A granular reflex fuzzy min-max neural network for classification.

Area of Science:

• Pattern Recognition
• Artificial Intelligence
• Computational Intelligence

Background:

• Traditional computing relies on numerical or symbolic data manipulation.
• Human recognition excels at processing information granules alongside numerical values.
• Classifying and clustering granular data is a growing challenge in pattern recognition.

Purpose of the Study:

• To propose a novel granular neural network (GNN) for granular data classification.
• To introduce the granular reflex fuzzy min-max neural network (GrRFMN) capable of learning and classifying granular data.
• To address the challenge of data granularity in classification tasks.

Main Methods:

• The GrRFMN utilizes hyperbox fuzzy sets to represent granular data.
• An architecture with a reflex mechanism inspired by the human brain is employed to manage class overlaps.
• Online training is supported for both granular and point data, with specialized neuron activation functions for diverse data granularities.
• A data granulation preprocessing technique is investigated to enhance classifier performance.

Main Results:

• The proposed GrRFMN demonstrates superior accuracy in classifying granular data of varying granularity compared to the general fuzzy min-max neural network (GFMN) and classical methods.
• Experimental results on real datasets validate the effectiveness of the GrRFMN.
• Data granulation as a preprocessing step was observed to improve classifier performance.

Conclusions:

• The GrRFMN offers an effective solution for granular data classification and clustering.
• The network's design, incorporating a reflex mechanism and adaptive activation functions, successfully handles complex data characteristics.
• The study highlights the potential of granular computing approaches in advancing pattern recognition.