A TUTORIAL ON HIDDEN MARKOV MODELS

Rakesh Dugad U. B. Desai

Signal Processing and Articial Neural Networks Laboratory
Department of Electrical Engineering
Indian Institute of Technology | Bombay
Powai, Mumbai 400 076, India
[email protected], [email protected]

Technical Report No. : SPANN-96.1

May 1996

Abstract
In this tutorial we present an overview of (i) what are HMMs, (ii) what are the dierent
problems associated with HMMs, (iii) the Viterbi algorithm for determining the optimal state
sequence, (iv) algorithms associated with training HMMs, and (v) distance between HMMs.

1 Introduction
[1]
Suppose a person has say three coins and is sitting inside a room tossing them in some sequence{
this room is closed and what you are shown (on a display outside the room) is only the outcomes of
his tossing TTHTHHTT.. . this will be called the observation sequence . You do not know the
sequence in which he is tossing the dierent coins, nor do you know the bias of the various coins.
To appreciate how much the outcome depends on the individual biasing and the order of tossing
the coins, suppose you are given that the third coin is highly biased to produce heads and all coins
are tossed with equal probability. Then, we naturally expect there to be far greater number of
heads than tails in the output sequence. Now if it be given that besides the bias the probability of
going to the third coin (state) from either the rst or the second coin (state) is zero; then assuming
that we were in the rst or second state to begin with the heads and tails will appear with almost
equal probability in spite of the bias. So we see that the output sequence depends very much on
the individual bias, the transition probabilities between various states, as well as on which state
is chosen to begin the observations. The three sets, namely, the set of individual bias of the three
coins, the set of transition probabilities from one coin to the next and the set of initial probabilities
of choosing the states characterize what is called as the HIDDEN MARKOV MODEL(HMM)
for this coin tossing experiment. We shall shortly see the problems and their solutions that come
under the framework of HMMs.
 Rakesh Dugad is currently with Beckman Institute, ECE Department, University of Illinois, 405 N. Mathews
Avenue, Urbana IL 61801, USA

1
2 Notations
In the above experiment the outcomes, of tossing of each coin, T or H are called the observation
symbols. Because we considered coins, only two observation symbols were possible. To be more
general, consider a set of N urns each consisting of a number of marbles. The marbles are of M
distinct colors. Within each urn the marbles are of distinct colors. Our experiment consists of
drawing marbles from these urns in some sequence; only the sequence of marbles drawn is shown
to us. Marbles, here, correspond to the T or H and urns to coins in the above experiment. We now
dene the following notation for our model :
N = number of states (urns) in the model
M = total number of distinct observation symbols (marbles of M distinct colors)
T = length of observation sequence i.e. the number of symbols observed
1,2,.. . ,N will denote the N urns respectively
it denotes the state in which we are at time t
V = fv1,. . . ,vM g the discrete set of possible observation symbols
 = fig; i = P(i1 = i), the probability of being in state i at the beginning of the experiment i.e.
at t=1
A = faij g where aij = P(it+1=j j it= i), the probability of being in state j at time t+1 given that
we were in state i at time t. We assume that aij 's are independent of time.
B = fbj (k)g,bj (k) = P(vk at t j it = j), the probability of observing the symbol vk given that we
are in state j.
Ot will denote the observation symbol observed at instant t
 = (A,B; ) will be used as a compact notation to denote an HMM
Using the model, an observation sequence O = O1 ,O2 ,.. . ,OT is generated as follows : We start
our experiment by choosing one of the urns (according to the initial probability distribution  ),
then we choose a marble (observation symbol) from this urn { this beginning instant of time is
taken as t=1 and the state and observation symbol chosen at this t=1 are denoted by i1 and O1
respectively. After this we choose an urn (may be same or dierent from the urn at t=1) according
to the transition probability distribution A and again select a marble (denoted by O2 ) from this urn
depending on the observation symbol probability distribution bj (k ) for that urn (state). Continuing
this up to time t=T, generates the observation sequence O = O1 ,O2,. . . ,OT .

3 The Three Problems for HMMs

Most applications of HMMs are nally reduced to solving three main problems. These are :
Problem 1 : Given the model  = (A,B; ) how do we compute P(Oj ) , the probability of
occurrence of the observation sequence O = O1,O2 ,. . . ,OT .
Problem 2 : Given the model  = (A,B; ) how do we choose a state sequence I = i1,i2,. . . ,iT so
that P(O,I j ) ,the joint probability of the observation sequence O = O1,O2 ,. . . ,OT and the
state sequence given the model is maximized.

