Seminar: Neuroevolution Trajectories and Landscapes

Presenter: Prof. Gabriela Ochoa, Division of Computing Science and Mathematics, University of Stirling, Scotland.

Abstract: Neuroevolution, the use of evolutionary algorithms to design neural networks, has a long tradition in evolutionary computation with roots in the late 1980s and early 1990s. Most neuroevolution systems optimise both the neural network topology and its weights. However, when scaling up to contemporary deep models with millions of weights for supervised learning tasks, gradient-based weight optimisation outperforms evolutionary methods. In consequence, many recent neuroevolution systems use gradient-based weight optimisation and only evolve the topology. This talk overviews our recent work on modelling neuroevolution systems with search trajectory networks (STNs) and local optima networks (LONs) with the aim of providing a visual and quantitative understanding of search and optimisation in this domain.