基于出错成本的隐马尔可夫模型主动学习-软件工程专业论文.docxVIP

Title： Title： A Cost-Based Active Learning Method for Hidden Markov Models Major： Soih,vare Engineering Narne： “Zhiwej Supervisor： Chen Juhua,Yin Jian ABSTRACT Since their inception in the late 1960’S，Hidden Markov models(H／VIJⅥS)have been quite widely applied in numerous real—world applications．However,when data collection is costly or time—consuming the need for training data becomes a key bottleneck． This dissertation presents a method to minimize the training data without losing the accuracy ofthe acquired information．Our approach combines the basic idea ofactive learning with the HMM theory．By automatically choose the most useful examples，it can significantly reduce the need for number of examples that the learning system is tralned on． In our approach，we first introduce a hidden-but-queryable observation in active learning HMM．It is treated as a special observation whose value is hidden but visible after a query．We then explore cost—based active learning method for two fundamental HMM problems：state estimation and model learning／parameter estimation，where methods are presented for actively choosing a time point to sample that is more efficient than random sampling． To assess the performance of the algorithm，we conducted two sets of simulation experiments and experimental results demonstrate that，in general，our approach is superior to other algorithms． Key words：Active Learning，Hidden Markov model，forward—backward procedure, Baum—Welch algorithm ．Ⅱ． 李志伟：基于出错成本的隐马尔可夫模型主动学习第1章引言 李志伟：基于出错成本的隐马尔可夫模型主动学习 第1章引言 1．1选题的意义 隐马尔可夫模型(Hidden Markov Models，HMMs)是一种在序列分析领域 应用广泛的统计模型，如语音识别，生物序列分析等。然而和许多机器学习过 程一样，收集数据往往费时费力甚至在某些情况下数据获取非常有限。因此通 过某些方式来降低对数据集大小的要求或者最大化对现有数据的利用，常常是 非常有益的。 在主动学习的框架下，我们允许算法更仔细地主动选择训练数据，而非均 匀取样(Uniform Sampling)或随机取样(Random Sampling)。我们不需要事先设定 选择训练数据的方式，而是算法根据历史数据主动地选择最有价值的数据进行 训练。本文提出的基于出错成本的隐马尔可夫模型主动学习算法就是通过降低 学习系统对训练数据集大小的依赖，以获得算法在效率以及／或者学习效果上的 改进。 1．2研究现状 起源于60年代末70年代初的隐马尔可夫模型‘29，”01021作为信号的一种统 计模型目前被广泛用于语音识别‘43,36,2q手写识别‘2，4t2棚图像信号处理125J以及生物 信息学