2
Problem 3 : How do we adjust the HMM model parameters  = (A,B; ) so that P(Oj ) (or
P(O,I j ) ) is maximized.
Problems 1 and 2 can be viewed as analysis problems while Problem 3 is a typical synthesis (or
model identication or training) problem.
Solutions to the Three Problems
3.1 Problem 1
A most straightforward way to determine P(O j ) is to nd P(O j I; ) for a xed state sequence
I = i1,i2,. . . ,iT then multiply it by P(I j ) and then sum up over all possible I's. We have
P(O j I; ) = bi1 (O1 )bi2 (O2 )


biT (OT ): (1)
P(I j ) = i1 ai1 i2 ai2i3


aiT ?1 iT : (2)
Hence we have:
X
P(O j ) = P(O j I; ) P(I j ) (3)
I
X
i1 bi1 (O1)ai1i2 bi2 (O2)


aiT ?1iT biT (OT ): (4)
I
where I = i1,i2,. . . ,iT .
From Eq. 4 we see that the summand of this equation involves 2T {1 multiplications and there
exists N T distinct possible state sequences I . Hence a direct computation from ( 4) will involve of
the order of 2TN T multiplications. Even for small values, N =5 and T =100, this means approxi-
mately 1072 multiplications which would take eons to complete even for a supercomputer. Hence
we see that a more ecient procedure is required to solve Problem 1; such a procedure exists and
is called the forward-backward procedure.
Forward-Backward Procedure
Consider the forward variable t(i ) dened as :
t(i) = P (O1; O2; : : :; Ot; it = i j ) (5)
i.e the probability of the partial observation sequence up to time t and the state i at time t, given
the model . t (i ) can be computed inductively as follows:
1.
1(i ) = ibi(O1) ; 1  i  N (6)
2. for t = 1,2,.. . ,T {1, 1  j  N
"N #
X
t+1(j) = t(i)aij bj (Ot+1) (7)
i=1

3. then we have:
X
N
P(O j ) = T (i ) (8)
i=1

3
In Step 2 we want to compute the probability of partial observation sequence up to time t+1 and
state j at time t+1; state j can be be reached (with probability aij ) independently from any of the
N states at time t. The summation in Eq. 7 refers to this fact. Also the summand gives observation
sequence up to time t; hence the bj (Ot+1 ) outside the brackets. In Step 3 we just sum up all
possible (independent) ways of realizing the given observation sequence. Let us illustrate this by
taking the case of
2(j) = P (O1; O2; i2 = j )
(We have dropped dependence on  for convenience). The sequence O1 ,O2 ,i2 =j occurs as: rst O1,
then state i2 , then object O2. But O1 can occur through any of the following mutually exclusive
and exhaustive ways : state 1 then O1 , state 2 then O1, and so on up to state N . We know that if
fSig be a set of mutually exclusive and exhaustive events then for any event E we have
X
P (E ) = P (E j Si)P (Si) (9)
i

Hence here we can write

2(j) = PX(O1; O2; i2 = j )
= P (O2; i2 = j j O1from state i)P (O1from state i)
i
X
= ([P (O2 j i2 = j; O1 from state i)P (i2 = j j O1from state i)] [P (O1 j i1 = i)P (i1 = i)])
i
X
= ([P (O2 j i2 = j )P (i2 = j j i1 = i)] [P (O1 j i1 = i)P (i1 = i)])
i
X
= ([bj (O2 ) aij ] [bi (O1 ) i ])
i
" #
X
= (i bi (O1 ) )aij bj (O2)
i
" #
X
= 1(i )aij bj (O2)
i
(using (6) above)
which is the same as (7) above. Hence this simple example gives an insight into the mathematical
justication of (7).
As for (8) using (9 ) we can write
X
P (O) = P (O j iT = i)P (iT = i)
i
X
= P (O; iT = i)
i
X
= T (i)
i
(using the denition given in (5))
Hence proved.
Now let us examine the number of computations involved in this algorithm. Step 1 involves N
multiplications. In Step 2 the summation involves N multiplications plus one for the out of bracket
bj (Ot+1) term|this has to be done for j=1 to N and t = 1 to T{1, making the total number of

