Sequential testing and change detection algorythms

顺序测试和变更检测算法

• 批准号: 36484-2007
• 负责人: Gombay, Edit
• 金额: $ 1.46万
• 依托单位: University of Alberta
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2007
• 资助国家: 加拿大
• 起止时间: 2007-01-01 至 2008-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=362389
• 关键词: Sequential; testing; change; detection; algorythms

I propose to do research in the areas of Sequential Statistical Algorithms and Change-Point Analysis.1. Sequential Testing is necessary when observations come in one by one over a period of time. Monitoring is ongoing to detect change in the data generating mechanism, such as change in climate, manufacturing quality, etc. In my earlier work new type of algorithms have been defined for the independent observations case. These are superior to existing methods as, although optimal for minimizing the reaction time after change had occurred, those have unacceptable type I error. This is the reason why many sequential testing methods remained mostly in the theoretical domain, and practitioners have not adapted them widely. The new algorithms have tight control over type I errors, easy to perform, and pay for this with only slightly increased delay time in raising alarm (declaring that the data generating mechanism has changed). Having solved this problem for independent observations other data structures have to be considered as the independence assumptions are often violated in practice. Work is ongoing and expanding for different time series observations that are arising in practice.  Data structures that arise in clinical trials are of special interest. The implementation of sequential monitoring for them is based on solving theoretically challenging problems.2. Change-Point Detection: In these algorithms we look at the data retrospectively to see if any disturbance has occurred. Such methods are essential, as if there was a structural change in the data generating procedure, and it was not detected, then all model fitting and forecast are erroneous. Potential and existing applications of these methods are in environmental data analysis, econometrics, biostatistics, etc.

我建议在序贯统计算法和变点分析领域进行研究。当观测在一段时间内一个接一个地进入时，顺序检验是必要的。监测正在进行中，以检测数据生成机制的变化，如气候变化，制造质量等，在我早期的工作中，新类型的算法已被定义为独立的观察情况。这些方法上级现有的方法，因为尽管对于最小化变化发生后的反应时间是最佳的，但这些方法具有不可接受的I型误差。这就是为什么许多序贯测试方法大多停留在理论领域，而实践者并没有广泛采用它们的原因。新算法对I类错误有严格的控制，易于执行，并且在发出警报（声明数据生成机制已更改）时仅略微增加延迟时间。在解决了独立观测的这个问题之后，必须考虑其他数据结构，因为在实践中经常违反独立性假设。工作正在进行中，并扩大不同的时间序列观察，在实践中出现的数据结构，在临床试验中出现的特别感兴趣。对它们实施顺序监测是基于解决理论上具有挑战性的问题.变点检测：在这些算法中，我们回顾性地查看数据，看看是否发生了任何干扰。这样的方法是必不可少的，因为如果在数据生成过程中有结构变化，而没有检测到，那么所有的模型拟合和预测都是错误的。这些方法在环境数据分析、计量经济学、生物统计学等方面有潜在的和现有的应用。