Flexible and Adaptive Statistical Modeling

[9971405]在20世纪80年代，试图模拟大脑学习的研究人员为多层非线性神经网络开发了几种新的学习算法。有些偶然的是，这些算法被证明是非常强大的自适应回归和分类技术，并且已经证明它们的有用性与它们是否是一个好的大脑模型无关。它们现在被应用于医学诊断、化学过程控制、形状识别和其他广泛的重要实际问题。与此同时，统计领域的适应性技术也取得了重大进展。新的统计方法比传统的线性回归和线性判别分析等方法更强大。最近开发的一些程序包括CART（分类和回归树），广义加性模型，MARS（多元加性回归样条），以及最近邻算法的复杂版本，该算法可以为输入空间学习适当的度量。尽管它们来自不同的领域，使用不同的术语，但这些方法有很多共同之处。该提案中的一个项目是撰写一本研究专著，试图将这些观点集中在一起，以统一的方式解释它们。另外两个项目探讨了最近改进分类器和自适应模型选择的一些新方法。这项工作旨在开发基于历史数据进行预测的新模型和方法。这些技术在许多不同的领域都很重要，包括医疗诊断、财务预测和工业过程控制。生物技术是这些方法的一个特别重要的应用领域。科学家们现在拥有同时测量数千个基因的基因表达水平的技术，这使得确定哪些人类基因与癌症和心脏病等疾病有关成为令人兴奋的可能性。对大量信息进行分类就像在大海捞针一样，而像这里研究的预测方法将为这种搜索提供重要的工具。

9971405During the 1980's, researchers who were trying to model learning in the brain developed several novel learning algorithms for multilayer non-linear neural networks. Somewhat incidentally, these algorithms turn out to be very powerful techniques for adaptive regression and classification, and have proven their usefulness independent of whether or not they are a good model for the brain. They are now being applied to medical diagnosis, chemical process control, shape recognition and a wide range of other important practical problems. At the same time, there have been significant advances in adaptive techniques in the field of statistics. The new statistical methods are more powerful than classical techniques such as linear regression and linear discriminant analysis. Some of the recently developed procedures include CART (Classification And Regression Trees), generalized additive models, MARS (Multivariate Additive Regression Splines), and sophisticated versions of nearest neighbor algorithms that learn an appropriate metric for the input space. Although they come from different fields using different terminologies, these methods have much in common. One item in this proposal is a research monograph that seeks to bring many of these ideas under one umbrella, explaining them in a unified fashion. Two other items explore some recent new methods for improving classifiers and for adaptive model selection.This work aims at developing new models and methods for making predictions based on historical data. These techniques are important in many different fields including medical diagnosis, financial forecasting and industrial process control. The area of biotechnology is an especially important application for these methods. Scientists now have techniques for measuring gene expression levels for thousands of genes at the same time, allowing the exciting possibility of determining which human genes are involved in a diseases such as cancer and heart disease. Sorting through the mass of information is like trying to find a needle in a haystack, and predictive methods like the ones studied here will provide an important tool in this search.