4
multiplications in Step 2 as (N+1)N(T{1). Step 3 involves no multiplications. Hence total number
of multiplications is N+N(N+1)(T{1) i.e. of the order of N 2T as compared to 2T
N T required for
the direct method1 . For N=5 and T=100 we need about 3000 computations for the forward method
as compared to 1072 required by the direct method{a saving of about 69 orders of magnitude!
Backward Procedure
In a similar manner we may dene a backward variable t(i ) as:
t(i ) = P (Ot+1; Ot+2; : : :; OT j it = i; ) (10)
i.e. the probability of the observation sequence from t+1 to T given the state i at time t and the
model . Note that here it=i has already been given (it wasn't in the case of the forward variable).
This distinction has been made to be able to combine the forward and the backward variables to
produce useful results, as we shall see soon. We can easily solve for t(i ) as done for t (i ) :
1.
T (i) = 1; 1  i  N (11)
2. for t = T{1,T{2,.. . ,1, 1  i  N
X
N
t(i ) = aij bj (Ot+1) t+1(j) (12)
j =1

3.
X
N
P(O j ) = ibi(O1) 1(i) (13)
i=1

The proof of (12) and (13) is similar to the one given for (7) and (8) and is left as an exercise
for the reader.
The computation of P(O j ) using t(i ) also involves of the order of N 2T calculations. Hence
both the forward as well as the backward method is equally ecient for the computation of P(O j )
. This solves Problem 1.
3.2 Problem 2
Here we have to nd a state sequence I = i1 ,i2 ,. . . ,iT such that probability of occurrence of the
observation sequence O = O1 ,O2,. . . ,OT from this state sequence is greater than that from any
other state sequence. In other words, our problem is to nd I that will maximize P(O,I j ) .
There is a famous algorithm to do this, called the Viterbi Algorithm [2]. It is an inductive
algorithm in which at each instant you keep the best (i.e. the one giving maximum probability)
possible state sequence for each of the N states as the intermediate state for the desired observation
sequence O = O1,O2 ,. . . ,OT . In this way you nally have the best path for each of the N states
as the last state for the desired observation sequence. Out of these, we select the one which has
highest probability.
In order to get an idea of the Viterbi algorithm as applied to the optimum state estimation
problem a simple reformulation of the problem will be useful
Consider the expression for P (O; I j); from (1)-(2) we have
1
On similar lines it can be seen that the number of additions involved is N(N{1)(T{1) + (N{1).

5
 P (O; I j) = P (OjI; )P (I j)
= i1 bi1 (O1)ai i bi (O2 )


ai
1 2 2 ?
T 1 iT bi (OT )
T

Now dene
"    #
X
T
U ( i1,i2,. . . ,iT ) = ? ln i1 bi1 (O1) + ln ait?1 it bit (Ot ) (14)
t=2
then it is easily seen that
P (O; I j) = exp(?U ( i1,i2,. . . ,iT ))
consequently the problem of optimal state estimation, namely,
max P (O; i1


iT j)
fit gTt=1
becomes equivalent to
min U ( i1 ,i2 ,. . . ,iT ) (15)
fit gTt=1
This reformulation now enables us to view terms like ? ln(aij ik bik (Ot )) as the cost associated in
going form state ij to state ik at time t.
The Viterbi algorithm to nd the optimum state sequence can now be described as follows:
Suppose we are currently in state i and we are considering visiting state j next. We shall say that
the weight on the path from state i to state j is ? ln(aij bj (Ot ) ) (i.e. the negative of logarithm of
probability of going from state i to state j and selecting the observation symbol Ot in state j ) where
Ot is the observation symbol selected after visiting state j {this is the same symbol that appears in
the given observation sequence O = O1,O2 ,. . . ,OT .When the initial state is selected as state i the
corresponding weight is ? ln(ibi (O1 ) ){ we shall call this the initial weight. We dene the weight
of a sequence of states as the sum of the weights on the adjacent states. Note that this actually
corresponds to multiplying the corresponding probabilities. Now nding the optimum sequence is
merely a matter of nding the path (i.e. a sequence of states) of minimum weight through which
the given observation sequence occurs.
Note that the Viterbi Algorithm is essentially a dynamic programming approach for minimizing
U( i1,i2,. . . ,iT ).
We shall illustrate with an example on how the minimization is done. Consider a three state
HMM. Fig.1(a) shows the weights assigned to various paths as described above. The numbers in
circles show the initial weights assigned to the corresponding states. To simplify understanding of
the algorithm we have assumed that the weights are symmetrical (i.e. the weight from state i to
state j is same as the weight from state j to state i). Also we have assumed the weights do not
change with time which in general will not be true but then once this simplied case is understood
the extension to practical case is straight forward.
Consider Fig. 1(b). We take an observation sequence O1 ,.. . ,O4 of length four. At time t=2
state 1 can be visited from any of the three states that we had at time t=1. We nd out the
weights on each of these paths and add corresponding weights to the initial weights ( we shall call
this cumulative weight at state 1 as the cost at state 1 at time t=2). Thus going from state 1 to
state 1 the cost at state 1 is 3 + 6 = 9. Similarly the cost at state 1 in going from state 2 to state 1

