Enhancing Fuzzy Rule-Based Anfis Neuronal Fuzzy Architecture Through Integration with Binary Particle Swarm Optimization Technique for Low-Dimensional Data Modeling

Abstract

Fuzzy rule-based systems are instrumental in data interpretation, especially in scenarios dominated by low-dimensional data. While deep learning has revolutionized areas like image and speech recognition, its effectiveness diminishes in sparse, unstructured, or low-dimensional data. Besides, it requires vast parameter sets and substantial datasets for practical training and is prone to overfitting in data-scarce situations. Conversely, rule-based systems, underpinned by fuzzy logic, benefit from inherent interpretability. They offer insights into decision-making, with each rule providing a transparent rationale. This clarity is invaluable in sectors such as healthcare and finance, where comprehending the logic behind decisions is essential. However, these systems often falter with intricate and high-dimensional data. A strategy to counter these challenges is to merge rule-based systems with other machine-learning techniques. A prime example is the Adaptive Neuro-Fuzzy Inference System (ANFIS), which combines the transparency of fuzzy rule-based systems with the adaptability of neural networks. However, traditional ANFIS has limitations, especially when using grid partitioning for rule generation. A significant drawback is the exponential growth in the rule count as the problem’s dimensionality increases despite its simplicity in implementation. In response to the challenges of rule-based systems, especially concerning ANFIS with grid partitioning, this research introduces two innovative models for strategic rule reduction. The first model employs Binary Particle Swarm Optimization (BPSO) with ANFIS as a feature selector for the normalized firing strength, effectively pruning the rule set while maintaining system integrity. The subsequent model incorporates Principal Component Analysis (PCA) on the normalized firing strengths first, transforming them into a linearly uncorrelated set of components. These components are then selectively optimized and evaluated using Binary Particle Swarm Optimization (BPSO), ensuring a comprehensive and impactful reduction. This method minimizes the rule set and ensures decision-making precision. Additionally, a custom parameter update mechanism fine-tunes specific ANFIS layers. Updating both Inertia and Acceleration Coefficients dynamically to adjust BPSO parameters, bypassing potential local minima issues. The effectiveness of these enhancements has been validated on standard datasets from UCI respiratory and keel, covering classification and regression tasks, and a real-world ischemic stroke dataset from Hamad Medical Corporation, emphasizing the models’ adaptability and practicality.