Simulation Based Optimization-1

Chapter 6
CONTROL OPTIMIZATION
WITH STOCHASTIC DYNAMIC
PROGRAMMING
1. Chapter Overview
This chapter focuses on a problem of control optimization, in
particular the Markov decision problem (or process). Our discussions
will be at a very elementary level, and we will not attempt to prove
any theorems. The central aim of this chapter is to introduce the
reader to classical dynamic programming in the context of solving
Markov decision problems. In the next chapter, the same ideas will be
presented in the context of simulation-based dynamic programming.
The main concepts presented in this chapter are (1) Markov chains,
(2) Markov decision problems, (3) semi-Markov decision problems,
and (4) classical dynamic programming methods.
2. Stochastic Processes
We begin with a discussion on stochastic processes. A stochastic
(or random) process, roughly speaking, is an entity that has a prop-
erty which changes randomly with time. We refer to this changing
property as the state of the stochastic process. A stochastic process
is usually associated with a stochastic system. Read Chap. 2 for a
definition of a stochastic system. The concept of a stochastic process
is best understood with an example.
Consider a queue of persons that forms in a bank. Let us assume
that there is a single server (teller) serving the queue. See Fig. 6.1.
The queuing system is an example of a stochastic system. We need
to investigate further the nature of this queuing system to identify
properties, associated with the queue, that change randomly with time.
© Springer Science+Business Media New York 2015 123

A. Gosavi, Simulation-Based Optimization, Operations
Research/Computer Science Interfaces Series 55,
DOI 10.1007/978-1-4899-7491-4 6
124 SIMULATION-BASED OPTIMIZATION
Let us denote
The number of customers in the queue at time t by X(t) and
The number of busy servers at time t by Y (t).
Then, clearly, X(t) will change its value from time to time and so
will Y (t). By its definition, Y (t) will equal 1 when the teller is busy
serving customers, and will equal 0 when it is idle.
Customer being
served
Server
Customers
in the queue
Figure 6.1. A single-server queue
Now if the state of the system is recorded after unit time, X(t)
could take on values such as: 3, 3, 4, 5, 4, 4, 3 . . . The set {X(t)|t =
1, 2, · · · , ∞}, then, defines a stochastic process. Mathematically, the
sequence of values that X(t) assumes in this example is a stochastic
process.
Similarly, {Y (t)|t = 1, 2, · · · , ∞} denotes another stochastic process
underlying the same queuing system. For example, Y (t) could take on
values such as 0, 0, 1, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1, . . .
It should be clear now that more than one stochastic process may be
associated with any given stochastic system. The stochastic processes
X and Y differ in their definition of the system state. For X, the
state is the number of customers in the queue and for Y , the state is
the number of busy servers.
An analyst selects the stochastic process that is of interest to
him/her. E.g., an analyst interested in studying the utilization of
the server (i.e., proportion of time the server is busy) will choose Y ,
while the analyst interested in studying the length of the queue will
choose X. See Fig. 6.2 for a pictorial explanation of the word “state.”
In general, choosing the appropriate definition of the state of a
system is a part of “modeling.” The state must be defined in a manner
suitable for the optimization problem under consideration. To under-
stand this better, consider the following definition of state. Let Z(t)
denote the total number of persons in the queue with black hair. Now,
Dynamic Programming 125
State = 3
Customers Server
State = 2
Customers Server
Customers Server
Figure 6.2. A queue in two different states: The “state” is defined by the number
in the queue
although Z(t) is a mathematically perfect example of a stochastic pro-

cess, this definition may contain very little information of use, when it
comes to controlling systems in a cost-optimal manner!
We have defined the state of a stochastic process. Now, it is time to
closely examine an important stochastic process, namely, the Markov
process.
3. Markov, Semi-Markov, and Decision

Processes
A Markov process is of special interest to us because of its
widespread use in studying real-life systems. In this section, we
will study some of its salient features.
A stochastic process, usually, visits more than one state. We will
assume throughout this book that the set of states visited by the
stochastic process is a finite set denoted by S.
Figure 6.3. Schematic of a two-state Markov chain, where circles denote states
An important property of a Markov process is that it jumps regu-

larly. In fact, it jumps after unit time. (Some authors do not use
the “unit time” convention, and we will discuss this matter in detail
later.) Hence, after unit time, the system either switches (moves) to
a new state or else the system returns to the current state. We will
refer to this phenomenon as a state transition.
To understand this phenomenon better, consider Fig. 6.3. The figure
shows two states, which are denoted by circles, numbered 1 and 2.
The arrows show the possible ways of transiting. This system has
two states: 1 and 2. Assuming that we first observe the system when
it is in state 1, it may for instance follow the trajectory given by:
1, 1, 2, 1, 1, 1, 2, 2, 1, 2, . . .
A state transition in a Markov process is usually a probabilistic, i.e.,
random, affair. Consider the Markov process in Fig. 6.3. Let us further
assume that in its first visit to state 1, from state 1 the system jumped
to state 2. In its next visit to state 1, the system may not jump to
state 2 again; it may jump back to state 1. This should clarify that
the transitions in a Markov chain are “random” affairs.
We now need to discuss our convention regarding the time needed
for one jump (transition). In a Markov process, how much time is spent
in one transition is really irrelevant to its analysis. As such, even if the
time is not always unity, or even if it is not a constant, we assume it
to be unity for our analysis. If the time spent in the transition becomes
an integral part of how the Markov chain is analyzed, then the Markov
process is not an appropriate model. In that case, the semi-Markov
process becomes more appropriate, as we will see below.
When we study real-life systems using Markov processes, it usu-
ally becomes necessary to define a performance metric for the real-life
system. It is in this context that one has to be careful with how the
unit time convention is interpreted. A common example of a perfor-
mance metric is: average reward per unit time. In the case of a Markov
process, the phrase “per unit time” in the definition of average reward
actually means “per jump” or “per transition.” (In the so-called semi-
Markov process that we will study later, the two phrases have different
meanings.)
Another important property of the Markov process needs to be
studied here. In a Markov process, the probability that the process
jumps from a state i to a state j does not depend on the states vis-
ited by the system before coming to i. This is called the memoryless
property. This property distinguishes a Markov process from other
stochastic processes, and as such it needs to be understood clearly.
Because of the memoryless property, one can associate a probability
with a transition from a state i to a state j, that is,
i −→ j.
We denote the probability of this transition by P (i, j). This idea is

best explained with an example.
Consider a Markov chain with three states, numbered 1, 2, and 3.
The system starts in state 1 and traces the following trajectory:
1, 3, 2, 1, 1, 1, 2, 1, 3, 1, 1, 2, . . .
Assume that: P (3, 1) = 0.2 and P (3, 2) = 0.8. When the system visits
3 for the first time in the above, it jumps to 2. Now, the probability of
jumping to 2 is 0.8, and that of jumping to 1 is 0.2. When the system
revisits 3, the probability of jumping to 2 will remain at 0.8, and that
of jumping to 1 at 0.2. Whenever the system comes to 3, its probability
of jumping to 2 will always be 0.8 and that of jumping to 1 be 0.2. In
other words, when the system comes to a state i, the state to which
it jumps depends only on the transition probabilities: P (i, 1), P (i, 2)
and P (i, 3). These probabilities are not affected by the sequence of
states visited before coming to i. Thus, when it comes to jumping to a
new state, the process does not “remember” what states it has had to
go through in the past. The state to which it jumps depends only on
the current state (say i) and on the probabilities of jumping from that
state to other states, i.e., P (i, 1), P (i, 2) and P (i, 3). In general, when
the system is ready to leave state i, the next state j depends only on
P (i, j). Furthermore, P (i, j) is completely independent of where the
system has been before coming to i.
We now give an example of a non-Markovian process. Assume that
a process has three states, numbered 1, 2, and 3. X(t), as before,
denotes the system state at time t. Assume that the law governing
this process is given by:
P{X(t + 1) = j|X(t) = i, X(t − 1) = l} = f (i, l, j), (6.1)
where f (i, l, j) denotes a probability that depends on i, l, and j. This