6
 6 4 7
state 1 state 2
3 3

3 4

state 3
8 2

(a)

state , t 1 2 3 4

1 3 5 9 11
4
3 3

2 3 6 9 12
3 3
3 3
4 4

3 2 6 8 12

(b)

Figure 1: (a) The state machine. The weights for transitions between the states have been shown.
The circled numbers are the initial weights of the corresponding states. (b) How a path of minimum
cost is traced out using the Viterbi algorithm. The cumulative weights are shown in bold and the
nal minimum cost path is also shown in bold.

is 3 + 4 = 7 and the cost in going from state 3 to state 1 is 2 + 3 = 5. Of these the cost 5 is
minimum. Hence we retain this cost at state 1 for further calculations. The minimum cumulative
weight paths are shown in the gure by arrowed lines. The cumulative weights have been shown in
bold alongside the respective states at each time instant. We repeat the same procedure for state 2
and state 3. We see that the minimum costs at state 2 and state 3 are 6 (through state 3 ) and 6
(through state 1) respectively.
We repeat the above procedure again for t=3 but now using the costs calculated above for each
state rather than the initial weights ( we used them above because we started with t=1). And the
same procedure is repeated for t=4. Now select out the state which has minimum cost of all the
states. We see that state 1 is the required state with a cost of 11. Back tracing the sequence of
states through which we got at state 1 at time t=4 gives the required sequence of states through
which the given observation sequence has highest probability of occurrence. As can be seen from
the gure this state sequence is state 3,state 1,state 3,state 1 . This sequence has been shown in
bold in the gure.
To prove our point suppose you were given that the length of observation sequence is required
to be two and that the last state is to be state 3. What path would you choose to minimize the
cost (i.e. to maximize the probability of the observation sequence O1 O2 ) ? The procedure outlined
above clearly shows that we would choose the path beginning at state 1 and ending at state 3
(at t=2) as that gave the minimum cost (viz. 6) at state 3. All other paths have a higher cost.
Similar argument applies if any other state were required to be the last state. Similarly if the
observation sequence were to be of length three and ending in say state 1 we would choose the path

7
state 1,state 3,state 1 as outlined in the gure and described above. This means that at any given
instant the path tracked upto any state by the above procedure is the minimum cost path if we
were to stop at that instant at that state. Proceeding to t=4 we see that we have the minimum
cost paths corresponding to stopping in state 1,state 2 or state 3 respectively. We just pick up the
smallest of these three because we are interested in the minimum cost and hence we choose to stop
in state 1 which gives us the minimum cost.
Now that we know how the Viterbi Algorithm works here is how it can be implemented [1, 3] :
Note : t (i ) denotes the weight accumulated when we are in state i at time t as the algorithm
proceeds and t (j) represents the state at time t ? 1 which has the lowest cost corresponding to
the state transition to state j at time t. For example in Fig. 1(b),
3(1 ) = 9; 3(2 ) = 9; 3(3 ) = 8 (16)
and
4 (1 ) = 3; 4 (2 ) = 3; 4 (3 ) = 1 (17)
1. Initialization
For 1  i  N
1(i) = ? ln(i) ? ln(bi(O1)) (18)
1 (i ) = 0 (19)
2. Recursive computation
For 2  t  T for 1  j  N
t(j) = 1min
iN
[t?1 (i ) ? ln(aij )] ? ln(bj (Ot )) (20)
t (j) = arg min [t?1 (i ) ? ln(aij )] (21)
1iN

3. Termination

P  = 1min
iN
[T (i )] (22)
qT = arg 1min
iN
[T (i )] (23)

4. Tracing back the optimal state sequence

For t = T ? 1; T ? 2; : : :; 1
qt = t+1 (qt+1 ) (24)
Hence exp(?P  ) gives the required state-optimized probability , and Q = fq1 ; q2; : : :; qT g is the
optimal state sequence.
A little re ection over the above steps will show that computationally the Viterbi Algorithm is
similar to the forward (-backward) procedure except for the comparisons involved for nding the
maximum value. Hence its complexity is also of the order of N 2 T . This solves Problem 2.

