Also appears in the Online Symposium for Electronics Engineer 2000

http://www.techonline.com/osee/

Hidden Markov Models:

Fundamentals and Applications
Part 1: Markov Chains and Mixture Models

Valery A. Petrushin
[email protected]
Center for Strategic Technology Research
Accenture
3773 Willow Rd.
Northbrook, Illinois 60062, USA.

Abstract
The objective of this tutorial is to introduce basic concepts of a Hidden Markov Model
(HMM) as a fusion of more simple models such as a Markov chain and a Gaussian mixture
model. The tutorial is intended for the practicing engineer, biologist, linguist or programmer
who would like to learn more about the above mentioned fascinating mathematical models
and include them into one’s repertoire. This lecture presents Markov Chains and Gaussian
mixture models, which constitute the preliminary knowledge for understanding Hidden
Markov Models.

Introduction
Why it is so important to learn about these models? First, the models have proved to be
indispensable for a wide range of applications in such areas as signal processing,
bioinformatics, image processing, linguistics, and others, which deal with sequences or
mixtures of components. Second, the key algorithm used for estimating the models – the so-
called Expectation Maximization Algorithm -- has much broader application potential and
deserves to be known to every practicing engineer or scientist. And last, but not least, the
beauty of the models and algorithms makes it worthwhile to devote some time and efforts to
learning and enjoying them.

1 Introduction into Markov chains

1.1 Paving the path to the shrine problem

Welcome to the planet Median! Median is the capital
planet of the Stochastic Empire, which conquered all
worlds of the observable part of the Universe and spread its
influence to the hidden parts. Ruled by the Glorious and
Fair ruler, Emperor Probabil the Great, the Empire reached
the most flourishing state of commerce, science and
industry ever.
Welcome to Bayes-City, the capital of the Empire! Today,
the Emperor’s heralds had the honor of announcing a new
giant project. The four major Provinces of the Empire have
to pave the paths from their residence cities on Median to
the St. Likely Hood Shrine, located on the Mean Square in
Figure 1.1. Paving problem Bayes-City. Each path is 1,000 miles long
and has to be paved using large tiles of the
following precious stones: ruby (red), emerald (green) and topaz (blue) in accordance to its
Province’s random process which states that the color of next tile depends only on the color of
the current tile. The square around the St. Likely Hood Shrine will be secretly divided into
sectors and paved by the teams from all four Provinces over one night (Figure 1.1). A pilgrim
will be allowed to enter the Shrine to enjoy the view of the Bayesian Belief Network, the log
that served as a chair to Likely Hood, and the other relics, and also to talk to the high priests
of science during the high confidence time intervals if and only if he or she can determine for
three sectors of pavement around the Shrine which sector has been paved by the team of
which Province.
Imagine you are a pilgrim who journeyed more than 1,000 miles on foot to see the Stochastic
Empire’s relics. How can you solve the problem?
1.2 Markov chains
We deal with a stochastic or random process which is characterized by the rule that only the
current state of the process can influence the choice of the next state. It means the process has
no memory of its previous states. Such a process is called a Markov process after the name of
a prominent Russian mathematician Andrey Markov (1856-1922). If we assume that the
process has only a finite or countable set of states, then it is called a Markov chain. Markov
chains can be considered both in discrete and continuous time, but we shall limit our tutorial
to the discrete time finite Markov chains.

(a) (b)
π = (0.2 0.6 0.2 )
0.25 0.5 0.25
π = (0.5 0.5) A = 0.3 0.4 0.3 

 
0.5 0.5  0.3 0.5 0.2 
A= 
0.3 0.7
Figure 1.2. Markov chain models for a biased coin (a), and the paving model MN
Such chains can be described by diagrams (Figure 1.2). The nodes of the diagram represent
the states (in our case, a state corresponds to a choice of a tile of a particular color) and the
edges represent transitions between the states. A transition probability is assigned to each
edge. The probabilities of all edges outgoing from a node must sum to one. Beside that, there
is an initial state probability distribution to define the first state of the chain.

