Prediction In Hidden Markov Models Using Sequential Monte Carlo Methods
Dongqing Zhang,Xuanxi Ning,Xueni Liu,Hongwei Ma
DOI: https://doi.org/10.1109/GSIS.2007.4443368
2007-01-01
Abstract:A novel method of multistep-ahead prediction based on joint probability distribution is proposed in this paper. Firstly, we introduce the basic theory of hidden Markov Model (HMM) and sequential Monte Carlo (SMC) method. Secondly, we make the joint multistep-ahead prediction using SMC method in HMM and then develop the corresponding on-line algorithm. At last, the data of monthly national air passengers in America, from Jan. 1990 to Aug. 2001, are analyzed, and experimental results demonstrate that the method proposed in this paper is effective.review of basic theory of HMM and SMC method respectively. In Section 4, approximate the probability density function interested using SMC rnethod in HMM and the corresponding on-line algorithm is developed. In section 5, the data of monthly national air passengers in America from Jan. 1990 to Aug. 2001, are analyzed and empirical results are presented. Finally, a surnmary is given in section 6. 11. HIDDEN MARKOVMODEI_ M ANY problems in time series analysis can be stated in hidden Markov models. To fix notation, consider the state-space model of the HMM[I 1: X, - g(x, X'_1) X') where x, I are unobserved states of the system and _-, I are observations, g(. I -) and f(. I -) are state evolution and observation densities. Cheng et al (2003) applied SMC methodology to address the problems of optimal filtering and smoothing in hidden Markov models, as in [2]. In this article we develop SMC method for prediction in HMM. In fact, a primary concern in HMM is one step-ahead prediction of state by estimating the probability distribution p(x, I zl:, _ 1) [I]-[3]and 11711.11tistep-allead prediction of state by estimating the distribution P(Xl,h 111, where 7,., = (71,. However, the predictive trajectory of P(X,,h I) may be incredible for nealecting the consistency principle sometimes and for many cases where predictive observations are required. Therefore, a novel method of multisep-ahead prediction of observations based on joint probability distributions, p(z, - - - 71+111 71:,_1), wh ich provides the most probable estimate for the predicted trajectory, is proposed. In this paper, a novel rnethod of multistep-ahead prediction based on joint probability distribution is proposed and we will focus on the joint multistep-ahead prediction using SMC rnethod concerning discrete-time HMM. The paper is organized as follows. In Section 2 and 3, we present a brief Dongqing ZFIANG is with College of Economics & Management, NanJing University of Aeronautics and Astronautics, Nanjing, CO 210016 C14INA (corresponding author phone:138138526241 e-mail: hollycase@1 sina.com). Xuanxi NING is with College of Economics & Management, Nanjing University of Aeronautics and Astronautics, Nanjing, CO 210016 CHINA. Xueni LIU is with College of Economics & Management, Nanjing University of Aeronautics and Astronautics, Nar1jing, CO 210016 CHINA. Hongwei Ma is with College of Engineering, Nanling Agricultural University, Nan.jing, CO 210031 CHINA. Hidden Markov models were originallv introduced as far back as 1957 and they have become increasingly popular in the last several years due to strong inathernatical structure and theoretical basis. A hidden Markov model is a doubly embedded stochastic process with an underlying stochastic process (the state sequence) that is nor observable or hidden. The state sequence is a Markov process and it is called hidden because it can only be observed through another set of stochastic processes that produce the sequence of observations, where each observation is a probabilistic function of the corresponding state[4]. A standard HMM is characterized as follows: X, - g(x, I x,-,, A) 7, - f (z, I x, B) where x, c S is hidden state at time t S=1Sl1S__1'-'1SA11'S thehidden state set with N elements: x, - g(x, J x,-,, A) is state evolution density and depends on state transition matrix A = la il. and = si I X, - -.f (z, I x, B) is measurement a pily t -1 = Si; 7, density at timet, B =fb,(v),l :5 i! NJ is the conditional probability distribution of observation variable with definition of bJv)=p z, Ix, =S,J. In addition.there is an initial state istribution z = Izil,;T, = P(x, = Si)J ! i: N in a standard HMM. It can be seen from the above discussion th at complete specification of a HMM requires 0 = tS, A., B, al. For convenience, we use the compact notation 0 = JA, B, irl. In additiom, there are three basic problems that must be solved to be useful in real-world applications for a HMM. The details of three problems and corresponding algorithms are given in the tutorial by Rabiner[4]. Ill. SEQUENTIAL MONTE CARLO METHODS SMC method was introduced in the 1960's and it is ail emerging and powerful methodology for coping with difficult 1-4244-1294-3/07/$25.00 (02007 IEEE 718