Machine learning algorithms for automated decision making under domain shift

领域转移下自动决策的机器学习算法

• 批准号: 2736505
• 金额: --
• 依托单位: University College London
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2736505
• 关键词: Machine; learning; algorithms; automated; decision

Machine Learning (ML) has made significant progress in recent years, powered by the development of new algorithms, availability of data, open-source software, and ever-increasing computational resources. However, most applications in which machine learning has been successful so far are limited to the cases where data is abundant and cheap to gather, such as images, advertising, website user interactions, and financial data. In contrast, in many real-world scientific applications, data is scarce and expensive to collect and label, which mandates the use of historic data. Unfortunately, such data is often not representative of novel data points on which we want to perform predictions, a problem more formally called domain shift. This is even exacerbated when models trained on historic data are used in conjunction with autonomous decision-making agents.In this project, you will work towards addressing these limitations of machine learning. The goal is to develop improved algorithms which are able to better extract relevant information from data instead of spurious correlations, are robust to domain shift and generalise better to novel situations, and can therefore be employed within decision making agents. To assess how the methods perform, you can address real-world examples from important chemistry problems, such as drug and materials discovery.

近年来，机器学习（ML）取得了重大进展，这得益于新算法的开发、数据的可用性、开源软件和不断增加的计算资源。然而，到目前为止，机器学习成功的大多数应用仅限于数据丰富且收集成本低的情况，例如图像，广告，网站用户交互和财务数据。相比之下，在许多现实世界的科学应用中，数据是稀缺的，收集和标记的成本很高，这就要求使用历史数据。不幸的是，这些数据通常不能代表我们想要进行预测的新数据点，这个问题更正式地称为域转移。当在历史数据上训练的模型与自主决策代理结合使用时，这一问题甚至会加剧。在本项目中，您将致力于解决机器学习的这些局限性。我们的目标是开发改进的算法，能够更好地从数据中提取相关信息，而不是虚假的相关性，是强大的域转移和更好地概括到新的情况，因此可以在决策代理。为了评估这些方法的性能，您可以从重要的化学问题中解决现实世界的例子，例如药物和材料发现。