M = π,A π = (π 1 ,π 2 ,...,π N ) A = {aij }i , j =1, N (1.1)

P(qk +1 | q1 ,..., qk ) = P(qk +1 | qk ) (1.2)

Thus, a Markov chain is uniquely defined by a pair (1.1), where π is the vector of initial
probability distribution and A is a stochastic transition matrix. The Markov process
characteristic property is represented by (1.2). Figure 1.2 presents Markov chain models for a
biased coin and tile generation.
1.3 Training algorithm
We assume that each Province uses its own Markov chain to generate sequences of tiles. We
also assume that all tiles have their ordinal numbers carved on them and we have no problem
determining the order of tiles in the sequence (in spite of the fact that the tiles cover the two-
dimensional surface of the path). The resulting paths could be visually different for pavements
of different Provinces. For example, North Province, which is famous for its deposits of
emerald, might have larger probabilities for transitions to emerald (green) state and generate
greenish paths. But the differences among models could be rather small and the paths
produced by the models could look very similar, especially for the paths of short length. That
is why we have to find a formal approach for estimating the model's parameters based on
data. Thanks to the Emperor, we have enough data to train our models -- 1,000 miles of data
per model! If a tile has the diameter of one yard and the tiles are lined up in rows of 20 tiles
per row then we have 35,200 tiles per mile. To estimate parameters of a model with 3 states,
we need to calculate 12 values: a 3-element vector of initial state probabilities and a 3 by 3
state transition probability matrix. Let Q = q q ...q be a training sequence of states.
1 2 L
As estimations for the initial state probabilities, we use frequencies of tiles ci in the training
sequence (1.3). To estimate transition probabilities, we have to count the number of pairs of
tiles for each combination of colors cij (i,j=1,N) and then normalize values of each row (1.4).
ci
πi = i = 1, N (1.3)
L
cij N N
aij = N i, j = 1, N ∑ ∑ cij = L − 1 (1.4)
∑ cij i =1 j =1

j =1
Table 1.1 presents three estimates of parameters for the increasing length of the training
sequence.

Table 1.1. Markov chain training results

True L=1000 L=10000 L=35200
π = (0.2 0.6 0.2 ) π = (0.25 0.45 0.3) π = (0.28 0.46 0.26 ) π = (0.29 0.45 0.26 )

0.25 0.5 0.25 0.20 0.53 0.27 0.26 0.49 0.25 0.24 0.52 0.24
A =  0.3 0.4 0.3  A = 0.26 0.36 0.38 A =  0.3 0.4 0.3  A =  0.3 0.4 0.3 
       
 0.3 0.5 0.2  0.27 0.51 0.22  0.3 0.5 0.2   0.3 0.49 0.21

Now you know what to do! You should walk along each path, collect data and use (1.3) and
(1.4) to create Markov chains for each Province.
1.4 Recognition algorithm
Let us assume that you have already built the models for each Province: MN, ME, MS, and
MW. How can you use them to determine which Province’s team paved a sector Q? A natural
approach to solve this problem is to calculate the conditional probability of the model M for
the given sequence of tiles Q: P(M|Q).
Using the Bayes’ Rule (1.5), we can see that P(Q) does not depend on the model. Thus, we
should pick up the model, which maximizes the numerator of the formula (1.5).

