Adaptive Methods for Nonparametric Classification and Regression/Supervised Learning, Inference in HMM and State Space Models and Inference in Semiparametric Models

非参数分类和回归/监督学习的自适应方法、HMM 和状态空间模型中的推理以及半参数模型中的推理

• 批准号: 0104075
• 负责人: Peter Bickel
• 金额: $ 63万
• 依托单位: University of California-Berkeley
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2001
• 资助国家: 美国
• 起止时间: 2001-08-01 至 2007-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0104075&HistoricalAwards=false
• 关键词: Adaptive; Methods; Nonparametric; Classification; Regression

In non and semiparametric inference Bickel, in collaboration with Ritov and others, proposes to study how an increasing but proportionally vanishingly small cross validation sample can be used systematically to optimize supervised learning (classification and regression) procedures, for example ADA BOOST. Further they propose to study a unified theory for testing of semiparametric hypotheses and develop efficient tests for bioequivalence. In dependent data models, they propose to extend previous results on Hidden Markov models to state space models and study how procedures obtained by fitting and use of approximate likelihoods such as particle filters behave.The investigator and collaborators propose to analyze and develop new effective methods for identifying (by machine) the type of a newly perceived object or predicting some feature from historical information. This ranges from machine reading of hand written zip codes to predicting travel times of cars from one destination to another to predicting tumor type from microarray data. In a similar direction they propose to see how well computer simulation based approximations to ideal prediction methods work in very complicated models applying to situations such as voice recognition. Further they propose to study methods of inference bearing on questions such as whether a new drug which may be more expensive and have side-effects is sufficiently better than drugs currently in use to be authorized for distribution.

在非参数和半参数推理中，Bickel与Ritov等人合作，提出研究如何系统地使用不断增加但按比例消失的小交叉验证样本来优化监督学习（分类和回归）程序，例如ADA BOOST。 此外，他们建议研究半参数假设检验的统一理论，并开发有效的生物等效性检验。 在依赖数据模型中，他们建议将隐马尔可夫模型的先前结果扩展到状态空间模型，并研究通过拟合和使用近似似然性（如粒子滤波器）获得的程序如何表现。研究者和合作者建议分析和开发新的有效方法，用于识别（通过机器）新感知对象的类型或从历史信息中预测某些特征。 这包括从机器阅读手写的邮政编码到预测汽车从一个目的地到另一个目的地的旅行时间，再到从微阵列数据预测肿瘤类型。 在类似的方向上，他们建议看看基于计算机模拟的理想预测方法在应用于语音识别等情况的非常复杂的模型中的效果如何。 此外，他们还建议研究与诸如可能更昂贵且具有副作用的新药是否比目前使用的药物好得多等问题有关的推理方法，以便批准销售。