A Bayesian Approach with Type-2 Student-tMembership Function for T-S Model Identification

Vikas Singh, Homanga Bharadhwaj, Nishchal K. VermaarXiv 2020 · arXiv

Abstract

This paper presents a robust Bayesian framework for identifying Takagi-Sugeno (T-S) fuzzy models, specifically optimized for sparse datasets. By integrating a novel Type-2 Student-t membership function with $L_2$ regularization and Stochastic Gradient Descent (SGD), the proposed approach effectively handles data uncertainty and prevents overfitting. Experimental results on house price prediction, non-linear plant modeling, and the Sinc function demonstrate superior performance compared to existing state-of-the-art methods.

Introduction

Takagi-Sugeno (T-S) fuzzy models are widely utilized for mapping complex, non-linear systems through a set of "if-then" linguistic rules. While the introduction of Type-2 fuzzy logic has enhanced these models' ability to manage data uncertainty, traditional identification methods—such as fuzzy c-regression—often struggle when data is sparse. In such scenarios, models frequently overfit to noise rather than capturing the underlying signal. Furthermore, standard Gaussian membership functions (MFs) lack the flexibility required to represent the distributions of sparse or outlier-heavy data. This research addresses these gaps by integrating Bayesian principles into the fuzzy modeling process.

Methodology

The proposed framework introduces several key innovations to improve model identification and generalization:

  1. Hybrid Membership Function (MF): The core innovation is a membership function that blends a Gaussian distribution with a Student-t distribution. The Student-t component, characterized by "heavier tails," is significantly better at quantifying the effects of sparse data and outliers. A hyperparameter ($\alpha$) allows the model to adjust the influence of the Student-t component based on data density.
  2. Bayesian MAP Estimation: Rather than relying on simple clustering, the authors employ Maximum A-Posteriori (MAP) estimation for fuzzy-space partitioning. This allows for the incorporation of a Gaussian prior on the regression coefficients, providing a probabilistic foundation for the model.
  3. $L_2$ Regularization: To further combat overfitting, $L_2$ regularization is applied during the identification of antecedent parameters, penalizing excessively large coefficient values.
  4. Optimization Pipeline: The model's consequent parameters are optimized using Stochastic Gradient Descent (SGD). For type-reduction—converting complex Type-2 fuzzy sets into crisp outputs—the researchers utilize the Karnik-Mendel (KM) algorithm.

Experimental Evaluation

The model was evaluated against three distinct benchmarks to test its versatility and precision.

House Prices Dataset

Using a real-world sparse dataset from Kaggle, the model predicted property sale prices. The proposed method achieved a Mean Square Error (MSE) of 0.008, significantly outperforming Logistic Regression (0.06) and previous Type-2 fuzzy classifiers (0.035). It also reached a high Coefficient of Determination ($R^2$) of 0.85.

Actual vs. Model Output for House Prices

Non-Linear Plant Modeling

To test dynamic stability, the framework was applied to a second-order non-linear difference equation. The model achieved an MSE of $7.2 \times 10^{-5}$, an order of magnitude improvement over the Modified Inter Type-2 FCRM ($1.02 \times 10^{-4}$).

Sinc Function Precision

The standard 1D non-linear Sinc function ($y = \sin(x)/x$) was used to evaluate regression precision. The proposed method achieved an MSE of $2.4 \times 10^{-3}$, consistently outperforming established methods such as EFCM and MIT2-FCRM.

Comparison of MSE across different methods for the Sinc function

Discussion and Conclusion

A primary insight of this work is that sparsity requires "heavier tails" in membership functions; the Student-t distribution acts as a natural regularizer, making the model less sensitive to gaps in sparse data. Additionally, the integration of $L_2$ regularization directly into the partitioning phase of T-S models significantly boosts generalization on unseen data.

While the results are state-of-the-art, the computational overhead of combining SGD with the iterative Karnik-Mendel algorithm warrants further investigation for large-scale applications. Future work will focus on automating hyperparameter tuning (currently handled via grid search) and exploring the scalability of this Bayesian approach in high-dimensional feature spaces.

Summary updated

BibTeX
@misc{singh2020bayesian,
  author = {Vikas Singh and Homanga Bharadhwaj and Nishchal K. Verma},
  title = {{A Bayesian Approach with Type-2 Student-tMembership Function for T-S Model Identification}},
  year = {2020},
  eprint = {2009.00822},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.00822},
}
Figure