P (Q | M k ) ⋅ P ( M k )
P( M k | Q) = (1.5)
P (Q )
where P(Mk|Q) is the posterior probability of model Mk for the given sequence Q; P(Q|Mk) is
the likelihood of Q to be generated by Mk ; P(Mk) is the prior probability of Mk ; P(Q) is the
probability of Q to be generated by all models.
This criterion is called maximum aposteriori probability (MAP) estimation and can be
formally expressed as (1.6). However, it is often the case that the prior probabilities are not
known or are uniform over all the models. Then we can reduce (1.6) to (1.7). This criterion is
called maximum likelihood (ML) estimation.
M k = arg max{P (Q | M k )}
* (1.6)
Mk

M k = arg max{P(Q | M k ) ⋅ P(M k )}

* (1.7)
Mk
How can you calculate the likelihood? Let us consider the sequence “323” and trace on the
diagram in Figure 2.1b to see how this sequence could be generated. The first state can be
generated with the probability π3 , the transition probability from 3 to 2 is a32, and from 2 to 3
is a23. The probability of the sequence is equal to the product of these three factors (1.8).
P(Q | M ) = π a a = 0.2 ⋅ 0.5 ⋅ 0.3 = 0.075 (1.8)
3 32 23
Generalizing our derivation for the sequence of arbitrary length we obtain (1.9). Typically the
logarithm of likelihood or log-likelihood is used for calculations (1.10).
L
P (Q | M ) = π q ⋅ aq q ⋅ aq q ⋅ ... ⋅ aq
1 1 2 2 3 L −1q L
= π q ⋅ ∏ aq
1 i −1qi (1.9)
i =2
L
log P (Q | M ) = log π q + ∑ log aq1 i −1qi
(1.10)
i =2
Table 1.2 presents the accuracy of recognition of 1,000 sequences of different length. The test
sequences were generated by the “true” models and recognized by the estimated models. For
recognition we used the ML criterion, i.e. we calculated the log-likelihood of the given
sequence for each of four models and picked up the model with the maximal value. The set of
models was trained on “one-mile” data for each model. You can see that the accuracy is poor
for the short sequences and gradually improves with growing length of sequences. For reliable
recognition we need sequences of more than 100 symbols.

Table 1.2. Recognition accuracy for test sequences of different length

Length 10 20 50 100 150 200 250 300 400 500
Accuracy (%) 55.7 64.7 82 92 96.4 98 99.4 99.7 99.8 100

True models:
North East South West
π = (0.2 0.6 0.2 ) π = (0.5 0.25 0.25) π = (0.2 0.2 0.6) π = (0.3 0.3 0 .4 )
0.25 0.5 0.25 0.2 0.3 0.5 0.3 0.3 0. 4 
0.4 0.3 0.3 
A =  0.3 0.4 0.3 

A = 0.3 0.3 0.4  A = 0.3 0.4 0.3
  A = 0.5 0.25 0.25    
 
 0.3 0.5 0.2  0.4 0.3 0.3  0.2 0.2 0.6 0.4 0.3 0.3

Estimated “one-mile” models:

North East South West
π = (0.28 0.45 0.27 ) π = (0.43 0.28 0.29) π = (0.23 0.24 0.53) π = (0.33 0.33 0.34 )
0.25 0.49 0.26  0.4 0.3 0.3  0.2 0.29 0.51  0.3 0.29 0.41
A =  0.3 0.4 0.3  A =  0.5 0.25 0.25 A = 0.3 0.3 0.4  A =  0.3 0.39 0.31
       
 0.3 0.5 0.2  0.41 0.29 0.3  0.2 0.2 0.6  0.4 0.3 0.3 

1.5 Generalizations
Two important generalizations of the Markov chain model described above are worth to
mentioning. They are high-order Markov chains and continuous-time Markov chains.
In the case of a high-order Markov chain of order n, where n > 1, we assume that the choice
of the next state depends on n previous states, including the current state (1.11).
P(qk + n | q1 , q2 ,..., qk ,..., qk + n −1 ) = P(qk + n | qk ,..., qk + n −1 ) (1.11)
In the case of a continuous-time Markov chain, the process waits for a random time at the
current state before it jumps to the next state. The waiting time at each state can have its own
probability density function or the common function for all states, for example, an exponential
distribution (1.12).

f T (t ) = λ ⋅ e − λt t≥0 (1.12)
1.6 References and Applications
The theory of Markov chains is well established and has vast literature. Taking all
responsibility for my slant choice, I would like to mention only two books. Norris’ book [1] is
a concise introduction into the theory of Markov chains, and Stewart’s book [2] pays attention
to numerical aspects of Markov chains and their applications. Both books have some
recommendations and references for further reading.