implies that if the process is in state i at time t (notice that X(t) = i in
the equation above is supposed to mean exactly this), and if it was in
state l at time (t − 1), then the probability that the next state (i.e.,
the state visited at time (t + 1)) will be j is a function of i, l, and j. In
other words, at any point of time, the past (i.e., X(t − 1)) will affect
its future course.
Thus the process described above is not a Markov process. In this
process, the state of the process at time (t−1) does affect the probabil-
ities of going to other states at time (t+1). The path of this stochastic
process is thus dependent on its past, not the entire past, but some of
its past.
The Markov process, on the other hand, is governed by the follow-

ing law:
P{X(t + 1) = j|X(t) = i} = f (i, j), (6.2)
where f (i, j) is the probability that the next state is j given that the
current state is i. Also f (i, j) is a constant for given values of i and j.
Carefully note the difference between Eqs. (6.2) and (6.1). Where
the process resides one step before its current state has no influence on
a Markov process. It should be obvious that in the Markov process,
the transition probability (probability of going to one state to another
in the stochastic process in one step) depends on two quantities:
the present state (i) and the next state (j). In a non-Markovian pro-
cess, such as the one defined by Eq. (6.1), the transition probability
depended on the current state (i), the next state (j), and the previous
state (l). An implication is that even if both the processes have the
same number of states, we will have to deal with additional probabil-
ities in the two-step stochastic process.
The quantity f (i, j) is an element of a two-dimensional matrix. Note
that f (i, j) is actually P (i, j), the one-step transition probability of
jumping from i to j, which we have defined earlier.
All the transition probabilities of a Markov process can be conve-
niently stored in a matrix. This matrix is called the one-step tran-
sition probability matrix or simply the transition probability ma-
trix, usually abbreviated as TPM. An example of a TPM with three
states is:
 
0.7 0.2 0.1
P = 0.4 0.2 0.4 . (6.3)
0.6 0.1 0.3
P (i, j) here denotes the (i, j)th element of the matrix, P, i.e., the
element in the ith row and the jth column of P. In other words, P (i, j)
denotes the one-step transition probability of jumping from state i to
state j. Thus, for example, P (3, 1), which is 0.6 above, denotes the
one-step transition probability of going from state 3 to state 1.
We will also assume that a finite amount of time is taken in any
transition and that no time is actually spent in a state. This is one
convention (there are others), and we will stick to it in this book. Also,
note that by our convention, the time spent in a transition is unity (1).
In summary, a Markov process possesses three important properties:
(1) the jumpy property, (2) the memoryless property, and (3) the unit
time property (by our convention).
3.1. Markov Chains

A Markov chain can be thought of as an entity that accompanies a
stochastic process. Examples of stochastic processes accompanied by
Markov chains that we will consider are the Markov process and the
semi-Markov process.
We associate a unique TPM with a given Markov chain. It should be
kept in mind that the Markov chain (not the Markov process) contains
no information about how much time is spent in a given transition.
Example 1. Consider Fig. 6.4. It shows a Markov chain with two
states, shown by circles, numbered 1 and 2. The arrow indicates a
possible transition, and the number on the arrow denotes the proba-
bility of that transition. The figure depicts the following facts. If the
process is in state 1, it goes to state 2 with a probability of 0.3, and
with a probability of 0.7, it stays in the same state (i.e., 1). Also, if
the process is in state 2, it goes to state 1 with a probability of 0.4 and
stays in the same state (i.e., 2) with a probability of 0.6. The TPM of
the Markov chain in the figure is therefore
0.7 0.3
P= .
0.4 0.6
0.6
0.7
0.3
2
1 0.4
Figure 6.4. Schematic of a two-state Markov chain, where circles denote states,
arrows depict possible transitions, and the numbers on the arrows denote the prob-
abilities of those transitions
Figures 6.5 and 6.6 show some more examples of Markov chains
with three and four states respectively. In this book, we will consider
Markov chains with a finite number of states.
Estimating the values of the elements of the TPM is often quite
difficult. This is because, in many real-life systems, the TPM is
very large, and evaluating any given element in the TPM requires
the setting up of complicated expressions, which may involve multiple
integrals. In subsequent chapters, this issue will be discussed in depth.
Figure 6.5. Schematic of a Markov chain with three states
3
1
Figure 6.6. Schematic of a Markov chain with four states
n-step transition probabilities. The n-step transition probability

of going from state i to state j is defined as the probability of starting at
state i and being in state j after n steps (or jumps/transitions). From
the one-step transition probabilities, it is possible to construct the two-
step transition probabilities, the three-step transition probabilities,
and the n-step transition probabilities, in general. The n-step transi-
tion probabilities are very often of great importance to the analyst.
The so-called Chapman-Kolmogorov theorem helps us find these
probabilities. The theorem states that the n-step transition proba-
bilities can be obtained by raising the one-step transition probability
matrix to the nth power. We next state the theorem without proof.
Theorem 6.1 If P denotes the one-step transition probability matrix
of a Markov chain and S = Pn , where Pn denotes the matrix P raised
to the nth power, then S(i, j) denotes the n-step transition probability
of going from state i to state j.
Basically, the theorem states says that the nth power of a TPM
is also a transition probability matrix, whose elements are the n-step
transition probabilities. Let us illustrate the meaning of this result

with an example. Consider the TPM given by:
0.7 0.3
P= .
0.4 0.6
Now,
0.7 0.3 0.7 0.3 0.61 0.39
P2 = = .
0.4 0.6 0.4 0.6 0.52 0.48
Here the value of P 2 (1, 1) is 0.61. The theorem says that P 2 (1, 1)
equals the two-step transition probability of going from 1 to 1. Let us
verify this from the basic principles of probability theory. Let Cx−y−z
denote the probability of going from state x to state z in two transitions
with y as the intermediate state. Consider the event of going from state
1 to state 1 in two steps. Then, clearly, the probability of this event
should equal:
C1−1−1 + C1−2−1 .
From the values of P, C1−1−1 = (0.7)(0.7) and C1−2−1 = (0.3)(0.4),
and therefore the required probability should equal:
(0.7)(0.7) + (0.3)(0.4) = 0.61,
which is equal to P 2 (1, 1). The verification is thus complete.
3.1.1 Regular Markov Chains

A regular Markov chain is one whose TPM satisfies the following
property: There exists a finite positive value for n, call it n∗ , such that
for all n ≥ n∗ , and all i and j:
P n∗ (i, j) > 0.
In other words, by raising the TPM of a regular Markov chain to some
positive power, one obtains a matrix in which each element is strictly
greater than 0. An example of a regular Markov chain is:
 
0.7 0.0 0.3
P = 0.4 0.6 0.0 .
0.2 0.7 0.1
It is not hard to verify that P can be raised to a suitable power to
obtain a matrix in which each element is strictly greater than 0. An ex-
ample of a Markov chain that is not regular is:
0 1
P= .
1 0
This is not regular because:

0 1 1 0
Pn = for odd n and Pn = for even n.
1 0 0 1
3.1.2 Limiting Probabilities

If one raises the TPM of a regular Markov chain to higher powers,
the elements in any given column start converging to (that is,
approaching) the same number. For example, consider P of the
previous section, raised to the 8th power:
0.5715 0.4285
P8 = .
0.5714 0.4286
The elements of a given column in Pn start approaching each other,

as we increase the power n, and notice that by n = 8, the elements are
very close to each other. It can be proved that for a regular Markov
chain as the power n tends to infinity, for every j, P n (i, j) starts
converging to a unique finite number. In other words, for every value
of j,
lim P n (i, j) exists and is unique.
n→∞
In the example under consideration, the limit appears to be 0.57 for
state 1 and 0.43 for state 2. We will denote the limit for state j by
Π(j). Mathematically,
Π(j) ≡ lim P n (i, j).

