Introduction
Presented at AAAI-91, this research establishes a foundational framework for neuro-fuzzy systems. It introduces a methodology to bridge the gap between the linguistic interpretability of fuzzy logic and the computational learning power of neural networks. This work laid the groundwork for what would eventually become the Adaptive Neuro-Fuzzy Inference System (ANFIS) architecture.
Problem Statement
In the early 1990s, modeling complex, non-linear systems—such as economic or ecological models—posed a significant challenge. Traditional mathematical models often lacked the flexibility to handle uncertainty, while fuzzy inference systems, which use "if-then" rules to mimic human reasoning, suffered from two primary limitations:
- Knowledge Engineering Bottleneck: Translating human expertise into precise fuzzy rules was a manual and subjective process.
- Lack of Adaptability: Existing fuzzy systems lacked standard algorithms to "learn" from data or automatically tune membership functions to minimize error.
The proposed solution automates the creation and refinement of these models by embedding fuzzy logic within a neural network structure.
Methodology: The Generalized Neural Network (GNN)
The core innovation is the Generalized Neural Network (GNN). Unlike traditional neural networks where weights are primarily associated with links, a GNN places functions and parameters directly within the nodes. The architecture maps a fuzzy inference system onto a five-layer feed-forward network:
- Layer 1 (Premise): Defines membership functions (e.g., "If temperature is High"). Parameters here determine the shape of the fuzzy sets.
- Layers 2 & 3 (Rule Processing): Calculate rule firing strengths and perform normalization.
- Layer 4 (Consequence): Implements linear equations of the inputs (e.g., "Then volume is Small").
- Layer 5 (Output): Aggregates the results into a single numerical output.
Hybrid Learning Algorithm
To optimize the GNN, the author proposed a hybrid learning approach:
- Gradient Descent: Employed to update the non-linear "premise" parameters in Layer 1.
- Kalman Filter Algorithm: Used to identify the "consequence" parameters in Layer 4. Since the final output is a linear combination of these parameters, the Kalman filter can identify optimal values much faster than standard gradient descent, significantly accelerating convergence and avoiding local minima.
Experimental Results
The GNN was evaluated against three synthetic datasets to test its function approximation capabilities.
- Example 1: A surface generated by a known fuzzy system. The model achieved an Average Percentage Error (APE) of 1.56% after 200 epochs.
- Example 2: A highly non-linear function involving trigonometric and exponential components. The GNN achieved an APE of 0.46%.
- Example 3 (Benchmark): A 3-input non-linear function used in previous landmark studies. The GNN achieved an APE of 0.01357%, outperforming previous state-of-the-art methods by Sugeno and Kang (0.59%) and Kondo (4.7%) by nearly 40 times.
Conclusion and Future Work
The GNN approach demonstrates that fuzzy inference systems are functionally equivalent to specific feed-forward neural networks. This allows for "gray-box" modeling, where prior human knowledge can be used to initialize the network, which is then fine-tuned through data-driven learning. While the current model requires a pre-specified number of rules, future work points toward using clustering techniques to automate structural identification, further reducing the need for manual intervention.
Summary updated