8
3.3 Problem 3
This problem deals with training the HMM such that it encodes the observation sequence in such
a way that if a observation sequence having many characteristics similar to the given one be
encountered later it should be able to identify it. There are two methods that we shall discuss
depending on which probability is chosen for identication:
The Segmental K-means Algorithm : In this method the parameters of the model  = (A,B; )
are adjusted to maximize P(O,I j ) where I here is the optimum sequence as given by the
solution to Problem 2 above.
The Baum-Welch re-estimation Formulas : Here parameters of the model  = (A,B; ) are
adjusted so as to increase P(O j ) until a maximum value is reached. As seen before cal-
culating P(O j ) involves summing up P(O,I j ) over all possible state sequences I . Hence
here our focus is not on a particular state sequence.
3.3.1 The Segmental K-means Algorithm ([3], [4])
This method takes us from k to k such that P (O,I k j k )  P (O,I k j k ) where, I k is the
+1
+1
+1

optimum sequence for O = O1 ,O2,. . . ,OT and k , found according to the solution to Problem 2 .
This criterion of optimization is called the maximum state optimized likelihood criterion.
This function P (O,I  j ) = maxI P (O,I j ) is called the state optimized likelihood function.
In this method to train the model a number of (training) observation sequences are required. Let
there be ! number of such sequences available. Each sequence O = O1 ,O2 ,.. . ,OT consists of T
observation symbols. Instead of ! number of such sequences we might be given just one very long
sequence; in that case we segment it into some conveniently chosen ! number of short sequences
each of length T. Each observation symbol (Oi ) is assumed2 to be a vector of dimension D ( 1).
The algorithm then consists of the following steps:
1. Randomly choose N observation symbols (vectors of dimension D) and assign each of the !T
observation vectors to one of these N vectors from which its Euclidean distance is minimum.
Hence we have formed N clusters each called a state (1 to N). This initial choice of clustering
vectors doesn't decide the nal HMM that we get but can decide the number of iterations
required for the HMM training. In our case we just pick up N equally spaced sequences of
feature vectors and pick one vector from each of these sequences. The rst vector is taken as
the rst vector of the rst of these sequences, the second as the second of the second of these
sequences and so on. Of course this is just to make the initial choice of clusters as widely
distributed as possible and one is at liberty to choose the vectors as per his need.
2. Calculate the initial probabilities and the transition probabilities : For 1  i  N :
Number of occurrences of fO1 2 ig
^i = Total (25)
number of occurrences of O1 (i.e. ! )
For 1  i  N and 1  j  N :
of occurrences of fOt 2 i andOt+1 2 j g for all t
^aij = NumberNumber (26)
of occurrences of fOt 2 ig for all t
2
If the observation symbols aren't actually given as numbers or vectors we shall have to characterize each of them
with a vector

9
 3. Calculate the mean vector and the covariance matrix for each state : For 1  i  N
^ = 1
X
i O Ni Ot2 i t (27)

