Neuro-Fuzzy Toolbox

A PyTorch-based library for the design, training, and analysis of ANFIS-based neuro-fuzzy models. The toolbox provides ready-to-use model variants and training algorithms, a structural adaptation algorithm, and utilities for rule inspection and local contribution analysis. Its modular design also makes it a flexible basis for building custom training procedures and deep neuro-fuzzy architectures.

Installation

pip install neuro-fuzzy-toolbox

Features

• Three ANFIS model variants: classical ANFIS, homogeneous h_ANFIS, and rule_reduced_ANFIS for high-dimensional settings.
• Multiple training strategies: hybrid learning algorithm, single-optimizer training, and dual-optimizer training with independent premise and consequent optimizers. All strategies integrate with PyTorch loss functions and support early stopping.
• Structural adaptation: a modified SONFIS algorithm for rule_reduced_ANFIS models, supporting rule growing, splitting, and pruning during training.
• Rule inspection and analysis: tabular export of premises and consequents, membership function visualization, intermediate layer access, and local rule-contribution analysis via RulesAnalyzer.
• Low-level API: direct access to premise and consequent parameter subsets for custom optimizer instantiation, and programmatic rule addition and removal at runtime.

Requirements

• torch >= 2.5 (tested with 2.5.1)
• numpy >= 2.2 (tested with 2.2.1)
• pandas >= 2.2 (tested with 2.2.3)
• matplotlib >= 3.10 (tested with 3.10.0)

Documentation

Full documentation including a usage guide, API reference, and end-to-end examples is available at: https://neuro-fuzzy-toolbox.readthedocs.io/en/latest/