Evolving Modular Neural Networks: Investigating the Role of Modularity and Linkage Learning in Neuroevolution

The School of EECS is hosting the following PhD Thesis Review Seminar:

Evolving Modular Neural Networks: Investigating the Role of Modularity and Linkage Learning in Neuroevolution

Speaker: Yukai Qiao
Host: A/Prof Marcus Gallagher

Abstract: Neuroevolution has emerged as a powerful approach for automating the design of neural networks. However, the performance of neuroevolution heavily depends on the ability of the variation operators to effectively navigate the complex search space. Modularity, a property widely observed in natural systems, has been identified as a key factor in the success of evolutionary algorithms.

This thesis investigates the role of modularity in advancing neuroevolution algorithms and proposes novel techniques that exploit this property to evolve efficient neural networks. The first work explores the Modular Encoding for Neural Networks based on Attribute Grammars (MENNAG) algorithm, demonstrating its ability to generate modular structures and improve the efficiency of the evolutionary search process. Building upon these findings, the second work introduces a modularity-based linkage learning model inspired by the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA).

The proposed model captures the dependencies between neural network components and guides the variation operators to preserve and recombine modular substructures effectively. The third work presents a theoretical analysis of the runtime and optimization dynamics of GOMEA on the concatenated trap function, validating its effectiveness in exploiting problem structure and providing insights into the impact of modularity on the performance of evolutionary algorithms.

The contributions of this thesis lie in the theoretical understanding of modularity in neuroevolution and the development of practical algorithms that leverage this property to evolve efficient neural networks. The findings highlight the importance of modularity in designing effective neuroevolution techniques and have broader implications for the field of evolutionary computation.

Bio: I am currently a PhD candidate in the School of EECS supervised by A/Prof Marcus Gallagher, focusing on models for neuroevolution.