Feedback Fuzzy Cognitive Maps

• Feedback Fuzzy Cognitive Maps are nonlinear dynamical systems defined as weighted, directed graphs with cycles that capture recursive causal mechanisms.
• They employ iterative update rules and fuzzy weights to simulate complex phenomena in policy analysis, time series forecasting, and explainable AI.
• Learning methods such as Hebbian updates, evolutionary algorithms, and gradient techniques help synthesize expert knowledge with data-driven insights.

Feedback Fuzzy Cognitive Maps (FCMs) are nonlinear, recurrent, graph-theoretic causal models in which concept nodes are interlinked by weighted, directed edges that explicitly encode both degree and direction of causal influence. The defining characteristic of feedback FCMs is the presence of cycles in the underlying weighted digraph, enabling direct modeling of feedback mechanisms in domains where acyclic graphical models, such as DAGs, are structurally inadequate. FCMs leverage fuzzy edge weights, iterative update rules, and a flexible combination framework to support simulation, analysis, and synthesis across a wide range of applications, including policy analysis, system identification, classification, time series forecasting, and explainable AI, often at scales and levels of interpretability inaccessible to probabilistic or strictly neural-network-based approaches (Osoba et al., 2019, Szwed, 2021, Mkhitaryan et al., 2021, Panda et al., 2024, Panda et al., 29 Sep 2025, Obiedat et al., 2022, Orang et al., 2022, Panda et al., 31 Dec 2025).

1. Mathematical Formalism and Dynamical Properties

A feedback Fuzzy Cognitive Map is defined as a signed, weighted directed graph $G=(C,E, W)$, where:

• $C = \{C_1, ..., C_n\}$ is the node set of concepts or variables;
• $E \subseteq C \times C$ is the set of directed edges;
• $W = [w_{ij}] \in [-1,1]^{n \times n}$ is the adjacency (weight) matrix; $w_{ij}>0$ denotes a positive causal influence from $C_i$ to $C_j$, $w_{ij}<0$ a negative (inhibitory) influence, and $w_{ij}=0$ absence of direct causality.

The system state at discrete time $t$ is the activation vector $x^{(t)} = (x_1^{(t)}, ..., x_n^{(t)}) \in [0,1]^n$, with $x_i^{(t)}$ interpreted as the fuzzy degree of activation of concept $C_i$.

The canonical synchronous update rule is: $x^{(t+1)} = f\left( x^{(t)} W + b \right)$ where $f(\cdot)$ is an elementwise nonlinearity (e.g., sigmoid, hard threshold, hyperbolic tangent) and $b \in \mathbb{R}^n$ is a bias vector. Feedback arises from cycles in $W$ (i.e., the existence of $k$, $i_1,...,i_k$ such that $w_{i_1,i_2} w_{i_2,i_3} \cdots w_{i_k,i_1}\neq 0$).

These recurrences render FCMs as nonlinear dynamical systems. Depending on $W$ and $f$, asymptotic behaviors include:

• Fixed-point attractors: $x^{(t)}\to x^*$;
• Limit cycles: periodic orbits in activation space;
• Chaotic attractors for certain $W$ and $f$ configurations (Osoba et al., 2019, Orang et al., 2022, Mkhitaryan et al., 2021).

Feedback loops permit modeling of inertia, memory, and recursive causation absent in DAG-based formalisms.

2. Causal Transitivity and Influence Structure

Transitive causal influence is a distinguishing property of FCMs (Osoba et al., 2019). For a simple directed path $C_{j_1} \rightarrow ... \rightarrow C_{j_{k+1}}$, if each activation is a smooth, nondecreasing function of its input, the infinitesimal effect of $C_{j_1}$ on $C_{j_{k+1}}$ is: $$\frac{\partial C_{j_{k+1}}}{\partial C_{j_1}} = \prod_{\ell=1}^k w_{j_\ell, j_{\ell+1}} \cdot \psi \quad \text{with} \quad \psi = \prod_{\ell=2}^{k+1} \frac{dC_{j_\ell}}{dx_{j_\ell}} \geq 0$$ The composite influence along all possible paths from $C_r$ to $C_s$ is the sum over all acyclic paths: $$D(C_r, C_s) = \sum_{m=1}^M \partial(C_r, C_s; \lambda_m)$$ for all M paths $\lambda_1,...,\lambda_M$. For sufficiently small $\|W\|$, the total effect admits the linear approximation $D = (I-W)^{-1} - I$. This explicit accounting for transitive, feedback-mediated influences is unavailable in general in probabilistic DAGs (Osoba et al., 2019).

3. Learning, Synthesis, and Expert Fusion

Learning the weight matrix $W$ from data or expert knowledge is critical. Three methodological classes dominate (Orang et al., 2022, Mkhitaryan et al., 2021):

• Hebbian-based updates: E.g., Nonlinear Hebbian Learning, Differential Hebbian Learning (DHL), and Active Hebbian Learning (AHL), iteratively adjust $w_{ij}$ proportional to product of activations or their differences at successive time steps;
• Population-based global optimization: Genetic algorithms, Particle Swarm Optimization, and other metaheuristics optimize $W$ against sequence error criteria;
• Gradient-based approaches: Backpropagation through the feedback layers, as in d-step classifiers, enables end-to-end training with loss functions such as cross-entropy or MSE (Szwed, 2021, Mkhitaryan et al., 2021).

