Machine learning for non-myopic decision support and knowledge discovery

用于非短视决策支持和知识发现的机器学习

• 批准号: 418645-2012
• 负责人: Lizotte, Daniel
• 金额: $ 1.6万
• 依托单位: University of Western Ontario
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2015
• 资助国家: 加拿大
• 起止时间: 2015-01-01 至 2016-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=573763
• 关键词: Machine; learning; non; myopic; decision

The goal of our proposed research is to develop methods for using data to help humans make good long-term decisions. As a motivating example, consider choosing a medical treatment for a patient with a chronic disorder: in practice, treatment decisions are made based on knowledge about the state of the patient, knowledge about the treatment options currently available, and knowledge about how future decisions will be influenced by the patient's progress over time. Another example is the control of a water reservoir: decisions about how much water to use to generate power or irrigate crops are made based on knowledge about current demand, the state of the reservoir, and about the potential useage decisions that might be made in the future. We say a decision is "non-myopic" if it is made based on knowledge of the potential for future decision-making. In both of these examples, ultimately the final decision rests in human hands; however, there is enormous potential for that decision to be guided by sources of sequential data that are becoming more and more ubiquitous. For example, we are now seeing the development of databases that record how thousands or even millions of patients respond to different sequences of treatments over time, and these have the potential to inform non-myopic medical decision making more effectively than previous studies. However, rigorous analysis methods for constructing non-myopic decision aids from data are still in their infancy. Analysis methods in reinforcement learning and machine learning have enormous potential, but in many ways are not suited to decision aid construction: current methods do not effectively account for user preferences, they do not provide appropriate measures of confidence in their recommendations, and they do not account for the cost of gathering new data. My research aims to develop new methods without these shortcomings that can be applied to produce useful decision aids from sequential data. In the long term, as the depth and breadth of sequential medical data increases, the methods will improve the delivery of health care in Canada by providing our medical doctors with new, high-quality evidence to aid them in choosing the best treatments for their patients.

我们提出的研究的目标是开发使用数据来帮助人类做出良好的长期决策的方法。作为一个激励人心的例子，考虑为慢性病患者选择一种医疗治疗：在实践中，治疗决定是基于对患者状态的了解、对现有治疗方案的了解，以及关于患者随着时间的推移将如何影响未来决定的知识。另一个例子是水库的控制：关于用多少水发电或灌溉农作物的决定是基于对当前需求、水库状态和未来可能做出的潜在使用决定的了解。如果一个决定是基于对未来决策潜力的了解而做出的，我们就说它是“非短视的”。在这两个例子中，最终的决定掌握在人的手中；然而，该决定有巨大的潜力由变得越来越普遍的顺序数据来源来指导。例如，我们现在正在看到数据库的发展，这些数据库记录了数千甚至数百万患者随着时间的推移对不同治疗序列的反应，这些数据库有可能比以前的研究更有效地为非近视眼的医疗决策提供信息。然而，从数据中构建非短视决策辅助工具的严格分析方法仍处于初级阶段。强化学习和机器学习中的分析方法具有巨大的潜力，但在许多方面不适合决策辅助构建：目前的方法没有有效地考虑到用户的偏好，它们没有对他们的建议提供适当的置信度衡量，也没有考虑到收集新数据的成本。我的研究旨在开发没有这些缺点的新方法，这些方法可以应用于从顺序数据中产生有用的决策辅助工具。从长远来看，随着连续医疗数据的深度和广度的增加，这些方法将通过为我们的医生提供新的高质量证据来帮助他们为他们的患者选择最佳治疗方法，从而改善加拿大的医疗保健提供。