n→∞
The quantity Π(j) will also be referred to as the limiting or steady-

state or invariant probability of the state j.
Now, Π(j), it must be understood, is the long-run probability of
entering the state j from any given state. For instance, in the
example given above, regardless of which state the Markov chain is
in, the long-run (that is, when n → ∞) probability of entering state 1
is 0.57. Similarly, the long-run probability of entering state 2 is 0.43.
From this, we can make an important inference:
Since the transitions are assumed to take unit time, 57 % of the

time will be spent by the process in transitions to state 1 and
43 % of the time in transitions to state 2.
For the Markov process, the time taken in any transition is equal,
and hence the limiting probability of a state also denotes the propor-
tion of time spent in transitions to that particular state.
We will now show how we can obtain the limiting probabilities from
the TPM without raising the TPM to large powers. The following
important result provides a very convenient way for obtaining the
limiting probabilities.
Theorem 6.2 Let Π(i) denote the limiting probability of state i, and
let S denote the set of states in the Markov chain. Then the limiting
probabilities for all the states in the Markov chain can be obtained
from the transition probabilities by solving the following set of linear
equations:
|S|
Π(i)P (i, j) = Π(j), for every j ∈ S (6.4)
i=1
|S|
and Π(j) = 1, (6.5)
j=1
where |S| denotes the number of elements in the set S.
Equations (6.4) and (6.5) are often collectively called the invariance
equation, since they help us determine the invariant (limiting) proba-
bilities. Equation (6.4) is often expressed in the matrix form as:
[Π(1), Π(2), . . . , Π(|S|)]P = [Π(1), Π(2), . . . , Π(|S|)],
or in the following abbreviated form: ΠP = Π, where Π is a row vector

of the limiting probabilities.
Although we do not present its proof, the above is an important
result from many standpoints. If you use the equations above to find
the limiting probabilities of a Markov chain, you will notice that there
is one extra equation in the linear system of equations defined by the
theorem. You can eliminate any one equation from the system defined
by (6.4), and then solve the remaining equations to obtain a unique
solution. We demonstrate this idea with the TPM given in (6.3).
From equations defined by (6.4), we have:
For j = 1 : 0.7Π(1) + 0.4Π(2) + 0.6Π(3) = Π(1). (6.6)

For j = 2 : 0.2Π(1) + 0.2Π(2) + 0.1Π(3) = Π(2). (6.7)
For j = 3 : 0.1Π(1) + 0.4Π(2) + 0.3Π(3) = Π(3). (6.8)
With some transposition, we can re-write these equations as:
− 0.3Π(1) + 0.4Π(2) + 0.6Π(3) = 0. (6.9)

0.2Π(1) − 0.8Π(2) + 0.1Π(3) = 0. (6.10)
0.1Π(1) + 0.4Π(2) − 0.7Π(3) = 0. (6.11)
Now from Eq. (6.5), we have:
Π(1) + Π(2) + Π(3) = 1. (6.12)
Thus we have four Equations: (6.9)–(6.12) and three unknowns:

Π(1), Π(2), and Π(3). Notice that the system defined by the three
Eqs. (6.9)–(6.11) actually contains only two independent equations
because any one can be obtained from the knowledge of the other
two. Hence we select any two equations from this set. The two along
with Eq. (6.12) can be solved to find the unknowns. The values are:
Π(1) = 0.6265, Π(2) = 0.1807, and Π(3) = 0.1928.
Remark. The history-independent property of the Markov chain is
somewhat misleading. You can incorporate as much history as you
want into the state space, by augmenting the state space with histori-
cal information, to convert a history-dependent process into a Markov
process. Obviously, however, this comes with a downside in that the
size of the state space in the synthesized Markov process is much larger
than that in the history-dependent process. Usually, trying to incor-
porate all the relevant history to transform the stochastic process into
a Markov process can produce an unwieldy stochastic process that is
difficult to analyze.
3.1.3 Ergodic Markov Chains

A state in a Markov chain is said to be recurrent if it is visited
repeatedly (again and again). In other words, if one views a Markov
chain for an infinitely long period of time, one will see that a recurrent
state is visited infinitely many times. A transient state is one which
is visited only a finite number of times in such an “infinite viewing.”
An example of a transient state is one to which the system does not
come back from any recurrent state in one transition. In Fig. 6.7, 1 is
a transient state because once the system enters 2 or 3, it cannot come
back to 1.
There may be more than one transient state in a Markov chain. In
Fig. 6.8, 1a and 1b are transient states. If the system starts in any
one of these two states, it can visit both but once it goes to 2, it can
never come back to 1a or 1b. A state that is not transient is called
a recurrent state. Thus 2 and 3 are recurrent states in both Figs. 6.7
and 6.8. Another type of state is the absorbing state. Once the system
enters any absorbing state, it can never get out of that state and it
remains there.
An ergodic Markov chain is one in which all states are recurrent
and no absorbing states are present. Ergodic chains are also called
irreducible chains. All regular Markov chains are ergodic, but the
converse is not true. (Regular chains were defined in Sect. 3.1.1.) For
instance, a chain that is not regular may be ergodic. Consider the
Markov chain with the following TPM:
0 1
.
1 0
This chain is not regular, but ergodic. It is ergodic because both states
are visited infinitely many times in an infinite viewing.
1
2
Transient State
Recurrent State
Figure 6.7. A Markov chain with one transient state
We will now discuss the semi-Markov process.
3.2. Semi-Markov Processes

A stochastic process that spends a random amount of time (which
is not necessarily unity) in each transition, but is otherwise simi-
lar to a Markov process, is called a semi-Markov process. Conse-
quently, underlying a semi-Markov process, there lurks a Markov chain
1b
2
1a
Transient State
Recurrent State
Figure 6.8. A Markov chain with two transient states
called the embedded Markov chain. The main difference between the
semi-Markov process and the Markov process lies in the time taken in
transitions.
In general, when the distributions for the transition times are
arbitrary, the process goes by the name semi-Markov. If the time in
every transition is an exponentially distributed random variable, the
stochastic process is referred to as a continuous time Markov process.
Some authors refer to what we have called the continuous time
Markov process as the “Markov process,” and by a “Markov chain,”
they mean what we have referred to as the Markov process.
There is, however, a critical difference between the Markov chain
underlying a Markov process and that underlying a semi-Markov pro-
cess. In a semi-Markov process, the system jumps, but not necessarily
after unit time, and when it jumps, it jumps to a state that is different
than the current state. In other words, in a semi-Markov process, the
system cannot jump back to the current state. (However, in a semi-
Markov decision process, which we will discuss later, jumping back to
the current state is permitted.) In a Markov process, on the other
hand, the system can return to the current state after one jump.
If the time spent in the transitions is a deterministic quantity, the
semi-Markov process has a transition time matrix analogous to the
TPM, e.g.,
− 17.2
.
1 −
For an example of the most general model in which some or all of the
transition times are random variables from any given distributions,
consider the following transition time matrix:
− unif (5, 6)
,
expo(5) −
where unif (min, max) denotes a random number from the uniform
distribution with parameters, min and max, and expo(µ) denotes the
same from the exponential distribution with parameter µ.
When we analyze a semi-Markov process, we begin by analyzing
the Markov chain embedded in it. The next step usually is to analyze
the time spent in each transition. As we will see later, the semi-
Markov process is more powerful than the Markov process in modeling
real-life systems, although very often its analysis can prove to be more
complicated.
3.3. Markov Decision Problems

