Augmenting Novelty Search with a Surrogate Model to Engineer Meta-Diversity in Ensembles of Classifiers

Cardoso, Rui P; Hart, Emma; Burth Kurka, David; Pitt, Jeremy

Abstract

Using Neuroevolution combined with Novelty Search to promote behavioural diversity is capable of constructing high-performing ensembles for classification. However, using gradient descent to train evolved architectures during the search can be computationally prohibitive. Here we propose a method to overcome this limitation by using a surrogate model which estimates the behavioural distance between two neural network architectures required to calculate the sparseness term in Novelty Search. We demonstrate a speedup of 10 times over previous work and significantly improve on previous reported results on three benchmark datasets from Computer Vision-CIFAR-10, CIFAR-100, and SVHN. This results from the expanded architecture search space facilitated by using a surrogate. Our method represents an improved paradigm for implementing horizontal scaling of learning algorithms by making an explicit search for diversity considerably more tractable for the same bounded resources.

Citation

Cardoso, R. P., Hart, E., Burth Kurka, D., & Pitt, J. (2022, April). Augmenting Novelty Search with a Surrogate Model to Engineer Meta-Diversity in Ensembles of Classifiers. Presented at EvoSTAR, Madrid

• Presentation Conference Type: Conference Paper (Published)
• Conference Name: EvoSTAR
• Start Date: Apr 20, 2022
• End Date: Apr 22, 2022
• Acceptance Date: Jan 26, 2022
• Online Publication Date: Apr 15, 2022
• Publication Date: 2022
• Deposit Date: Feb 7, 2022
• Publicly Available Date: Apr 16, 2023
• Publisher: Springer
• Pages: 418-434
• Series Title: Lecture Notes in Computer Science
• Series Number: 13224
• Book Title: Applications of Evolutionary Computation: EvoApplications 2022
• ISBN: 978-3-031-02461-0
• DOI: https://doi.org/10.1007/978-3-031-02462-7_27
• Keywords: diversity, ensemble, novelty search, surrogate
• Public URL: http://researchrepository.napier.ac.uk/Output/2842322