Markov Chains have a wide range of applications. One of the most important and elaborated
areas of applications is analysis and design of queues and queuing networks. It covers a range
of problems from productivity analysis of a carwash station to the design of optimal computer
and telecommunication networks. Stewart’s book [2] gives a lot of details on this type of
modeling and describes a software package that can be used for queuing networks design and
analysis (http://www.csc.ncsu.edu/faculty/WStewart/MARCA/marca.html).

Historically, the first application of Markov Chains was made by Andrey Markov himself in
the area of language modeling. Markov was fond of poetry and he applied his model to
studies of poetic style. Today you can have fun using the Shannonizer which is a Web-based
program (see http://www.nightgarden.com/shannon.htm) that can rewrite your message in the
style of Mark Twain or Lewis Carroll.

Another example of Markov chains application in linguistics is stochastic language modeling.

It was shown that alphabet level Markov chains of order 5 and 6 can be used for recognition
texts in different languages. The word level Markov chains of order 1 and 2 are used for
language modeling for speech recognition (http://www.is.cs.cmu.edu/ISL.speech.lm.html).

DNA sequences can be considered as texts in the alphabet of four letters that represent the
nucleotides. The difference in stochastic properties of Markov chain models for coding and
non-coding regions of DNA can be used for gene finding. GeneMark is the first system that
implemented this approach (see http://genemark.biology.gatech.edu/GeneMark/).

The other group of models known as branching processes was designed to cover such
application as modeling chain reaction in chemistry and nuclear physics, population genetics
(http://pespmc1.vub.ac.be/MATHMPG.html), and even game analysis for such games as
baseball (http://www.retrosheet.org/mdp/markov/theory.htm), curling, etc.

And the last application area, which I’d like to mention, is so called Markov chain Monte
Carlo algorithms. They use Markov chains to generate random numbers that belong exactly to
the desired distribution or, speaking in other words, they create a perfectly random sampling.
See the http://dimacs.rutgers.edu/~dbwilson/exact/ for additional information about these
algorithms.

2 Introduction into Gaussian mixture models

2.1 Revealing the old monk’s secret problem

All respectful families on Median have their traditional vast gardens of sculptures. The
gardens, which are covered by sand composed of microscopic grains of precious stones,
exhibit both the sculptures of concrete things and persons and abstract concepts. The color of
the sand is as important as the artistic and scientific value of sculptures. The sand recipes are
kept in secret and passed from one generation to another.
A new rich man (let us call him Mr. Newrich), who interpreted the slogan “Keep off
Median!” as permission to explore and exploit poor developed suburbs of the Empire,
returned to the capital planet and settled himself in a prestigious area. But when he started to
design his garden of sculptures, he ran into a problem of sand recipe. After several
unsuccessful attempts to buy a recipe he decided to steal some sand from the workshop of an
old monk who served as a sandkeeper for a noble and respectful, but not too rich family, and
reveal the secret of the recipe using the stolen sand as data. Mr. Newrich’s helpers measured
the wavelength of reflected light for 10,000 sand grains. Figure 2.1 presents the histogram for
the data. Three concentrations of data near 400 nm, 420 nm and 450 nm allow us to form a
hypothesis that sand is a mixture of three components.

Figure 2.1. Histogram of sand

data.

Let us pretend that you are the chief scientist at Allempire Consulting, Inc., and Mr. Newrich
pays you a fortune to decode the data. How can you determine the characteristics of each
component and proportion of the mixture? It seems that stochastic modeling approach called
Gaussian Mixture Models can do the job.