V^i = N1
X
(Ot ? ^i )T (Ot ? ^i ) (28)
i Ot2 i

4. Calculate the symbol probability distributions for each training vector for each state as (we
assume Gaussian distribution { there is no particular reason for this ,just change the formulas
below for the particular probability distribution that sui ts your problem ): For 1  i  N
^bi(Ot ) = 1 exp[ ? 1 (Ot ? ^i )V^i?1 (Ot ? ^i )T ] (29)
(2 )D=2 ^
jVij 1=2 2
5. Find the optimal state sequence I  (as given by the solution to Problem 2) for each training
sequence using ^ i = (A^i ,B^i ; ^i) computed in steps 2 to 4 above. A vector is reassigned a state
if its original assignment is dierent from the corresponding estimated optimum state; for
example suppose O2 of the 5th training sequence was assigned to state 3 (i.e. the 3rd cluster)
and now we nd that the in the optimal state sequence I  corresponding to the 5th training
sequence, i2  is not 3 but say 4. Hence now we reassign O2 of the 5th training sequence to
state 4. Hence we assign Ot (of say kth training sequence ) to state i if it  (corresponding to
the kth training sequence ) is state i and this is done for all training sequences (i.e. for k=1
to ! ).
6. If any vector is reassigned a new state in Step 5, use the new assignment and repeat Step 2
through Step 6; otherwise, stop.
It can be shown [4] that the segmental K-means algorithm converges to the state-optimized likeli-
hood function for a wide range of observation density functions including the Gaussian density we
have assumed. Please refer [4] for the proof.
3.3.2 The Baum-Welch re-estimation Formulas [1] :
This method assumes an initial HMM model which is improved upon by using the formulas (given
below) so as to maximize P(O j ) . An initial HMM can be constructed in any way but we may
use the rst ve steps of the Segmental K-means Algorithm described above to give us a reasonable
initial estimate of the HMM. Henceforth we shall assume that an initial HMM is known. This
algorithm maximizes P(O j ) by adjusting the parameters of . This optimization criterion is called
the maximum likelihood criterion. The function P(O j ) is called the likelihood function.
Before we get down to the actual formulas we shall introduce some concepts and notations that
shall be required in the nal formulas. Consider t(i ) dened as follows:
t(i ) = P (it = i jO; )
i.e. the probability of being in state i at time t given the observation sequence O = O1 ,O2 ,. . . ,OT and
the model  = (A,B;  ) . By Bayes law we have :
t = i; O j)
t(i ) = P (iP( Oj ) (30)
= P(t (i )t(i )
Oj ) (31)

10
where t (i ) and t(i ) are dened by (5) and (10) respectively. t (i ) accounts for O1 ,. . . ,Ot and
state i at time t, and t(i ) accounts for Ot+1 ,. . . ,OT given state i at time t .
We also dene t (i,j) as :
t(i,j) = P (it = i; it+1 = j j O; )
i.e. the probability of being in state i at time t and making a transition to state j at time t + 1,
given the observation sequence O = O1,O2 ,. . . ,OT and the model  = (A,B;  ) . Using Bayes law
it is seen that
t(i,j) = P (it = i;P(it+1 = j; O j ) (32)
Oj )
But O = O1 ,O2,. . . ,OT . Hence
The Numerator = P (it = i; O1 ,. . . ,Ot ; Ot+1 ,. . . ,OT ; it+1 = j j ) (33)
= P (it = i; O1 ,. . . ,Ot j )P ( Ot+1 ,. . . ,OT ; it+1 = j j ) (34)
(since the Markov chain is causal) (35)
Consider the second term. The observation symbol at time t + 1 is Ot+1 at which time we are
required to be in state j ; this means that the observation symbol Ot+1 has to be picked up from
state j .Hence we have
P ( Ot+1,. . . ,OT ; it+1 = j j ) = P (it+1 = j; Ot+1 j )P ( Ot+1,. . . ,OT j it+1 = j; ) (36)
= aij bj (Ot+1 ) t+1 (j) (37)
(since it =i is known from (34)) (38)
Hence combining (5),(32),(34) and (37) we get
t(i,j) = t(i )aij bP(
j (Ot+1 ) t+1(j )
Oj ) (39)
Here t (i ) accounts for O1 ,. . . ,Ot ,aij accounts for the transition to state j , bj (Ot+1 ) picks
up the symbol Ot+1 from state j ,and t+1 (j) accounts for the remaining observation sequence
Ot+2,. . . ,OT .
If we sum up t(i ) from t = 1 to T we get a quantity which can be viewed as the expected
number of times state i is visited or if we sum up only up to T ? 1 then we shall get the expected
number of transitions out of state i (as no transition is made at t = T ). Similarly if t (i,j) be
summed up from t = 1 to T ? 1 , we shall get the expected number of transitions from state i to
state j . Hence
?1
TX
t(i) = Expected number of transitions from state i
t=1
?1
TX
t(i,j) = Expected number of transitions from state i to state j
t=1
Now we are prepared to present the Baum-Welch re-estimation formulas:
^i = t(i ); 1  i  N (40)
?1
TX
,T ?1
X
^aij = t(i; j) t(i) (41)
t=1 t=1
,T
X
T X
^bj (k ) = t(j ) t(j ) (42)
t=1 t=1
Ot =k

