Neuroevolution Trajectory Networks: illuminating the evolution of artificial neural networks

Citation

Sarti S (2024) Neuroevolution Trajectory Networks: illuminating the evolution of artificial neural networks. Doctor of Philosophy. University of Stirling.

Abstract
Neuroevolution is the discipline whereby ANNs are automatically generated using EC. This field began with the evolution of dense (shallow) neural networks for reinforcement learning task; neurocontrollers capable of evolving specific behaviours as required. Since then, neuroevolution has been used to discover architectures and hyperparameters of Deep Neural Networks, in ways never before conceived by human experts, with many achieving state-of-the-art results. Similar to other types of EAs, there is a wide variety of neuroevolution algorithms constantly being introduced. However, there is a lack of effective tools to examine these systems and assess whether they share underlying principles. This thesis proposes Neuroevolution Trajectory Networks (NTNs), an advanced visualisation tool that leverages complex networks to explore the intrinsic mechanisms inherent in the evolution of neural networks. In this research the tool was developed as a specialised version of Search Trajectory Networks, and it was particularly instantiated to illuminate the behaviour of algorithms navigating neuroevolution search spaces. Throughout the progress, this technique has been progressively applied from systems of shallow network evolution, to deep neural networks. The examination has focused on explicit characteristics of neuroevolution system. Specifically, the learnings achieved highlighted the importance of understanding the role of recombination in neuroevolution, revealing critical inefficiencies that hinder overall algorithm performance. A relation between neurocontrollers' diversity and exploration exists, as topological structures can influence the behavioural characterisations and the diversity generation of different search strategies. Furthermore, our analytical tool has offered insights into the favoured dynamics of transfer learning paradigm in the deep neuroevolution of Convolutional Neural Networks; shedding light on promising avenues for further research and development. All of the above have offered substantial evidence that this advanced tool can be regarded as a specialised observational technique to better understand the inner mechanics of neuroevolution and its specific components, beyond the assessment of accuracy and performance alone. This is done so that collective efforts can be concentrated on aspects that can further enhance the evolution of neural networks. Illuminating their search spaces can be seen as a first step to analysing neural network compositions.

Keywords
Neuroevolution, NTNs, STNs, NEAT, Fast-DENSER, QD, Explainability, Evolutionary Computing, Complex Networks