2.2 Gaussian mixture model

Let us assume that components of the mixture have Gaussian distribution with mean µ and
variance σ2 and probability density function (2.1). The choice of the Gaussian distribution is
natural and very common when we deal with a natural object or phenomenon, such as sand.
But the mixture model approach can be applied to the other distributions from the exponential
family.
2
−( x − µ )
1
f ( x; µ , σ ) = e 2σ
2
(2.1)
σ 2π
We let o1,o2,…,oL denote our observations or data points. Then each data point is assumed to
be an independent realization of the random variable O with the three-component mixture
probability density function (2.2) and log-likelihood function (2.3). With the maximum
likelihood approach to the estimation of our model M = ν , N (µ , σ ),..., N (µ , σ ) , we
1 1 K K

need to find such values of νi, µi and σi that maximize the function (2.3). To solve the
problem, the Expectation Maximization (EM) Algorithm is used.
K K
f (o; M ) = ∑ν i f i (o; µ i ,σ i ) ∑ν i = 1 (2.2)
i =1 i =1
L K
log L( M ) = ∑ log(∑ν i f (o j ; µ i ,σ i )) (2.3)
j =1 i =1

The EM Algorithm is also known as the “Fuzzy k-mean” algorithm. The key idea of the
algorithm is to assign to each data point oi a vector wi that has as many elements as there are
components in the mixture (three in our case). Each element of the vector wij (we shall call
them weights) represents the confidence or probability that the i-th data point oi belongs to the
j-th mixture component. All weights for a given data point must sum to one (2.4). For
example, a two-component mixture weight vector (1, 0) means that the data point certainly
belongs to the first mixture component, but (0.3, 0.7) means than the data point belongs more
to the second mixture component than to the first one.

K
wij = P ( N ( µ j ,σ j ) | oi ), ∑ wij = 1 (2.4)
j =1
For each iteration, we use the current parameters of the model to calculate weights for all data
point (expectation step or E-step), and then we use these updated weights to recalculate
parameters of the model (maximization step or M-step).

Let us see how the EM algorithm works in some details. First, we have to provide an initial
(0)
approximation of model M using one of the following ways:
(1) Partitioning data points arbitrarily among the mixture components and then calculating the
parameters of the model.
(2) Setting up the parameters randomly or based on our knowledge about the problem.
After this, we start to iterate E- and M-steps until a stopping criterion gets true.

For the E-step the expected component membership weights are calculated for each data point
based on parameters of the current model. This is done using the Bayes’ Rule that calculates
the weight wij as the posterior probability of membership of i-th data point in j-th mixture
component (2.5). Compare the formula (2.5) to the Bayes’ Rule (1.5).
K
wij = ν j ⋅ f (oi ; µ j ,σ j ) ∑ν h ⋅ f (oi ; µ h ,σ h ) (2.5)
h =1
For the M-step, the parameters of the model are recalculated using formulas (2.6), (2.7) and
(2.8), based on our refined knowledge about membership of data points.

L
ν (j t +1) = ∑ wij L j = 1, K (2.6)
i =1
L L
µ (j t +1) = ∑ wij ⋅ oi ∑ wij (2.7)
i =1 i =1
L L
σ 2j ( t +1) = ∑ wij ⋅ (oi − µ j ) 2 ∑ wij (2.8)
i =1 i =1
A nice feature of the EM algorithm is that the likelihood function (2.3) can never decrease;
hence it eventually converges to a local maximum. In practice, the algorithm stops when the
difference between two consecutive values of likelihood function is less than a threshold
(2.9).

log L( M ( t +1) ) − log L( M ( t ) ) < ε (2.9)

In order to better understand how the EM algorithm works, let us trace it for the case of two
mixture components with a small amount of data points. Suppose we have 10 data points that
are arbitrarily partitioned between two components (see Figure 2.3a). The initial model has
large variances and means that lie close to each other (see Table 2.1). Using this model, the
algorithm calculates a new model with larger distance between means and smaller variances.
After 22 iterations the algorithm produces the model that maximizes the likelihood function
and optimizes the membership probability distributions for each data point (see Figure 2.3 and
Table 2.1). The resulting solution is depicted as two green bell-shaped curves on Figure 2.3d.
You can see that we have perfect mixture proportion, and good approximation for the mean of
the first component and variance of the second component. In general, the results are
amazingly good for a data set of ten points.
Figure 2.3. EM algorithm convergence.