We will now discuss the topic that forms the central point of control
optimization in this book. Thus far, we have considered Markov chains
in which the transition from one state to another is governed by only
one transition law, which is contained in the elements of the TPM.
Such Markov chains are called uncontrolled Markov chains, essentially
because in such chains there is no external agency that can control the
path taken by the stochastic process.
We also have systems that can be run with different control mech-
anisms, where each control mechanism has its own TPM. In other
words, the routes dictated by the control mechanisms are not the
same. The control mechanism specifies the “action” to be selected
in each state. When we are faced with the decision of choosing from
more than one control mechanism, we have what is called a Markov
decision problem, which is often abbreviated as an MDP. The MDP
is also commonly called the Markov decision process.
The MDP is a problem of control optimization. In other words, it
is the problem of finding the optimal action to be selected in each
state. Many real-world problems can be set up as MDPs, and before
we discuss any details of the MDP framework, let us study a simple
example of an MDP.
Example. Consider a queuing system, such as the one you see in

a supermarket, with a maximum of three counters (servers). You
have probably noticed that when queues become long, more counters
(servers) are opened. The decision-making problem we will consider
here is to find the number of servers that should be open at any given
time.
The people who function as servers also have other jobs to perform,
and hence it does not make business sense to have them wait on the
counters when there are no customers at the counters. At the same
time, if very long queues build up but more counters are not opened
when there is capacity, customers do not feel very happy about it,
and may actually go elsewhere the next time. Hence in this situation,
one seeks an optimal strategy (i.e., a control mechanism) to control
the system. Next, we will discuss the idea of control mechanisms or
policies with some examples.
Consider a system which has a maximum of three counters. Let us
assume that the state of this system is defined by the number of people
waiting for service. Let us further assume that associated with this
state definition, a Markov process exists. (See Fig. 6.9 for a picture of
the underlying Markov chain.) Thus when the system enters a new
state of the Markov chain, one out of the following three actions can
be selected:
{Open 1 counter, Open 2 counters, and Open 3 counters.}
One possible control mechanism (or policy) in this situation would
look like this:
State = Number waiting for service Action

0 Open 1 counter
1 Open 1 counter
2 Open 1 counter
3 Open 1 counter
4 Open 2 counters
5 Open 2 counters
6 Open 2 counters
7 Open 3 counters
8 Open 3 counters
. Open 3 counters
. Open 3 counters
. Open 3 counters
0 1 2 3 .............
Figure 6.9. A Markov chain underlying a simple single-server queue
Another possible control mechanism could be:

State = Number waiting for service Action
0 Open 1 counter
1 Open 1 counter
2 Open 1 counter
3 Open 1 counter
4 Open 1 counter
5 Open 1 counter
6 Open 2 counters
7 Open 2 counters
8 Open 3 counters
. Open 3 counters
. Open 3 counters
. Open 3 counters
Note that the two control mechanisms are different. In the first,
two counters are opened in state 4, while in the second, the same is
done in state 6. From these two examples, it should be clear that there
are, in fact, several different control mechanisms that can be used, and
the effect on the system—in terms of the net profits generated—may
differ with the control mechanism used. This because:
A control mechanism that allows big queues to build up may lead
to a cost reduction in some sense, since fewer employees may be
necessary to run the system. But it may also lead to reduced profits,
because in the future customers may choose to go elsewhere where
queues are shorter.
On the other hand, a control mechanism which is over-designed
with a large number of servers (where customers hardly ever have
to wait) may be expensive to maintain, because of the costs incurred
in hiring a large number of servers. Eventually the high costs of
running this system will be transmitted to the customers through

higher-priced products. The latter is also likely to drive customers
away.
It naturally makes business sense to use the control mechanism

that produces the greatest net profits. Formulation of this problem
as a Markov decision problem will help us identify the best control
mechanism.
From our discussion above, we can conclude that each control
mechanism is likely to have unique costs and profits. Further, since
the system considered above is stochastic, associated with each con-
trol mechanism, a distinctive pattern of behavior is likely to emerge.
Finally, the problem is one of finding the right control mechanism.
The Markov decision framework is a sophisticated operations research
model designed to solve this problem. We now provide details.
3.3.1 Elements of an MDP

The Markov decision framework is designed to solve the so-called
Markov decision problem (MDP). The framework is made up of five
important elements. They are: (1) A decision maker, (2) policies,
(3) transition probability matrices, (4) transition reward matrices, and
(5) a performance metric (objective function).
Decision maker. The decision maker is an entity that selects the

control mechanism. It is also called the agent or controller.
Policies. The control mechanism is usually referred to as a policy.

A policy for an MDP with n states is an n-tuple. Each element of
this n-tuple specifies the action to be selected in the state associated
with that element. For example, consider a 2-state MDP in which two
actions are allowed in each state. An example of a policy for this MDP
is: (2, 1). This means that by adhering to this policy, the following
would occur: In state 1, action 2 would be selected, and in state 2,
action 1 would be selected. Thus in general, if µ̂ denotes a policy, the
ith element of µ̂, that is, µ(i), denotes the action selected in the ith
state for the policy µ̂.
In this book, unless otherwise stated, the word “policy” will imply a
stationary, deterministic policy. The word stationary means that
the policy does not change with time. This implies that if a policy
dictates an action a be taken in a state x, then no matter how long
the system has been operating, every time the system visits state x, it
is action a that will be selected.
The word deterministic implies that in any given state, we can

choose only one (1) action (out of the multiple actions allowed).
In other words, with a probability of 1, a given action is selected. We
will deal with stochastic policies in the context of learning automata
and actor critics in Chap. 8 for control optimization. However, in this
chapter, we will primarily consider stationary, deterministic policies.
We will assume throughout this book that the set of actions allowed
in each state is a finite set. The set of actions allowed in state i will be
denoted by A(i). We will also assume the set of states in the system
to be a finite set, which will be denoted by S.
Since the number of actions allowed in each state and the number
of states themselves are finite quantities, we must have a finite number
of policies. For instance in an MDP with two states and two actions
allowed in each state, we have 22 = 4 policies, which are:
(1, 1), (1, 2), (2, 1), and (2, 2).
An MDP, let us reiterate, revolves around finding the most suitable

policy.
A few more words about the term “state” are in order. We may
encounter systems in which decisions are not made in every state. In
other words, in some states, there is only one allowable action. As such,
there is no decision making involved in these states. These states are
called non-decision-making states. A state in which one has to choose
from more than one action is hence called a decision-making state. In
this book, by “state,” we refer to a decision-making state. When we
have models in which some states are not of the decision-making kind,
we will distinguish between the two by the qualifiers: decision-making
and non-decision-making.
Transition probability matrices. The decision maker executes
the action to be used in each state of an MDP. Associated with
each action, we usually have a transition probability matrix (TPM).
Associated with each policy, also, we have a unique TPM. The TPM
for a policy can be constructed from the TPMs associated with the
individual actions in the policy.
We illustrate this construction with a 2-state MDP that has 2
actions allowed in each state. Let the TPM associated with action a
be denoted by Pa . Let
0.7 0.3 0.1 0.9

P1 = , and P2 = .
0.4 0.6 0.8 0.2
Now consider a policy µ̂ = (2, 1). The TPM associated with this
policy will contain the transition probabilities of action 2 in state 1
and the transition probabilities of action 1 in state 2. The TPM of
policy µ̂ is thus
0.1 0.9
Pµ̂ = .
0.4 0.6
P1 P2
0.7 0.3 0.1 0.9
0.4 0.6 0.8 0.2
Matrix for action 1 Matrix for action 2
Pu
0.1 0.9
Matrix for
0.4 0.6
policy (2,1)
Construction of the TPM for policy (2,1)

from the TPMs of action 1 and action 2.
Figure 6.10. Schematic showing how the TPM of policy (2, 1) is constructed from
the TPMs of action 1 and 2
See Fig. 6.10 for a pictorial demonstration of the construction pro-

cess.
In general, we will use the following notation to denote a transition
probability:
p(i, a, j).
This term will denote the one-step transition probability of going from
state i to state j when action a is selected in state i. Now, if policy µ̂
is followed, then the action selected in state i will be denoted by µ(i),
and as a result the transition probability of going from state i to state
j will be denoted by
p(i, µ(i), j).
Then p(i, µ(i), j) will define the element in the ith row and the jth
column of the matrix Pµ̂ —the TPM associated with the policy µ̂.
In this section, we have used the phrase: “a transition probability
under the influence of an action.” The significance of this must
be noted with care. In our previous discussions, we spoke of transition