11
The re-estimation formula for i is simply the probability of being in state i at time t. The formula
for aij is the ratio of expected number of time of making a transition from state i to state j to the
expected number of times of making a transition out of state i. The formula for bk (i )s the ratio of
the expected number of times of being in state j and observing symbol Ok to the expected number
of times of being in state j . Note that the summation in the last formula goes up to t = T .
If we denote the initial model by  and the re-estimation model by ^ consisting of the parameters
estimated above, then it can be shown that either:
1. The initial model  is a critical point of the likelihood function in which case ^ = , or
2. P (O j ^ ) > P(O j ); i.e. we have found a better model from which the observation sequence
O = O1,O2,. . . ,OT is more likely to be produced.
Hence we can go on iteratively computing until P(O j ) is maximized. This solves Problem 3.
We have seen that the Baum-Welch algorithm maximizes P(O j ) ; given by (4). Each term in
the summand of this equation is between 0 and 1 and hence the summand (which is a product of
many such terms) is going to be very small and there are as many such small terms as the number
of possible state sequences { for a computer to be able to compute such terms individually, it should
be able to store within its precision the smallest of these terms. If T =100 there would be around
200 numbers (each between 0 and 1) to be multiplied and hence each term could comfortably go
beyond the precision of any computer. Taking logarithm does not help here since P(O j ) is given
by a summation of such small terms. Perhaps some clever scaling procedure could be used to get
over this problem.
On the other hand the Segmental k-means algorithm maximizes P(O,I j ) which is calculated
using the Viterbi Algorithm as given by (18){(24). Here we can use the logarithm to take care of
product of many small numbers as no summation is involved. The Segmental k-means algorithm
is also preferred because most of the time we are interested in the occurrence of the observation
sequence from the best possible (i.e. the optimum) state sequence and considering the occurrence of
the observation sequence through all the possible state sequences is not the intended representation
of the problem in most cases. Moreover the Segmental k-means algorithm requires much less
computation as compared to the Baum-Welch method. For all these reasons we have used the
Segmental k-means algorithm in our work.

4 Distance Between HMMs

If we want to compare two HMMs then we need a measure for distance between two HMMs. Such
a measure has been developed by Juang and Rabiner [5]. Their development is based on the
Kullback-Leibler distance between two probability distribution function; thus before presenting the
measure for distance between tow HMMs we present a very brief introduction to Kullback-Leibler
distance measure.
4.1 Kullback-Leibler Distance Measure
Let p1 (x) and p2(x) be two probability density functions (or probability mass functions) then, the
Kullback-Leibler distance measure can be used to judge how close the two probability distributions
are.
Denition The Kullback-Leibler distance measure for determining how close the probability den-
sity function p2(x) is to p1 (x) with respect to p1 (x) is

12
 Z1
I (p1; p2) = p1(x) ln( pp1((xx)) ) dx
?1 2

In case p1(x) and p2 (x) are probability mass functions then

p1(x) ln( pp1((xx)) )

X
I (p1; p2) =
all x 2

Note that the Kullback-Leibler distance measure is not symmetric, I (p1; p2) 6= I (p2; p1). If our
objective is to simply compare p1 and p2 we can dene a symmetric distance measure as
Is(p1; p2) = 12 [I (p1; p2) + I (p2; p1)]

A Typical Approximation Problem: Given the density functions p(x), we would like to
approximate it with another density function p^(x) (where p^(x) could be a parameterized function).
Then for obtaining p^(x), we consider the problem
p^ = arg min
p^
[I (p; p^)]
R R
Using I (p; p^) = x p(x) ln p(x) ? x p(x) ln p^(x) the above minimization is equivalent to
Z
p^ = arg max
p^
p(x) ln p^(x)
x

Moreover I (p; p^) can also be interpreted as

I (p; p^) = Ep(x)[ln p(x)] ? Ep(x)[ln p^(x)]
where Ep(x) is the expectation operator with respect to the density function p(x).
4.2 Distance between HMMs
It is the use of the Kullback-Leibler distance measure which leads to the denition of the distance
measure between two HMMs.
Let 1 and 2 be two HMMs. For example 1 could be the model HMM and 2 could be the
HMM based on the observations. We are interested in comparing how close they are.
Oj1] )
I (1; 2) = P [Oj1] ln( PP [[Oj
X
O 2 ]
In practice we cannot compute th above sum, because it would be impossible to generate all
possible sequences. Suppose we have generated an observation sequence of length T using the
HMM i ; denote this sequence by OTi . Now, corresponding to this sequence we can talk about
computing P [OT1 j2]. For this computation we will have to assume some state sequence, which at
least is not available for HMM 2. It is likely that we may have a state sequence for 1, since
the observation sequence OT1 was generated using HMM 1. Nevertheless, we do not use this
information. What one does is to compute the optimal state sequence using the Viterbi algorithm

