Multiobjective genetic algorithm partitioning for hierarchical learning of high-dimensional pattern spaces: a...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• High-dimensional data presents challenges for hierarchical learning.
• Existing methods often rely on competitive learning for input space partitioning.
• A need exists for pre-processing strategies that optimize data for subsequent classification tasks.

Purpose of the Study:

• To present a novel approach for partitioning pattern spaces using a multiobjective genetic algorithm.
• To identify near-optimal subspaces for hierarchical learning through a "learning-follows-decomposition" strategy.
• To optimize data partitions explicitly for mapping onto hierarchical classifiers.

Main Methods:

• Utilized a multiobjective genetic algorithm to partition pattern spaces.
• Implemented a "learning-follows-decomposition" approach, where input space is partitioned before hierarchical neural processing.
• Generated clusters based on fitness of purpose for subsequent mapping onto a hierarchical classifier.

Main Results:

• The proposed strategy effectively partitions pattern spaces.
• Preprocessing data and optimizing partitions reduced learning complexity and classification time.
• No degradation in overall classification error rate was observed.

Conclusions:

• The "learning-follows-decomposition" strategy is a generic solution for complex, high-dimensional problems.
• Optimized data partitioning enhances hierarchical learning efficiency.
• Neural modules demonstrate superiority in learning localized decision surfaces and offer better generalization within these partitions.