probabilities without reference to any action. That was because we
were dealing with uncontrolled Markov chains—which had only one
action in each state. Now, with the introduction of multiple actions in
each state, we must be careful with the phrase “transition probability,”
and must specify the action along with a transition probability.
Now, as stated previously, the MDP is all about identifying the
optimal policy (control mechanism). The TPM, we will see shortly,
will serve an important purpose in evaluating a policy and will be
essential in identifying the best policy. The other tool that we need
for evaluating a policy is discussed in the next paragraph.
Transition reward matrices. With each transition in a Markov
chain, we can associate a reward. (A negative value for the reward is
equivalent to a cost.) We will refer to this quantity as the immediate
reward or transition reward. The immediate reward helps us incor-
porate reward and cost elements into the MDP model. The immediate
reward matrix, generally called the transition reward matrix (TRM),
is very similar to the TPM. Recall that the (i, j)th element (the ele-
ment in the ith row and jth column) of the TPM denotes the transition
probability from state i to state j. Similarly, the (i, j)th element of the
TRM denotes the immediate reward earned in a transition from state
i to state j. Just as we have TPMs associated with individual actions
and policies, we have TRMs associated with actions and policies. Let
us examine some examples from a 2-state MDP, next.
Let Ra be the TRM associated with action a, and let:
11 −4 45 80
R1 = and R2 = .
−14 6 1 −23
Now consider a policy µ̂ = (2, 1). Like in the TPM case, the TRM
associated with this policy will contain the immediate reward of action
2 in state 1 and the immediate rewards of action 1 in state 2. Thus
the TRM of policy µ̂ can be written as
45 80
Rµ̂ = .
−14 6
The TPM and the TRM of a policy together contain all the informa-
tion one needs to evaluate the policy in an MDP. In terms of notation,
we will denote the immediate reward, earned in going from state i to
state j, under the influence of action a, by:
r(i, a, j).
When policy µ̂ is followed, the immediate reward earned in going from

state i to state j will be denoted by:
r(i, µ(i), j)
because µ(i) is the action that will be selected in state i when policy
µ̂ is used.
Performance metric. To compare policies, one must define a perfor-
mance metric (objective function). Naturally, the performance metric
should involve reward and cost elements. To give a simple analogy, in
a linear programming problem, one judges each solution on the basis
of the value of the associated objective function. Any optimization
problem has a performance metric, which is also called the objective
function. In this book, for the most part, the MDP will be studied
with respect to two performance metrics. They are:
1. Expected reward per unit time calculated over an infinitely long
trajectory of system states: We will refer to this metric as the
average reward.
2. Expected total discounted reward calculated over an infinitely long
trajectory of system states: We will refer to this metric as the
discounted reward.
It is the case that of the two performance metrics, average reward
is easier to understand, although the average reward MDP is more
difficult to analyze for its convergence properties. Hence, we will begin
our discussion with the average reward performance criterion. Dis-
counted reward will be defined later.
We first need to define the expected immediate reward of a state
under the influence of a given action. Consider the following scenario.
An action a is selected in state i. Under the influence of this action,
the system can jump to three states: 1, 2, and 3 with probabilities of
0.2, 0.3, and 0.5,
respectively. The immediate rewards earned in these three possible

transitions are, respectively,
10, 12, and − 14.
Then the expected immediate reward that will be earned, when

action a is selected in state a, will clearly be:
0.2(10) + 0.3(12) + 0.5(−14) = −1.4.

The expected immediate reward is calculated in the style shown

above. (Also see Fig. 6.11). In general, we can use the following exp-
ression to calculate the expected immediate reward.
r̄(i, a) = p(i, a, j)r(i, a, j). (6.13)

j∈S
(0.2,10)
1
(0.3,12)
i
2
(0.5,-14)
Legend: 3
(x,y)
x= transition probability
y=transition reward
Figure 6.11. Calculation of expected immediate reward
The average reward of a given policy is the expected reward earned

per unit time by running the Markov chain associated with the policy
for an infinitely long period of time. It turns out that there is a very
convenient way for representing the average reward of a policy, which
employs the limiting probabilities associated with the policy. We will
explain this with an example, and then generalize from there to obtain
a generic expression for the average reward.
Consider a 3-state MDP with states numbered 1, 2, and 3. Let us
assume that the system follows a fixed policy µ̂ and that the limiting
probabilities of the three states with respect to this policy are: Πµ̂ (1) =
0.3, Πµ̂ (2) = 0.5, and Πµ̂ (3) = 0.2. Let us further assume that the
expected immediate rewards earned in the three states are:
r̄(1, µ(1)) = 10, r̄(2, µ(2)) = 12, and r̄(3, µ(3)) = 14.
Now from our discussion on limiting probabilities, we know that the
limiting probability of a state denotes the proportion of time spent
in transitions to that particular state in the long run. Hence if we
observe the system over k transitions, kΠµ̂ (i) will equal the number
of transitions to state i in the long run. Now, the expected reward
earned in each visit to state i under policy µ̂ is r̄(i, µ(i)). Then the
total long-run expected reward earned in k transitions for this MDP
can be written as:
kΠµ̂ (1)r̄(1, µ(1)) + kΠµ̂ (2)r̄(2, µ(2)) + kΠµ̂ (3)r̄(3, µ(3)).
Consequently the average reward associated with policy µ̂ can be

written as:
kΠµ̂ (1)r̄(1, µ(1)) + kΠµ̂ (2)r̄(2, µ(2)) + kΠµ̂ (3)r̄(3, µ(3))
ρµ̂ =
k
3
= Πµ̂ (i)r̄(i, µ(i))
i=1
Then, in general, the average reward of a policy µ̂ can be written as:
ρµ̂ = Πµ̂ (i)r̄(i, µ(i)), where (6.14)

i∈S
Πµ̂ (i) denotes the limiting probability of state i when the system
(and hence the underlying Markov chain) is run with the policy µ̂
S denotes the set of states visited in the system
And r̄(i, a) denotes the expected immediate reward earned in the

state i when action a is selected in state i
In the next section, we will discuss a simple method to solve the

MDP. But before that we conclude this section by enumerating the as-
sumptions we will make about the control optimization problem we will
solve in this book.
Assumption 6.1 The state space S and the action space A(i) for
every i ∈ S is finite (although possibly quite large).
Assumption 6.2 The Markov chain associated with every policy in

the problem is regular.
Assumption 6.3 The immediate reward earned in any state transi-

tion under any action is finite, i.e., for all (i, a, j), |r(i, a, j)| < ∞.
3.3.2 Exhaustive Enumeration

The method that we will discuss in this section goes by the name
exhaustive enumeration or exhaustive evaluation. Conceptually, this
is the easiest method to understand, although in practice we can use
it only on small problems. The method is based on the following
idea: Enumerate every policy that can possibly be selected, evaluate
the performance metric associated with each policy, and then declare
the policy that produces the best value for the performance metric to
be the optimal policy. We now explain this method with a simple
example of an MDP that has just two states and two actions in each
state. This example will be used repeatedly throughout the remainder
of this book.
Example A. There are two states numbered 1 and 2 in an MDP, and
two actions, which are also numbered 1 and 2, are allowed in each state.
The transition probability matrices (TPM) associated with actions 1
and 2 are:
0.7 0.3 0.9 0.1

P1 = and P2 = .
0.4 0.6 0.2 0.8
The TRM for actions 1 and 2 are:
6 −5 10 17
R1 = and R2 = .
7 12 −14 13
Pictorially, the MDP is represented in Fig. 6.12.

In this MDP, there are four possible policies that can be used to
control the system. They are:
µ̂1 = (1, 1), µ̂2 = (1, 2), µ̂3 = (2, 1), and µ̂4 = (2, 2).
The TPMs and TRMs of these policies are constructed from the
individual TPMs and TRMs of each action. The TPMs are:
0.7 0.3 0.7 0.3

Pµ̂1 = ; Pµ̂2 = ;
0.4 0.6 0.2 0.8
0.9 0.1 0.9 0.1

Pµ̂3 = ; Pµ̂4 = .
0.4 0.6 0.2 0.8
The TRMs are

6 −5 6 −5
Rµ̂1 = ; Rµ̂2 = ,
7 12 −14 13
10 17 10 17
Rµ̂3 = ; Rµ̂4 = .
7 12 − 14 13
(1,0.7,6) (2,0.8,13)
(1,0.3,-5)
(2,0.1,17)
1 2
(1,0.4,7)
(2,0.2,-14)
(2,0.9,10)
(1,0.6,12)
Legend:
(a,p,r): a = action
p = transition
probability
r = immediate
reward
Figure 6.12. A two-state MDP
From the TPMs, using Eqs. (6.4) and (6.5), one can find the limiting
probabilities of the states associated with each policy. They are:
Πµ̂1 (1) = 0.5714 and Πµ̂1 (2) = 0.4286;
Πµ̂2 (1) = 0.4000 and Πµ̂2 (2) = 0.6000;
Πµ̂3 (1) = 0.8000 and Πµ̂3 (2) = 0.2000;
Πµ̂4 (1) = 0.6667 and Πµ̂4 (2) = 0.3333.