13
for 1 and 2 corresponding to OT1 . Let the optimal state sequences be I 1(OT1 ) and I 2(OT1 ). Now,
for closeness between 1 and 2, we consider
P [OT1 ; I 1(OT1 )j1]fln(P [OT1 ; I 1(OT1 )j1]) ? ln(P [OT1 ; I 2(OT1 )j2])g
We would like to compute the above for a large value of T , the length of the observation sequence.
The probability function outside the braces can be ignored since it is common to both the terms
inside the braces. The distance between two HMMs is dened as
1 2 = lim 1 j ln(P [O1 ; I 1(O1 )j ]) ? ln(P [O1 ; I 2(O1 )j ])j
D( ;  ) def T
T !1
T T 1 T T 2

Once again we see that D(1; 2) is not symmetric. Nevertheless a symmetric distance between
HMMs is dened as
= 21 [D(1; 2) + D(2; 1)]
Ds(1; 2) def
What we have given above is a heuristic derivation of the formula for distance between two
HMMs. The formal derivation is given in Juang and Rabiner [5].
4.3 How Large Should T be?
In practice one cannot have T ! 1. To get over this problems we did some experimentation
as suggested by [5] to nd out how large a value of T should be chosen. The details of how the
experiment was performed are as follows:
Let two HMMs 1 and 2 , the distance between which is to be found, be given. We generate
feature sequences from the HMM 1.
Let S be the given covariance matrix. Hence S is symmetric and semi-positive denite. Since
we generate a large number of feature vectors, the covariance matrix in almost all cases is also
non-singular and hence can be assumed to be positive denite. Using Cholesky decomposition, S
can be written as
S = LLT (43)
where L is a lower triangular matrix.
Now suppose we want to generate n zero mean vectors each of dimension p with covariance
matrix as S . Generate an n  p matrix A whose elements are Gaussian random variables with zero
mean and standard deviation of 1 i.e. N(0,1). Then the rows of
B = AL (44)
give the required n zero mean vectors with covariance matrix S . Note that L is a p  p matrix.
Now simply add the required mean to all the generated vectors.
Next, we calculate the probability of the observation sequence generated from HMM 1 against
itself and also against the HMM 2 and compute the asymmetric distance D(1; 2). Similarly we
compute the probability of the observation sequence generated from 2 against itself and 1 and
again compute the corresponding asymmetric distance D(2; 1). Taking the average of these two
distances gives the required distance.
The result of such an experimentation using two HMMs 1 and 2 and a seed is shown in Fig.2.
From experimentation with dierent HMMs it was found that as T increases the distance ap-
proaches a limit. Also the distance has reasonably settled after T = 212 i.e. around 4000.

14
 HMM Distance vs. Feature Sequence Length
7.5
"distance.data"
7

6.5
HMM distance ->

5.5

4.5

3.5
4 6 8 10 12 14 16 18 20
log2(T) ->
Figure 2: Convergence of the distance between two HMMs as the length of the observation sequence
is increased

15
5 Conclusion
It is hoped that this would serve as good introduction to HMMs for anyone to venture further into
the realm of HMMs.

References
[1] L. R. Rabiner and B. H. Juang, \An introduction to hidden Markov models," IEEE ASSP
Mag., pp 4{16, Jun. 1986.
[2] G.D.Forney Jr., \The Viterbi Algorithm," Proc. IEEE, vol. 61, no. 3, pp. 263{278, Mar. 1973.
[3] Yang He and Amlan Kundu, \2-D shape classication using HMMs," IEEE Trans. Patt. Anal.
Machine Intell., vol. 13, no. 11, pp. 1172{1184, Nov. 1991.
[4] B. H. Juang and L. R. Rabiner, \The segmental k-means algorithm for estimating the pa-
rameters of hidden Markov models," IEEE Trans. Accoust., Speech, Signal Processing, vol.
ASSP-38,no. 9, pp. 1639{1641, Sept. 1990.
[5] B. H. Juang and L. R. Rabiner, \A Probabilistic Distance measure between HMMs", Vol. 64,
no. 2, pp. 391-408, Feb. 1985.

16