Rival penalized competitive learning for clustering analysis, RBF net, and curve detection

Area of Science:

• Machine Learning
• Artificial Intelligence
• Pattern Recognition

Background:

• Frequency Sensitive Competitive Learning (FSCL) performance degrades with incorrect unit selection.
• Competitive Learning (CL) algorithms require careful parameter tuning for optimal results.

Purpose of the Study:

• Introduce Rival Penalized Competitive Learning (RPCL) to address FSCL limitations.
• Develop an algorithm that automatically allocates an appropriate number of units for input data.
• Improve performance in unsupervised classification, RBF network training, and image curve detection.

Main Methods:

• Proposed RPCL algorithm modifies the winner unit and delearns the rival unit with a smaller learning rate.
• RPCL is an unsupervised extension of Kohonen's supervised LVQ2 algorithm.
• Evaluated RPCL against FSCL in unsupervised classification, RBF network training, and curve detection tasks.

Main Results:

• RPCL demonstrates superior performance compared to FSCL across all tested applications.
• The algorithm effectively handles the automatic allocation of units for diverse datasets.
• RPCL shows significant improvements in accuracy for classification and curve detection tasks.

Conclusions:

• RPCL offers a robust alternative to FSCL, particularly when unit number selection is critical.
• The proposed method enhances the adaptability and efficiency of competitive learning algorithms.
• RPCL provides a more effective approach for unsupervised learning tasks and neural network training.