We will next find the average reward of each of these four policies.
We first evaluate the average immediate reward in each possible tran-
sition in the MDP, using Eq. (6.13). For this, we need the TPMs and
the TRMs of each policy.
r̄(1, µ1 (1)) = p(1, µ1 (1), 1)r(1, µ1 (1), 1) + p(1, µ1 (1), 2)r(1, µ1 (1), 2)
= 0.7(6) + 0.3(−5) = 2.7.
r̄(2, µ1 (2)) = p(2, µ1 (2), 1)r(2, µ1 (2), 1) + p(2, µ1 (2), 2)r(2, µ1 (2), 2)
= 0.4(7) + 0.6(12) = 10.
r̄(1, µ2 (1)) = p(1, µ2 (1), 1)r(1, µ2 (1), 1) + p(1, µ2 (1), 2)r(1, µ2 (1), 2)
= 0.7(6) + 0.3(−5) = 2.7.
r̄(2, µ2 (2)) = p(2, µ2 (2), 1)r(2, µ2 (2), 1) + p(2, µ2 (2), 2)r(2, µ2 (2), 2)
= 0.2(−14) + 0.8(13) = 7.6.
r̄(1, µ3 (1)) = p(1, µ3 (1), 1)r(1, µ3 (1), 1) + p(1, µ3 (1), 2)r(1, µ3 (1), 2)
= 0.9(10) + 0.1(17) = 10.7.
r̄(2, µ3 (2)) = p(2, µ3 (2), 1)r(2, µ3 (2), 1) + p(2, µ3 (2), 2)r(2, µ3 (2), 2)
= 0.4(7) + 0.6(12) = 10.
r̄(1, µ4 (1)) = p(1, µ4 (1), 1)r(1, µ4 (1), 1) + p(1, µ4 (1), 2)r(1, µ4 (1), 2)
= 0.9(10) + 0.1(17) = 10.7.
r̄(2, µ4 (2)) = p(2, µ4 (2), 1)r(2, µ4 (2), 1) + p(2, µ4 (2), 2)r(2, µ4 (2), 2)
= 0.2(−14) + 0.8(13) = 7.6.
Now, using these quantities, we can now calculate the average

reward of each individual policy. We will make use of Eq. (6.14).
Thus:
ρµ̂1 = Πµ̂1 (1)r̄(1, µ1 (1)) + Πµ̂1 (2)r̄(2, µ1 (2))
= 0.5741(2.7) + 0.4286(10) = 5.83,
ρµ̂2 = Πµ̂2 (1)r̄(1, µ2 (1)) + Πµ̂2 (2)r̄(2, µ2 (2))

= 0.4(2.7) + 0.6(7.6) = 5.64,
ρµ̂3 = Πµ̂3 (1)r̄(1, µ3 (1)) + Πµ̂3 (2)r̄(2, µ3 (2))
= 0.8(10.7) + 0.2(10) = 10.56, and

ρµ̂4 = Πµ̂4 (1)r̄(1, µ4 (1)) + Πµ̂4 (2)r̄(2, µ4 (2))
= 0.6667(10.7) + 0.3333(7.6) = 9.6667.
It is clear that policy µ̂3 = (2, 1) is the best policy, since it produces
the maximum average reward.
Drawbacks of exhaustive enumeration. Clearly, exhaustive enu-
meration can only be used on small problems. Consider a problem
with 10 states and 2 actions in each state. On this, if we were to use
exhaustive enumeration, we would have to evaluate 210 different poli-
cies. As such the computational burden of this method can quickly
overwhelm it. We now turn to dynamic programming, which has a
lower computational burden.
4. Average Reward MDPs and DP

The method of dynamic programming (DP), in the context of
solving MDPs, was developed in the late 1950s with the pioneering
work of Bellman [24] and Howard [144]. Dynamic programming has
a considerably lower computational burden in comparison to exhaus-
tive enumeration. The theory of DP has evolved significantly since
the 1950s, and a voluminous amount of literature now exists on this
topic. DP continues to be a main pillar of control optimization in
discrete-event systems. Although the theory of DP is mathematically
sophisticated, the main algorithms rest on simple systems of equations.
As such, it is very easy to understand its basic principles, which form
the main focus of the remainder of this chapter.
In this chapter and in the next, we will endeavor to present the
main equations underlying DP without worrying about how the equa-
tions were derived. Also, we will present the main algorithms but will
not attempt to prove that they generate optimal solutions (Chap. 11

presents details of that nature). Links to computer codes can be found
at [121].
The system of equations, to which we were referring above, is often
called the Bellman equation. The Bellman equation has several forms.
In this section, we will concentrate on the average reward MDP, and
in the next, on discounted reward. For both average and discounted
reward, there are two different forms of the Bellman equation, based on
which we have two different DP methods that go by the following
names:
Policy iteration, which uses the Bellman policy equation or the
Poisson equation
Value iteration, which uses the Bellman optimality equation
4.1. Bellman Policy Equation

We have seen above that associated with every given policy, there
exists a scalar called the average reward of the policy. Similarly, associ-
ated with every policy, there exists a vector called the value function
vector for the policy. The dimension of this vector is equal to the num-
ber of elements in S, the set of decision-making states. The values of
this vector’s components can be determined by solving a linear system
of equations, which is collectively called the Bellman equation for a
policy or the Bellman policy equation. This equation, as stated
above, is also sometimes called the Poisson equation.
A natural question at this stage is: what purpose will be served by
finding the values of the elements of this vector? Much of DP revolves
around this vector. The answer is that the values of these elements
can actually help us find a better policy. The Bellman policy equation
in the average reward context is:
|S|
hµ̂ (i) = r̄(i, µ(i)) − ρµ̂ + p(i, µ(i), j)hµ̂ (j) for each i ∈ S. (6.15)
j=1
The above is a system of linear equations in which the number of

equations equals the number of elements in the set S, i.e., |S|. The
unknowns in the equations are the hµ̂ terms. They are the elements
of the value function vector associated with the policy µ̂. The other
terms are defined below:
µ(i) denotes the action selected in state i under the policy µ̂. Since
the policy is known, each µ(i) is known.
r̄(i, µ(i)) denotes the expected immediate reward in state i under

policy µ̂. Each of these terms can be calculated from the TPMs
and the TRMs.
p(i, µ(i), j) denotes the one-step transition probability of jumping
from state i to state j under the policy µ̂. Again, these terms can
be obtained from the TPMs.
ρµ̂ denotes the average reward associated with the policy µ̂, and it
can be obtained, when the policy is known, from the TPMs and
the TRMs as discussed in the context of exhaustive enumeration.
We will now discuss the celebrated policy iteration algorithm [144]
to solve the average reward MDP.
4.2. Policy Iteration

The basic idea underlying policy iteration is to start with an
arbitrarily selected policy, and then switch to a better policy in every
iteration. This continues until no further improvement is possible.
The advantage of policy iteration over exhaustive enumeration is that
the solution is usually obtained using comparatively fewer iterations.
When a policy is selected, the Bellman policy equation is used to
obtain the value function vector for that policy. This is called the policy
evaluation stage, since in this stage we evaluate the value function
vector associated with a policy. Then the value function vector is used
to find a better policy. This step is called the policy improvement step.
The value function vector of the new (better) policy is then obtained.
In other words, one returns to the policy evaluation step. This contin-
ues until the value function vector obtained from solving the Bellman
policy equation cannot produce a better policy. (If you are familiar
with the simplex algorithm of linear programming, you will see an
analogy. In simplex, we start at a solution (corner point) and then
move to a better point. This continues until no better point can be
obtained.)
4.2.1 Steps in the Algorithm