Table 2.1 EM algorithm convergence (10 points)

Applying the EM algorithm to Mr. Newrich’s data, we obtain the solution presented in Table
2.2. Due to a rather good initial model it took only 17 iterations to converge. The old monk’s
secret is revealed!

Table 2.2. Monk’s secret problem

The mixture model approach and the EM algorithm have been generalized to accommodate
multivariate data and other kind of probability distributions. Another method of generalization
is to split multivariate data into several sets or streams and create a mixture model for each
stream, but have the joint likelihood function.

The special beauty of mixture models is that they work in cases when the components are
heavily intersected. For example, for two-dimensional data of 1000 points and three-
component mixture with two components having equal means but different variances, we can
get a good approximation in 58 iterations (see Figure 2.4). The dark side of the EM algorithm
is that it is very slow and is considered to be impractical for large data sets.

Figure 2.4. Three-component mixture with two components having equal means.

The EM algorithm was invented in late 1970s, but the 1990s marked a new wave of research
revived by growing interest to data mining. Recent research is going in two directions:
(1) Generalization of the EM algorithm and extension to other probability distributions, and
(2) Improving the speed of EM algorithm using hierarchical data representations.

2.3 References and Applications

There is a sufficient body of literature devoted to the mixture models. Most of it is intended
for researches with a strong mathematical background. I refer to two books [3-4] written by
Australian researchers. There are also several commercial and free software packages devoted
to mixture models, most of which are available on the Web.
Commercial software:
• MPLUS http://www.statmodel.com/index.html
• MIX http://icarus.math.mcmaster.ca/peter/mix/mix.html
Freeware:
• EMMIX http://www.maths.uq.edu.au/~gjm/emmix/emmix.html
• MCLUST http://www.stat.washington.edu/fraley/mclust/
• MIXCLUS2 http://biometrics.ag.uq.edu.au/software.htm
• NORMIX http://alumni.caltech.edu/~wolfe/
• CAMAN http://ftp.ukbf.fu-berlin.de/sozmed/caman.html

Mixture Models have been in use for more than 30 years. It is a very powerful approach to
model-based clustering and classification. It has been applied to numerous problems. Some of
the more interesting applications are described below.
Mixture models are widely used in image processing for image representation and
segmentation, object recognition, and content-based image retrieval :
• http://www.cs.toronto.edu/vis/
• http://www.maths.uwa.edu.au/~rkealley/diss/diss.html
Mixture models also were used for text classification and clustering:
• http://www.cs.cmu.edu/~mccallum/
Gaussian mixture models proved to be very successful for solving the speaker recognition and
speaker verification problems:
• http://www.ll.mit.edu/IST/

There is growing interest in applying this technique to exploratory analysis of data or data
mining, and medical diagnosis:
• http://www.cs.helsinki.fi/research/cosco/
• http://www.cnl.salk.edu/cgi-bin/pub-search/
You can learn more about these applications by following the URLs presented above.

References
[1] J.R. Norris (1997) Markov Chains. Cambridge University Press.
[2] W.J. Stewart (1994) Introduction to the Numerical Solutions of Markov Chains.
Princeton University Press.
[3] G.J. McLachlan and K.E. Basford (1988) Mixture Models: Inference and Application to
Clustering. New York: Marcel Dekker.
[4] G.J. McLachlan and T. Krishnan (1997) The EM Algorithm and Extensions. New York:
Wiley.