Takagi–Sugeno Fuzzy Model

Updated 27 February 2026

The Takagi–Sugeno fuzzy model is a rule-based framework that blends local affine submodels using normalized fuzzy membership functions to approximate nonlinear dynamics.
It is widely applied in nonlinear control, system identification, and classification, offering real-time adaptation and interpretability in complex systems.
Model identification and optimization are achieved through clustering, least squares regression, and LMI-based stability analysis for robust and efficient performance.

A Takagi–Sugeno (TS) fuzzy model is a structured, rule-based modeling framework in which a nonlinear mapping is defined via a collection of local affine (or more generally, linear-in-parameters) submodels blended by normalized weights derived from membership functions over premise variables. Each rule encapsulates system dynamics valid in a specific region of the input or state space; overall model output is computed through a convex combination of these local predictions. TS fuzzy models are widely used for identification, nonlinear control, and classification, especially in applications demanding interpretability, robust stability margins, or online adaptation.

1. Canonical Structure and Mathematical Formulation

A standard r-rule TS fuzzy model is a set of IF–THEN rules, each with a fuzzy antecedent and an affine consequent, for either algebraic regression or state-space evolution. The general continuous-time state-space TS model is

$\dot x(t) = \sum_{i=1}^r h_i\bigl(z(t)\bigr)\bigl[A_i\,x(t) + B_i\,u(t)\bigr]$

subject to

$\sum_{i=1}^r h_i\bigl(z(t)\bigr) = 1,\quad h_i(z)\ge0,$

where $z(t)$ is a vector of premise (scheduling) variables, $A_i$ , $B_i$ are the local linear system matrices for rule $i$ , and $h_i(z)$ are normalized weights computed from the product of membership functions (typically Gaussian or triangular) associated with each premise variable (Aldarraji et al., 2021, Mozelli et al., 2024, Brunner et al., 2024, Ganji et al., 2024).

For regression or classification, the rule set is

$R_i\!:~\text{IF}~x_1~\text{is}~A_{i1},~\dots,~x_n~\text{is}~A_{in}~\text{THEN}~y_i(x) = p_{i0} + \sum_{j=1}^n p_{ij} x_j.$

The aggregated output is

$y(x) = \sum_{i=1}^r \omega_i(x)\,y_i(x),$

where $\omega_i(x) = h_i(x) / \sum_{k=1}^r h_k(x)$ (Alves et al., 28 Apr 2025, Xu et al., 2019, Shi et al., 2020).

2. Fuzzy Rule Base: Antecedents, Consequents, and Inference

Rule Antecedents

The antecedent of each rule is a conjunction of $n$ fuzzy sets ( $A_{ij}$ ) on the premise variables.
Membership functions can be Gaussian, triangular, or other parametric forms, designed by data clustering (e.g., fuzzy c-means (Shi et al., 2020)) or expert heuristics.
Type-1 (point-valued) and interval type-2 (interval-valued) MFs are both employed. Interval type-2 models capture uncertainty and improve robustness under measurement noise or sparse data (Bouhentala et al., 2022, Sarbaz, 2022, Singh et al., 2020).

Rule Consequents

The local model can be zero-order (constant), first-order (affine), or higher order polynomials in the inputs or state.
For dynamic systems, ARX or NARX-type local models are used, as in

$y_i(k+1) = \sum_{j=1}^n a_{ij} y_i(k-j+1) + \ldots + c_i$

(Taeib et al., 2013).

Inference and Aggregation

The firing strength of rule $i$ is $h_i(x) = \prod_j \mu_{A_{ij}}(x_j)$ .
The output is the weighted sum (weighted average defuzzification), expressing the system as a global convex blend of local submodels.
For multi-output systems, either rules are constructed with vector-valued consequents (Beauchemin-Turcotte et al., 2017) or decoupled as parallel MISO submodels (Taeib et al., 2013).

3. Identification, Parameter Learning, and Structural Optimization

Model Identification

Antecedent parameters (MF centers/spreads) are typically initialized by clustering (e.g., fuzzy c-means (Shi et al., 2020), enhanced soft subspace clustering (Xu et al., 2019)).
Consequent parameters are estimated by least squares (for weighted regression), recursive least squares (Taeib et al., 2013), or Bayesian MAP with $\ell_2$ -regularization (Singh et al., 2020).
Rule-base sparsity is enforced through $\ell_1$ -norm constraints (LASSO) or proximal algorithms, enhancing interpretability and preventing overfitting (Xu et al., 2019, Lou et al., 2023).

Feature and Rule Selection

Soft subspace clustering and wrapper-based feature selection (e.g., genetic algorithms) are used to select relevant features per rule, reducing model complexity (Xu et al., 2019, Alves et al., 28 Apr 2025).
Ensemble architectures (random subspace bagging, random forest of TS models) further improve generalization on large, high-dimensional data (Alves et al., 28 Apr 2025).

