DombiTSK
DombiTSK replaces product aggregation with an adaptive Dombi t-norm in the antecedent, providing more flexible high-dimensional rule activation behavior.
Reference
G. Xue, L. Hu, J. Wang and S. Ablameyko, "ADMTSK: A High-Dimensional Takagi–Sugeno–Kang Fuzzy System Based on Adaptive Dombi T-Norm," in IEEE Transactions on Fuzzy Systems, vol. 33, no. 6, pp. 1767-1780, June 2025, doi: 10.1109/TFUZZ.2025.3535640.
Mathematical Formulation
Antecedent
Each rule-term membership is computed with a Gaussian function:
where \(c_{r,d}\) is the center and \(\sigma_{r,d}>0\) is the spread.
Dombi aggregation
DombiTSK uses the Dombi t-norm to combine the antecedent membership values:
The hyperparameter \(\lambda > 0\) controls the aggregation shape:
- small \(\lambda\) makes the aggregation softer and closer to product-like behavior,
- large \(\lambda\) makes the aggregation sharper and closer to minimum-like behavior.
Normalization
Rule firing strengths are normalized by sum normalization:
Consequent
DombiTSK uses a first-order TSK consequent for both classification and regression.
For classification:
For regression:
Output aggregation
The final prediction is the normalized weighted sum of rule consequents:
- Classification:
- Regression:
Code ↔ Paper Correspondence
| Equation | Class / Method | Description |
|---|---|---|
| Dombi aggregation | DombiTSKClassifier / DombiTSKRegressor |
Fixed-λ Dombi antecedent aggregation with t_norm="dombi" |
| Normalization | SumBasedDefuzzifier |
Sum-based rule strength normalization |
| Consequent | ClassificationConsequentLayer / RegressionConsequentLayer |
First-order linear consequents |
| Membership functions | GaussianMF |
Standard Gaussian antecedent MFs |
Implementation notes
DombiTSKClassifierandDombiTSKRegressoruse a fixed Dombi parameterlambda_ > 0in the antecedent and default toSumBasedDefuzzifier.DombiTSKClassifierEstimatorandDombiTSKRegressorEstimatorare sklearn-compatible wrappers that build the rule base and membership functions frominput_configs,n_mfs,mf_init, andsigma_scale.- The estimators default to
mf_init="kmeans"andsigma_scale=1.0. - The default
rule_basefor estimator-built models is"coco"withmf_init="kmeans"and"cartesian"withmf_init="grid". CompositeGaussianMFis available when a positive lower bound on antecedent membership values is desired, supporting ADMTSK-style stability.AdaptiveDombiRuleLayeris implemented in the codebase and provides per-rule adaptive Dombi exponents, but there is currently no dedicatedADMTSKwrapper class exposing this behavior directly.
Alignment with the paper
- The paper defines a Dombi TSK baseline with Dombi antecedent aggregation and first-order consequent structure.
- highFIS implements this baseline directly through
DombiTSKClassifierandDombiTSKRegressor. - Rule strengths are normalized by sum-based defuzzification, matching the paper's TSK output aggregation.
- The package also includes building blocks for the ADMTSK extension:
CompositeGaussianMFandAdaptiveDombiRuleLayer. - A full ADMTSK-style model would require combining these components with an adaptive λ selection mechanism, which is not currently wrapped by a single model class.