v0.1.0 - Initial release

Neuro-Fuzzy Toolbox v0.1.0 — Initial Release

This is the first public release of the Neuro-Fuzzy Toolbox, a PyTorch-based library for the design, training, and analysis of ANFIS-based neuro-fuzzy models.

Overview

The toolbox provides a modular framework that explicitly separates model definition, internal layers, membership functions, learning algorithms, and rule-analysis utilities, making each component independently accessible and extensible.

Models

• ANFIS — General model with an arbitrary number of membership functions per input variable.
• h_ANFIS — Homogeneous variant in which all input variables share the same number of membership functions, improving computational efficiency.
• rule_reduced_ANFIS — Rule-reduced variant that avoids the classical combinatorial construction of rules, resulting in a more flexible and efficient model suited for high-dimensional inputs.

Training strategies

• Hybrid learning algorithm — Inspired by the classical ANFIS scheme proposed by Jang (1993), combining least-squares estimation for consequent parameters with gradient descent for premise parameters.
• Basic optimizer training — Single-optimizer gradient-based training using any PyTorch optimizer.
• Double optimizer training — Separate optimizers for premise and consequent parameters.
• SONFIS — Self-organizing algorithm for structural adaptation of rule-reduced models, incorporating rule growing, splitting, and pruning during training.

Membership functions

• Gaussian (Gaussian_MF)
• Generalized Bell (GeneralizedBell_MF)
• High-Slope Bell (HighSlopeBell_MF)

Rule inspection and analysis

• Tabular export of premises and consequents as pandas DataFrames via get_rules_structure()
• Visualization of membership functions via plot_premises()
• Access to intermediate layer outputs
• Local rule-contribution analysis via RulesAnalyzer

Low-level API

For users requiring finer control, model parameters are organized and accessible by component, so premises and consequents can be retrieved independently and passed to PyTorch optimizers over arbitrary parameter subsets. Rules can also be added (add_rules()) and removed (remove_rules()) programmatically at runtime (only for the rule_reduced_ANFIS class).

Installation

pip install neuro-fuzzy-toolbox

Requirements

• Python 3.10–3.12 (tested in 3.12.3)
• PyTorch >= 2.5 (tested in 2.5.1)
• NumPy >= 2.2 (tested in 2.2.1)
• Pandas >= 2.2 (tested in 2.2.3)
• Matplotlib >= 3.10 (tested in 3.10.0)

Documentation

Full documentation is available at: https://neuro-fuzzy-toolbox.readthedocs.io/en/latest/

License

MIT License

Note: The installation page in the docs shipped in Spanish in this release.
This has been corrected in the main branch. See the
latest documentation.