
C_1,\ldots,C_n;\quad W = [w_{ij}] \in [-1,1]^{n\times n}\ (\text{causal weights})Each concept has an activation value. At every step, a concept's new value is its previous value plus the weighted sum of the activations of all concepts pointing into it. This is the propagation of causal influence through the network's feedback structure.
Formula
A_i^{(t+1)} = f\!\left(A_i^{(t)} + \sum_{j\neq i} w_{ji}\,A_j^{(t)}\right)The summed influence is passed through a squashing (threshold) function — commonly a sigmoid with steepness λ, or a sign/hyperbolic-tangent function — to keep activations in a bounded range (e.g., [0,1] or [−1,1]) and to introduce nonlinearity, so concepts cannot grow without limit.
Formula
f(x) = \dfrac{1}{1+e^{-\lambda x}}\quad (\text{sigmoid squashing})The update is iterated until the activations stop changing (a fixed-point attractor), settle into a repeating limit cycle, or are declared chaotic. The converged pattern is the scenario's outcome; comparing converged states across different initial settings or weight edits is how FCMs answer what-if and policy questions.
Formula
\text{iterate until } A^{(t+1)} \approx A^{(t)}\ (\text{fixed point / limit cycle})Common pitfalls
- Reading converged concept values as precise predictions rather than relative tendencies.
- Ignoring how the choice of threshold function and λ changes the steady state.
- Building maps with unvalidated expert weights and treating them as ground truth.
- Overlooking non-convergence (limit cycles) and reporting a transient state as the result.
Frequently asked questions
How is an FCM different from a Bayesian network?
Both are directed graphs over concepts, but a Bayesian network encodes conditional probability distributions and is acyclic, supporting probabilistic inference. An FCM uses signed fuzzy weights, allows cycles (feedback), and is simulated dynamically to a steady state. FCMs are better for feedback-rich what-if scenarios; Bayesian networks for rigorous probabilistic reasoning.
Where do the weights come from?
Typically from expert judgement (often elicited as linguistic terms mapped to numbers), from historical data, or from learning algorithms such as Hebbian or evolutionary weight tuning. Because outcomes depend on the weights, their source and validation should always be reported.
What does it mean if the FCM doesn't converge?
Instead of settling to a fixed point, the system can enter a limit cycle (repeating pattern) or chaotic behaviour, depending on the weights and squashing function. Non-convergence is itself informative about the system's feedback structure, but a transient state should not be reported as the scenario outcome.
Related methods
Sources
Kosko, B. (1986). Fuzzy cognitive maps. International Journal of Man-Machine Studies, 24(1), 65–75.
Open DOI referenceCopied.Papageorgiou, E. I., & Salmeron, J. L. (2013). A review of fuzzy cognitive maps research during the last decade. IEEE Transactions on Fuzzy Systems, 21(1), 66–79.
Open DOI referenceCopied.