一种GMMHMM隐状态与高斯混合成份初始化算法
2019-06-07张军超蒋强荣
张军超 蒋强荣



摘 要:为了解决传统隐马尔可夫模型应用通常将隐状态数和混合成份数看作一致的弊端,更客观地描述问题,使模型研究适合现实的数据分布,参数设定更为精准,从而使算法效果达到最优,提出一种基于高斯混合分布、聚类思想和OEHS准则的适应数据分布且自动确定参数的算法。因隐马尔可夫学习算法由EM算法实现,但EM是局部最优算法,严重依赖初始值,从跳出局部最优的角度出发,对两个参数进行初始设定。与传统的随机初始化方法进行比较,实验结果表明,该算法能得到更好的结果。
关键词:隐马尔可夫模型;GMM混合成份;隐状态;自适应
DOI:10. 11907/rjdk. 181494
中图分类号:TP312文献标识码:A文章编号:1672-7800(2019)001-0081-05
Abstract:In order to solve the problem that in traditional application of hidden Markov model, the number of hidden states and the number of mixed components are usually regarded as the same, and to describe the problem more objectively so that the model research can be very suitable for the actual data distribution, and the parameters are set more accurately to make the algorithm achieve the best results, an algorithm based on Gaussian mixture distribution, clustering idea and OEHS criterion is proposed. At the same time, the hidden Markov learning algorithm is implemented by EM algorithm, but EM is a local optimal algorithm, which depends heavily on the initial value. From the point of jumping out of local optimum, so that the initial setting of the two parameters is conducted, which can adapt to the data distribution and automatically determine the parameters. Compared with the traditional random initialization method, the experimental results show that the proposed algorithm can get better results.
0 引言
20世紀60年代,鲍姆提出了隐马尔科夫模型(Hidden Markov Model,HMM),在语音识别领域使用,被广大科研人员熟知。文艺复兴科技公司是国际著名投资机构,因从1989年开始保持超高年回报率被业界誉为最高效的对冲基金,通过独特的数学模型,捕捉市场机会进行量化投资,而隐马尔科夫模型是主要工具之一[1]。可以看到,HMM在各个领域都有应用,具有广泛影响。常用的HMM模型可以根据观测状态分为两个类别:连续型和离散型。例如,连续型的GaussianHMM、GMMHMM,离散观测状态的有MultinomialHMM。
离散型模型使用较为基础,有几个重要的参数需要设置,例如:“startprob_”表示隐状态的初始分布概率,“transmat_”表示不同状态间的转移概率,“emissionprob_”表示观测值的发射概率矩阵。许博等[2]为了实时、准确地识别多种P2P应用流,提出采用离散型HMM的P2P 流识别技术,能减少模型建立时间,提高识别未知流的实时性和准确性,并能较好地适应网络环境变化。……