Step 1. Set k = 1. Here k will denote the iteration number. Let the
set of states be S. Select any policy in an arbitrary manner. Let us
denote the policy selected in the kth iteration by µ̂k . Let µ̂∗ denote
the optimal policy.
Step 2. (Policy Evaluation) Solve the following linear system of

equations. For i = 1, 2, . . . , |S|,
|S|
k k
h (i) = r̄(i, µk (i)) − ρ + p(i, µk (i), j)hk (j). (6.16)
j=1
Here one linear equation is associated with each value of i ∈ S. In

this system, the unknowns are the hk terms and ρk . The number
of unknowns exceeds the number of equations by 1. Hence to solve
the system, one should set any one of the hk terms to 0, and then
solve for the rest of the unknowns. The term ρk should not be set
to 0.
Step 3. (Policy Improvement) Choose a new policy µ̂k+1 such that
 
|S|
µk+1 (i) ∈ arg max r̄(i, a) + p(i, a, j)hk (j) for all i ∈ S.
a∈A(i) j=1
If possible, one should set µk+1 (i) = µk (i) for each i. The signifi-
cance of ∈ in the above needs to be understood clearly. There may
be more than one action that satisfies the argmax operator. Thus
there may be multiple candidates for µk+1 (i). However, the latter
is selected in a way such that µk+1 (i) = µk (i) if possible.
Step 4. If the new policy is identical to the old one, that is, if
µk+1 (i) = µk (i) for each i, then stop and set µ∗ (i) = µk (i) for
every i. Otherwise, increment k by 1, and go back to the second
step.
Policy iteration on Example A. We used policy iteration on

Example A from Sect. 3.3.2. The results are shown in Table 6.1. The
optimal policy is (2, 1). See also the case study on a preventive main-
tenance problem at the end of the chapter. We will next discuss an
alternative method, called value iteration, to solve the average reward
MDP.
Table 6.1. Calculations in policy iteration for average reward MDPs on Example A
Iteration (k) Policy selected Values ρk

1 µ̂1 = (1, 1) h1 (1) = 0 5.83
h1 (2) = 10.43
2 µ̂2 = (2, 1) h2 (1) = 0 10.56
h2 (2) = −1.4
4.3. Value Iteration and Its Variants

The value iteration algorithm is another very useful algorithm that
can solve the MDP. It uses a form of the Bellman equation that is
different than the form used in policy iteration. The idea of value
iteration is due to Bellman [24].
Another major point of difference between value and policy iteration
is that in value iteration, one does not solve any equations. This
makes it easy to write computer program for the algorithm, which has
contributed to its popularity. However, another advantage of value
iteration will be realized in reinforcement learning, the main topic of
Chap. 7.
We begin by presenting the Bellman optimality equation. A variant
of this will be used in the value iteration and the relative value iteration
algorithms.
 
|S|
∗ ∗
J (i) = max r̄(i, a) − ρ + p(i, a, j)J ∗ (j) for each i ∈ S,
a∈A(i)
j=1
(6.17)
where A(i) denotes the set of actions allowed in state i and ρ∗ denotes
the average reward of the optimal policy. The J ∗ terms are the
unknowns and are the components of the optimal value function vec-
tor J ∗ . The number of elements in the vector J ∗ equals the number
of states in the MDP.
Equation (6.17) is the famous Bellman optimality equation for
average reward MDPs. Since the equation contains the max operator,
it cannot be solved using linear algebra techniques such as Gauss elim-
ination. It is considered to be non-linear because of the max operator.
4.3.1 Value Iteration: A Natural Version

Value iteration in its most natural form seeks to use the Bellman
optimality equation in which ρ∗ is forcibly set to zero. Note that ρ∗ is
unknown at the start. We will start with some arbitrary values for the
value function vector. And then we will apply a transformation, i.e., an
updating mechanism, derived from the Bellman optimality equation,
on the vector repeatedly. This mechanism keeps updating (chang-
ing) the elements of the value function vector. The values themselves
will not converge in this algorithm, i.e., the vectors generated in suc-
cessive iterations will not approach each other. Therefore, we need
some mechanism to terminate the algorithm. A suitable termination
mechanism here employs the span seminorm, also called the span.
We will denote the span seminorm of a vector in this book by sp(.)
and define it as:
sp(x) = max x(i) − min x(i).

i i
The span of a vector is by definition non-negative (take simple

examples of vectors to verify this), and it denotes the range of values
in the different components of the vector. It is important to inform the
reader that a vector may keep changing in every iteration although its
span may not change! This is because the span measures the range,
and a changing vector may have a constant range. When the span
stops changing, we will terminate our algorithm. We will establish
later in Chap. 11 that the span actually converges to a finite num-
ber under certain conditions. We now present steps in natural value
iteration.
Step 1: Set k = 1. Select any arbitrary values for the elements of a
vector of size |S|, and call the vector J 1 . Specify > 0. Typically,
is set to a small value such as 0.01.
Step 2: For each i ∈ S, compute:
 
|S|
J k+1 (i) = max r̄(i, a) + p(i, a, j)J k (j) .
a∈A(i)
j=1
Step 3: If
sp(J k+1 − J k ) < ,
go to Step 4. Otherwise increase k by 1, and go back to Step 2.
Step 4: For each i ∈ S, choose
 
|S|
d(i) ∈ arg max r̄(i, a) + p(i, a, j)J k (j) ,
a∈A(i) j=1
ˆ
and stop. The -optimal policy is d.
The implication of -optimality (in Step 4 above) needs to be
understood. The smaller the value of , the closer we get to the
optimal policy. Usually, for small values of , one obtains policies
very close to optimal. The span of the difference vector (J k+1 − J k )
keeps getting smaller and smaller in every iteration, and hence for a
given positive value of , the algorithm terminates in a finite number
of steps. Also note that in Step 2, we have used a transformation (see

Chap. 1 for a definition) derived from the Bellman equation. Recall
that a transformation takes a vector and produces a new vector from
the vector supplied to it.
Do note that in Step 2 had we used the value of ρ∗ (assuming that
the value of ρ∗ is known) instead of replacing it by zero, we would have
obtained
The same value for the span in Step 3 and
The same sequence of maximizing actions in Step 2.
Together these two facts suggest that value iteration discussed above
should yield the results obtained from using the Bellman transforma-
tion in Step 2. We will show in Chap. 11 that the Bellman transfor-
mation leads one to an optimal solution, and hence, it can be inferred
that the value iteration algorithm above should also lead us to the
optimal policy.
And yet a major difficulty with value iteration presented above is
that the one or more of the values can become very large or very small
during the iterations. This difficulty can be overcome by using relative
value iteration, which we discuss in the next subsection. Before that,
we present some numerical results to show the difficulty encountered
with the value iteration algorithm above.
Value Iteration on Example A. See Table 6.2 for some sample
calculations with value iteration on Example A (from Sect. 3.3.2). Note
that the values gradually become larger with every iteration.
4.3.2 Relative Value Iteration

As discussed above, the major obstacle faced in using value iteration
is that some of the values can become very large or very small. On
problems with a small number of states with a value for that is not
very small, one can sometimes use value iteration without running into
this difficulty. However, usually, in larger problems, one of the values
becomes very large or very small, and then the natural form of value
iteration cannot be used. Relative value iteration (RVI), due to White
[320], provides us with an elegant way out of this difficulty that we now
discuss.
If the Markov chain of every policy is regular, which we have
assumed above, we will select any state in S and call it the dis-
tinguished state. The updating mechanism (see Step 3 below) will
use this distinguished state and keep the values from becoming too
large or too small. Recall that in the natural form of value iteration,
the corresponding updating mechanism (see Step 2 of natural value

iteration) is derived from the Bellman equation by setting ρ∗ = 0. We
now present steps in RVI.
Step 1: Select any arbitrary state from S to be the distinguished state
i∗ . Set k = 1, and select arbitrary values for the vector J 1 . Specify
the termination value > 0.
Step 2: Compute for each i ∈ S:
 
|S|
J k+1 (i) = max r̄(i, a) + p(i, a, j)J k (j) .
a∈A(i)
j=1
After calculations in this step for all states are complete, set ρ =
J k+1 (i∗ ).
Table 6.2. Calculations in value iteration for average reward MDPs: Note that the
values get unbounded but the span of the difference vector gets smaller with every
iteration. We start with J 1 (1) = J 1 (2) = 0
k J k (1) J k (2) Span