Combining multiple FCMs from experts or systems is performed by weighted matrix averaging: $$\bar{W} = \sum_{k=1}^m \alpha_k W^{(k)}, \quad \alpha_k \geq 0, \sum_k \alpha_k = 1$$ This operation is closed in the space of valid FCMs and enables coherent aggregation, as shown empirically to approximate the ground-truth equilibrium dynamics as the expert sample grows (Osoba et al., 2019). Multi-expert mixtures with phantom node estimation support scalable approximation to large systems with latent structure via convex combination of augmented maps. Learning is performed to fit the dynamical equilibria (fixed points or limit cycles) of the underlying system, often in parallelizable subdivided steps (Panda et al., 2024).

4. Interpretability, Explainable AI, and LLM-based Extraction

Recent research demonstrates FCMs' capacity for explainable modeling and extraction from unstructured data. LLM agents, guided by multi-stage prompting (noun extraction, concept filtering, fuzzy edge inference), recover feedback FCMs from raw text, enabling explainable AI workflows (Panda et al., 31 Dec 2025, Panda et al., 29 Sep 2025). In these pipelines:

• Detected concept nodes are those denoting variables that admit magnitude changes and engage in causal statements.
• Edge extraction links concepts with polarity and fuzzy-valued strength mapped from verb phrases.
• The resulting FCMs display equilibria (attractors, limit cycles) consistent with human-crafted maps, even with discrepancies in granularity.

An autoencoder-like mapping (FCM→text→FCM) is realized, with the LLM acting as a transparent encoder/decoder. Strong edges are prioritized and preserved, while weak links may be pruned during naturalization of the textual latent, resulting in a controllable trade-off between fidelity and human readability. Reconstruction error is quantified by $\ell_1$, $\ell_2$, $\ell_\infty$ norms on adjacency matrices and edge-preservation rate, with empirical preservation of nearly all strong cycles (Panda et al., 29 Sep 2025).

5. Applications in Policy Analysis, Classification, and Forecasting

Feedback FCMs have demonstrated versatility across diverse application domains:

• Policy scenario analysis: US–China Thucydides-trap dynamics, public support for insurgency scenarios, and water crisis mitigation. Large stakeholder-derived FCMs (hundreds of nodes) can be condensed via graph-theoretic centrality (Consensus Centrality Measure) and aggregated using fuzzy 2-tuple linguistic frameworks. Policy interventions are simulated via clamping activation to track system response and rank strategies by impact using multi-criteria fuzzy appropriateness measures (Obiedat et al., 2022, Osoba et al., 2019).
• Classification and feature transformation: FCM classifiers function as recurrent maps of fixed depth $d$; feedback-induced nonlinear feature transformation enhances separability, and gradient-based learning enables competitive accuracy on benchmark datasets. FCMs interoperate as pre-processing engines, boosting the performance of linear and density-based classifiers (Szwed, 2021).
• Time series prediction: By leveraging feedback, FCMs model inertia, accumulation, and higher-order causal structure in sequence modeling tasks. Hybrid designs (e.g., FCM+ARIMA, high-order FCMs, wavelet/EMD decompositions) explicitly embed feedback to stabilize complex, nonlinear system prediction (Orang et al., 2022).
• Intervention and “what-if” analysis: Feedback propagates policy shocks throughout the causal web, optionally quantified by difference in equilibria on intervention vs. baseline (Mkhitaryan et al., 2021).

6. System Integration, Software, and Computational Considerations

Open-source frameworks such as FCMpy provide full-lifecycle construction, simulation, learning, and intervention tools, supporting feedback through arbitrary cycles in $W$ and broad algorithmic support (NHL, AHL, RCGA, deterministic methods) (Mkhitaryan et al., 2021). Numerical stability is controlled via bounded transfer functions and stopping criteria on designated outputs.

For large systems or mixtures, computational complexity is mitigated by:

• Parallelizing expert FCM tuning and incremental phantom node estimation;
• Exploiting sparse block structures;
• Efficiently fusing sparse matrices.

LLM-based extraction and combination pipelines for FCMs introduce new dimensions of scalability and explainability, leveraging both data- and knowledge-driven modalities (Panda et al., 31 Dec 2025, Panda et al., 29 Sep 2025).

7. Advantages, Limitations, and Theoretical Distinctions

Feedback FCMs offer interpretability (semantic nodes and edges), direct modeling of feedback/inertia, and compositionality—features generally lacking in probabilistic DAGs and black-box neural architectures. They support both knowledge engineering and data-driven optimization. Transitivity of causal influence is mathematically axiomatic, in contrast to non-transitive probabilistic influence in Bayesian settings (Osoba et al., 2019).

However, FCMs exhibit:

• Sensitivity to weight scale and transfer function in dynamic stability, with possible emergence of unwanted cycles or chaos;
• Scalability limitations for naïvely full-matrix weight learning in large systems unless structure is exploited (Orang et al., 2022);
• Reduced numerical precision relative to full probabilistic inference;
• Risk of interpretability loss when feedback structures become excessively dense or when natural language extraction prunes weak, but real, edges (Panda et al., 29 Sep 2025, Panda et al., 31 Dec 2025).

Nevertheless, feedback FCMs remain a uniquely suited framework for simulating, synthesizing, and explaining causally dense dynamical systems, especially when human interpretability, flexible learning, and feedback-rich dynamics are essential.