Model Inversion

For multivariable models with affine consequents, analytical inversion (rule-wise left-inverse of the consequent matrix) allows explicit fuzzy inverse modeling, facilitating controller design and iterative learning (Beauchemin-Turcotte et al., 2017).

4. Control Design: Parallel Distributed Compensation and LMI Synthesis

Parallel Distributed Compensation (PDC)

The PDC strategy assigns a local state-feedback gain $K_i$ to each rule, constructing a global feedback via

$u(t) = \sum_{i=1}^r h_i(z(t))\,K_i\,x(t)$

(Aldarraji et al., 2021, Mozelli et al., 2024, Ganji et al., 2024).

Extensions include augmentation with membership-derivative weighted gains, yielding a two-term PDC controller with strictly less conservative LMI-based synthesis conditions (Mozelli et al., 2024).

Stability and Performance via LMIs

Lyapunov-based conditions are formulated as linear matrix inequalities (LMIs) on local or global Lyapunov matrices $P_i$ .
Additional LMIs address $H_\infty$ performance, positivity (via co-positive Lyapunov functions), region pole-placement, and delay robustness (Sarbaz, 2022, Ahmadi et al., 2019, Ganji et al., 2024).
For systems with type-2 uncertainty and time-varying delays, Razumikhin–Lyapunov approaches avoid state augmentation and maintain tractable online optimization for model predictive control (Sarbaz, 2022).

5. Interval Type-2 and Advanced Membership Function Designs

Interval Type-2 TS Models

Interval-valued MFs encode input uncertainty, yielding interval-valued firing strengths and so interval-valued model outputs (Bouhentala et al., 2022, Sarbaz, 2022).
Type-2 frameworks employ upper and lower MFs; final outputs are type-reduced (e.g., Karnik–Mendel algorithm (Singh et al., 2020)).
These methods have shown improved robustness to noise, outliers, and sparse data, with Bayesian MAP estimation procedures superior in regularizing consequent weights (Singh et al., 2020).

Student-t and Hybrid MFs

Heavy-tailed Student-t MFs are integrated to avoid zero assignment for outliers or sparse samples, increasing coverage and reducing sensitivity (Singh et al., 2020).

Subspace and Conciseness Strategies

Soft subspace clustering creates elastic rule antecedents, each operating on a locally optimal feature subspace, which greatly reduces rule length and enhances interpretability (Xu et al., 2019).

6. Application Domains and Representative Case Studies

TS fuzzy modeling has been demonstrated in diverse applications:

Process and systems control: PID and advanced feedback controller synthesis for nonlinear MIMO/NARX processes (Taeib et al., 2013), wind turbine modeling and control (Brunner et al., 2024), robotic manipulator stabilization (Aldarraji et al., 2021), blood glucose/diabetes management (Ganji et al., 2024), quadrotor attitude control with adaptive indirect SMC (Bouhentala et al., 2022), and cancer treatment positive systems (Ahmadi et al., 2019).
Time-series and regression: Renewable energy forecasting via new NTSK and ensemble fuzzy regressors (Alves et al., 28 Apr 2025), feature extraction in spectroscopy via TS-seeded broad learning networks (Wang et al., 2022), regression modeling with stochastic rule dropout and adaptive optimizers (Shi et al., 2020).
Classification: Multi-label extension of TSK for multi-label classification tasks, with joint label-correlation and sparsity optimization (Lou et al., 2023), concise models for medical and high-dimensional data (Xu et al., 2019).

Empirical evidence consistently indicates that TS fuzzy models, especially with interval type-2 and advanced clustering or regularization, offer high performance and interpretability, rivaling black-box ML models while remaining computationally tractable for real-time or embedded scenarios.

7. Algorithmic Innovations, Limitations, and Further Directions

Recent identification methods decouple the number of rules from the input dimensionality via independent MFs or target-increment–based rule partitioning, crucial for scalability (Shi et al., 2020, Alves et al., 28 Apr 2025).
Model pruning and feature selection—using soft subspace, clusterwise weights, or genetic optimization—achieve drastic parameter count reduction, addressing interpretability and overfitting (Xu et al., 2019, Alves et al., 28 Apr 2025).
Control-oriented TS models benefit from advances in structure-preserving positivity and customized Lyapunov synthesis (copositive, type-2, Razumikhin) (Ahmadi et al., 2019, Sarbaz, 2022).
Interval type-2 and Bayesian regularized structures are robust to noise, outliers, and data sparsity but incur extra computation in type-reduction and fuzzifier calibration (Bouhentala et al., 2022, Singh et al., 2020).
A plausible implication is that integrating data-driven, ensemble, and type-2 mechanisms is essential for future TS systems to remain competitive with modern deep learning architectures, particularly in highly uncertain regimes.