A self-organized, distributed, and adaptive rule-based induction system

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Learning Classifier Systems (LCSs) are established rule-based inductive learning methods.
• Supervised Classifier System (UCS) is a specific LCS variant for classification tasks.
• Current LCS approaches can be computationally intensive for large datasets.

Purpose of the Study:

• To develop an adaptive framework integrating UCS with Self-Organized Maps (SOM) for enhanced classification.
• To evaluate the performance of this distributed approach against traditional single-system models.
• To assess the framework's adaptability and efficiency using diverse synthetic and real-world data.

Main Methods:

• An adaptive framework was created by layering UCS onto a SOM neural network.
• The SOM dynamically decomposes classification problems into subproblems, each managed by a dedicated UCS.
• The framework's robustness was tested by substituting UCS with feedforward Artificial Neural Networks (ANNs).

Main Results:

• The proposed distributed framework achieved classification accuracy comparable to or exceeding non-distributed methods.
• Execution speed was significantly improved due to problem decomposition and smaller UCS populations.
• The system demonstrated adaptive problem decomposition capabilities, maintaining or enhancing accuracy and speed.

Conclusions:

• The adaptive SOM-UCS framework offers an efficient and effective solution for complex classification tasks.
• Problem decomposition in a distributed environment leads to increased system throughput and faster processing.
• This approach provides a scalable and high-performing alternative for data mining and machine learning applications.