Scholarlink

<p>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. We have proposed a method to overcome this limitation by using a <i>surrogate</i> model which estimates the behavioural distance between two neural network architectures, required to calculate novelty scores. This has demonstrated a speedup of 10 times over previous work, significantly improving on previous reported results on three benchmark datasets from Computer Vision—CIFAR-10, CIFAR-100, and SVHN. This method makes an explicit search for diversity considerably more tractable <i>for the same bounded resources</i>. Here we investigate a range of search methods that span the full spectrum of favouring accuracy, diversity, or different combinations of both. Surprisingly, we show that multiple unique combinations between a diversity metric and accuracy give rise to similar results. This enables us to posit the existence of a diversity-accuracy duality in ensembles of classifiers, which suggests that there might not be a need to find a trade-off between the two.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving Novelty Search with a Surrogate Model and Accuracy Objectives to Build High-Performing Ensembles of Classifiers

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月11日

查看全文
  • Rui P. Cardoso,
  • Emma Hart,
  • Jeremy V. Pitt

摘要

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. We have proposed a method to overcome this limitation by using a surrogate model which estimates the behavioural distance between two neural network architectures, required to calculate novelty scores. This has demonstrated a speedup of 10 times over previous work, significantly improving on previous reported results on three benchmark datasets from Computer Vision—CIFAR-10, CIFAR-100, and SVHN. This method makes an explicit search for diversity considerably more tractable for the same bounded resources. Here we investigate a range of search methods that span the full spectrum of favouring accuracy, diversity, or different combinations of both. Surprisingly, we show that multiple unique combinations between a diversity metric and accuracy give rise to similar results. This enables us to posit the existence of a diversity-accuracy duality in ensembles of classifiers, which suggests that there might not be a need to find a trade-off between the two.