An adaptive multiobjective approach to evolving ART architectures.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Adaptive Resonance Theory (ART) neural networks are widely used for classification tasks.
• Existing ART architectures often require manual parameter tuning, limiting their efficiency and effectiveness.
• The need for automated and improved ART network design is evident.

Purpose of the Study:

• To evolve Adaptive Resonance Theory (ART) neural network architectures using a multiobjective optimization approach.
• To simultaneously optimize weights and topology for Fuzzy ARTMAP (FAM), Ellipsoidal ARTMAP (EAM), and Gaussian ARTMAP (GAM).
• To introduce the novel MO-GART (Multiobjective-ART) framework for enhanced classification.

Main Methods:

• A multiobjective evolutionary algorithm was employed to optimize three ART architectures: FAM, EAM, and GAM.
• The optimization process simultaneously evolved network weights and topology.
• The resulting architectures were collectively named MO-GART (MO-GFAM, MO-GEAM, MO-GGAM).

Main Results:

• MO-GART produces multiple classification solutions with varying trade-offs between accuracy (generalization) and size (number of categories).
• MO-GART demonstrates superior elegance (no user intervention for parameters), effectiveness (higher accuracy, smaller size), and efficiency (faster convergence).
• MO-GART shows competitive performance against established classifiers like Classification and Regression Trees (CART) and Support Vector Machines (SVMs).

Conclusions:

• MO-GART offers an automated and effective approach to designing ART neural networks for classification.
• The multiobjective optimization strategy successfully balances classification accuracy and network complexity.
• MO-GART represents a significant advancement in ART-based classification, outperforming prior ART models and rivaling other popular methods.