Modular Grammatical Evolution for the Generation of Artificial Neural Networks

  • 2022-08-04 18:38:49
  • Khabat Soltanian, Ali Ebnenasir, Mohsen Afsharchi
  • 0

Abstract

This paper presents a novel method, called Modular Grammatical Evolution(MGE), towards validating the hypothesis that restricting the solution space ofNeuroEvolution to modular and simple neural networks enables the efficientgeneration of smaller and more structured neural networks while providingacceptable (and in some cases superior) accuracy on large data sets. MGE alsoenhances the state-of-the-art Grammatical Evolution (GE) methods in twodirections. First, MGE's representation is modular in that each individual hasa set of genes, and each gene is mapped to a neuron by grammatical rules.Second, the proposed representation mitigates two important drawbacks of GE,namely the low scalability and weak locality of representation, towardsgenerating modular and multi-layer networks with a high number of neurons. Wedefine and evaluate five different forms of structures with and withoutmodularity using MGE and find single-layer modules with no coupling moreproductive. Our experiments demonstrate that modularity helps in finding betterneural networks faster. We have validated the proposed method using tenwell-known classification benchmarks with different sizes, feature counts, andoutput class count. Our experimental results indicate that MGE providessuperior accuracy with respect to existing NeuroEvolution methods and returnsclassifiers that are significantly simpler than other machine learninggenerated classifiers. Finally, we empirically demonstrate that MGE outperformsother GE methods in terms of locality and scalability properties.

 

Quick Read (beta)

loading the full paper ...