Modular Grammatical Evolution for the Generation of Artificial Neural Networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• NeuroEvolution methods often struggle with scalability and generating structured neural networks.
• Existing Grammatical Evolution (GE) techniques face limitations in representation locality and scalability.

Purpose of the Study:

• To introduce and validate Modular Grammatical Evolution (MGE) for efficient neural network generation.
• To test the hypothesis that restricting the solution space to modular and simple neural networks enhances efficiency and structure.
• To improve upon state-of-the-art Grammatical Evolution methods.

Main Methods:

• Developed MGE with a modular gene representation mapped to neurons via grammatical rules.
• Mitigated GE's drawbacks of low scalability and weak representation locality.
• Defined and evaluated five structural forms, including modular and non-modular networks.
• Tested MGE on ten diverse classification benchmarks.

Main Results:

• Single-layer modules with no coupling proved most productive.
• Modularity significantly accelerated the discovery of effective neural networks.
• MGE achieved superior accuracy compared to existing NeuroEvolution methods.
• Classifiers generated by MGE were substantially simpler than those from other machine learning approaches.
• MGE demonstrated enhanced locality and scalability over other GE methods.

Conclusions:

• Modularity is a key factor in efficiently generating smaller, structured, and accurate neural networks.
• MGE offers a significant advancement in NeuroEvolution and Grammatical Evolution, addressing critical limitations.
• The proposed method provides a more scalable and locally representative approach to evolving neural networks.