Research on an integration methodology for combining many competing models with a dynamical mixing and/or switching

我们研究了经验贝叶斯和伪贝叶斯过程的集成方法，并集中在开发信息标准来选择最佳的元模型，每个元模型是通过组合许多竞争模型来增强预测能力的性能。我们作出了努力，以开发新的方法，在贝叶斯计算的顺序蒙特卡罗方法，吸引了许多利益的框架中的restaurant程序。通过研究人工智能社区中模型平均方法和提升之间的算法相似性，还研究了理论工作。我们考虑了一个应用程序中的混合和/或切换机制的许多竞争模型中起着重要的作用，在提高预测能力以及灵活性，同时保持鲁棒性的程序。至于具体的应用实例，我们已经进行了信号自动处理的研究项目， ...更多信息 通过人造卫星获得的数据、高频金融数据的分析、自动目标跟踪和识别视觉和传感器融合技术、销售点（POS）数据的分析以及GPS网络数据的时间相关反演。我们证实了所提出的算法的有效性，自动创建的模型，通过将这种方法应用于真实的世界的问题，并攻击新的挑战性问题，以响应社会需求。特别是，我们开发了DNA阵列数据的新分析技术，可以在基因组科学的一个实验中测量数千个基因的表达水平。在DNA阵列数据的分析中，我们面临着困难，其中大部分困难来自这样一个事实，即当案例的数量（样本大小）至多为100时，特征向量的维度（基因的数量）的范围从几千到几万。因此，由于过度学习，不可能应用传统的聚类方法。为了克服这种困难，我们开发了一种新的聚类方法，并提供了一个名为ArrayCluster的软件，该软件可以在http：//www.ism.ac.jp/jihiguchi/arraycluster.htm下载。为了促进建模科学的发展，我们组织了关于统计建模方法学进展的国际研讨会“建模科学：<th>信息标准30周年（AIC 2003）”。少

We studied an integration method of the empirical Bayesian and pseudo-Bayesian procedures and concentrated on developing information criteria to choose the best meta-model each of which is built up by combining many competing models to enhance a performance on the predictive ability. We made an effort to develop the new methods for resampling procedures in a framework of the sequential Monte Carlo methods that has drawn many interests in Bayesian computation. A theoretical work has been also studied by investigating the similarities in algorithms between the model averaging approach and boosting in Artificial Intelligence community. We considered an application of the procedures in which a mixing and/or switching mechanisms of many competing models plays an important role in boosting the predictive ability as well as the flexibility while keeping the robustness. As for the concrete examples of application, we have conducted the research projects on an automatic transaction of signals o … More btained by the artificial satellites, an analysis of high-frequent financial data, automatic target tracking and recognition for vision and sensor fusion technique, analysis of Point of Sales (POS) data, and time-dependent inversion from GPS network data. We confirmed an effectiveness of the proposed algorithms for an automatic creation of models by applying this methodology to real world problems, and attacked the new challenging problems in response to the social demand. In particular, we developed the new analysis technique for DNA array data that measures the expression level of thousands of genes in a single experiment in genome science. In the analysis of DNA array data, we are faced with difficulties most of which arise from the fact such that while the number of cases (sample size) is at most 100, a dimension of the feature vector (number of genes) ranges from several thousands up to a few ten thousands. It is therefore impossible to apply the conventional clustering methods, because of over-learning. To overcome such difficulty, we developed a novel clustering method, and provided a software called ArrayCluster which can be downloaded at http://www.ism.ac.jp/〜higuchi/arraycluster.htm. We organized an international symposium on advancement in the statistical modeling methodology, "Science of Modeling : The 30^<th> Anniversary of the Information Criterion (AIC2003)" for promoting the Science of Modeling. Less