highFIS

highFIS is a modern PyTorch library for high-dimensional Takagi-Sugeno-Kang (TSK) fuzzy systems. It delivers differentiable, trainable fuzzy inference for classification and regression, with sklearn-compatible estimators for fast experimentation.

Why highFIS?

• Built for high-dimensional data and numerical stability.
• Supports adaptive and gated fuzzy inference, including feature selection and rule extraction.
• Includes HDFIS variants for product T-norm and minimum T-norm high-dimensional inference.
• Ships sklearn-compatible estimators (*Classifier / *Regressor) for every model family.
• Works seamlessly with Pipeline, GridSearchCV, and standard scikit-learn workflows.

High-level overview

highFIS models combine:

• differentiable membership functions for antecedent fuzzification,
• configurable rule bases and T-norm aggregation,
• normalized rule weights via defuzzification,
• built-in metrics and evaluation utilities for regression and classification,
• task-specific consequent layers for classification or regression.

Use a model class from highfis.models for custom PyTorch pipelines, or import an estimator directly from highfis for sklearn-compatible training.

Quick Start

pip install highfis

from highfis import HTSKClassifier

clf = HTSKClassifier(
    n_rules=10,
    mf_init="kmeans",
    epochs=150,
    learning_rate=1e-3,
    random_state=42,
)
clf.fit(X_train, y_train)
print(f"Test accuracy: {clf.score(X_test, y_test):.4f}")

Models Available

highFIS includes the following concrete TSK model families:

• TSK — vanilla TSK with product antecedent aggregation and sum-based normalization.
• HTSK — high-dimensional TSK with geometric mean aggregation and log-space softmax normalization.
• LogTSK — inverse-log normalization of log-domain rule weights for stable high-dimensional inference.
• HDFIS — high-dimensional inference with both product-DMF aggregation (HDFIS-prod) and minimum frozen-antecedent inference (HDFIS-min).
• DombiTSK — Dombi parametric aggregation with a learnable shape parameter.
• ADMTSK — adaptive Dombi TSK with dimension-dependent Gaussian membership functions.
• AYATSK — Yager-style aggregation for more flexible antecedent behavior.
• ADATSK — adaptive softmin aggregation with dynamic rule weighting.
• ADPTSK — adaptive double-parameter softmin aggregation with stable normalized rule weights.
• FSRE-ADATSK — gated feature selection and rule extraction inside an adaptive inference pipeline.
• DGTSK — double-gated training for simultaneous feature selection and rule extraction.
• DGALETSK — adaptive Ln-Exp softmin with embedded feature and rule gates for sparse high-dimensional modeling.
• MHTSK — multihead sparse subantecedents for high-dimensional rule extraction and scalable TSK learning.

Each model family exposes both classifier and regressor variants.

Model selection guide

• Choose TSK for a baseline vanilla fuzzy model.
• Choose HTSK or LogTSK for high-dimensional problems where numerical stability is critical.
• Choose DombiTSK or AYATSK when you want more control over antecedent aggregation behavior.
• Choose HDFIS when you need high-dimensional inference with either a dimension-dependent product antecedent or a frozen minimum antecedent.
• Choose ADATSK, FSRE-ADATSK, DGTSK, or DGALETSK when you need adaptive sparsity, feature gating, or rule extraction.

Documentation

Get Started

Import estimators directly from highfis for sklearn-compatible usage, or access PyTorch model classes via highfis.models for custom training loops.