2 10.7 10.0 0.7
3 21.33 20.28 0.35
4 31.925 30.70 0.175
. ... ... ...
. ... ... ...
. ... ... ...
153 1,605.40 1,604.00 0.001367
154 1,615.96 1,614.56 0.000684

J k+1 (i) ← J k+1 (i) − ρ.
Step 4: If
sp(J k+1 − J k ) < ,
go to Step 5. Otherwise increase k by 1 and go back to Step 2.
 
|S|
d(i) ∈ arg max r̄(i, a) + p(i, a, j)J k (j) ,
a∈A(i)
j=1
ˆ ρ is the optimal reward’s

and stop. The -optimal policy is d;
estimate.
Essentially, the only difference between relative value iteration and
regular value iteration is in the subtraction of a value. And yet, what a
difference a subtraction makes! What is remarkable is that the values
remain bounded (see Table 6.3). We would also like to state without
proof here that:
Remark 1. The algorithm converges to the policy that value iteration
would have converged to.
Remark 2. As k tends to ∞, ρ converges to ρ∗ .
Remark 2 implies that when the algorithm converges, the values will
converge to a solution of the Bellman optimality equation. Once J(i∗ )
converges to ρ∗ , it is clear from Step 3 that one would in fact be using
a transformation based on the Bellman optimality equation. We will
formally show in Chap. 11 that under some conditions, relative value
iteration converges in the span to the optimal solution, i.e., a solution
of the Bellman optimality equation.
Example A and RVI. In Table 6.3, we show the calculations with
using relative value iteration on Example A from Sect. 3.3.2.
4.3.3 Average Reward: A General Expression

It is good to remind the reader here that we have assumed the time
spent in a transition of a Markov process to be unity although the
actual time spent may not be unity. This assumption is perfectly valid
for the MDP where the transition time is not required to measure the
performance metric. But if the time spent in each transition is used to
measure the performance metric, which is the case in the semi-Markov
decision process (SMDP), then this assumption is invalid.
We now present a general expression for the average reward per unit
time in an MDP. Thus far, the expression used for average reward
has used limiting probabilities (assuming regular Markov chains for
all policies). The following expression, although equivalent, will not
contain the limiting probabilities. If the time spent in a transition is
not unity, one can view the expression below as that for the average
reward per jump or transition. If the transition time is unity, per unit
time and per jump are clearly equivalent.
The average reward per unit time calculated over an infinite time
horizon, starting at state i and using a policy µ̂, can be expressed as:
k
E s=1 r(xs , µ(xs ), xs+1 )|x1 =i
ρ(i) = lim inf , where
k→∞ k
k denotes the number of transitions (or time assuming that each tran-
sition takes unit time) over which the system is observed, xs denotes
the state from where the sth jump or state transition occurs under the
policy µ̂, and E denotes the expectation operator over all trajectories
that start under the condition specified within the square brackets.
If you have trouble understanding why we use lim inf here, you may
replace it by lim at this stage.
It can be shown that for policies with regular Markov chains, the
average reward is independent of the starting state i, and hence, ρ(i)
can be replaced by ρ. Intuitively, the above expression says that the
average reward for a given policy is
the expected sum of rewards earned in a very long trajectory
.
the number of transitions in the same trajectory
In the above, we assume that the associated policy is pursued within
the trajectory. We now discuss the other important performance
metric typically studied with an MDP: the discounted reward.
5. Discounted Reward MDPs and DP

The discounted reward criterion is another popular performance
metric that has been studied extensively by researchers in the con-
text of MDPs. In this section, we will focus on the use of DP methods
Table 6.3. Calculations in Relative value iteration for average reward MDPs: =
0.001; -optimal policy found at k = 12; J 1 (1) = J 1 (2) = 0
Iteration (k) J k (1) J k (2) Span ρ

2 0 −0.7 0.7 10.760
3 0 −1.05 0.35 10.630
4 0 −1.225 0.175 10.595
5 0 −1.3125 0.0875 10.578
6 0 −1.35625 0.04375 10.568
7 0 −1.378125 0.021875 10.564
8 0 −1.389063 0.010938 10.562
9 0 −1.394531 0.005469 10.561
10 0 −1.397266 0.002734 10.561
11 0 −1.398633 0.001367 10.560
12 0 −1.399316 0.000684 10.560
to find the policy in an MDP that optimizes discounted reward. Links

to computer codes for some DP algorithms are presented at [121].
The idea of discounting is related to the fact that the value of money
reduces with time. To give a simple example: a dollar tomorrow is
worth less than a dollar today. The discounting factor is the fraction
by which money gets devalued in unit time. So for instance, if I earn $3
today, $5 tomorrow, $6 the day after tomorrow, and if the discounting
factor is 0.9 per day, then the present worth of my earnings will be:
3 + (0.9)5 + (0.9)2 6.
The reason for raising 0.9 to the power of 2 is that tomorrow, the
present worth of day-after-tomorrow’s earning will be 0.9(6). Hence
today, the present worth of this amount will be 0.9[0.9(6)] = (0.9)2 6.
In general, if the discounting factor is λ, and if e(t) denotes the
earning in the tth period of time, then the present worth of earnings
over n periods of time can be denoted by
e(0) + λe(1) + λ2 e(2) + · · · + λn e(n). (6.18)
Note that the discounting factor is in fact defined as follows:

1
λ= ,
1+γ
where γ is the rate of interest (see any standard text on engineering
economics) expressed as a ratio. Since 0 < γ < 1, we have that λ < 1.
We now define the discounted reward.
5.1. Discounted Reward

By the discounted reward of a policy, we mean the expected total
discounted reward earned in the Markov chain associated with the
policy, when the policy is pursued over an infinitely long trajectory
of the Markov chain. A technical definition for the discounted reward
earned with a given policy µ̂ starting at state i is
k
vµ̂ (i) = lim E λs−1 r(xs , µ(xs ), xs+1 )|x1 = i , where (6.19)
k→∞
s=1
λ denotes the discounting factor, k denotes the number of transitions

(or time assuming that each transition takes unit time) over which the
system is observed, xs denotes the state from where the sth jump or
transition of system state occurs under the policy µ, and E denotes
the expectation operator over all trajectories that operate under the
condition specified in the square brackets.
In Eq. (6.19),
k
E λs−1 r(xs , µ(xs ), xs+1 )|x1 = i =
s=1
E r(i, µ(i), x2 ) + λr(x2 , µ(x2 ), x3 ) + · · · + λk−1 r(xk , µ(xk ), xk+1 ) .

This should make it obvious that the discounted reward of a policy is
measured using the format discussed in Eq. (6.18).
5.2. Discounted Reward MDPs

In the discounted reward MDP, we have more than one perfor-
mance metric (objective function), and hence the performance metric
is expressed in the form of a vector. Remember that in the average
reward case, the performance metric is the unique scalar quantity (ρ):
the average reward of a policy. In the discounted problem, associated
with every state, there exists a scalar quantity; we strive to max-
imize each of these quantities. These quantities are the elements of
the so-called value function vector. There is no quantity like ρ in the
discounted problem, rather there are as many quantities as there are
states.
The value function for discounted reward was defined in Eq. (6.19).
The term vµ̂ (i) in Eq. (6.19) denotes the ith element of the value func-
tion vector. The number of elements in the value function vector is
equal to the number of states. The definition in Eq. (6.19) is associ-
ated with a policy µ̂. The point to be noted is that the value function
depends on the policy. (Of course it also depends on the discounting
factor, the transition probabilities, and the transition rewards.)
Solving an MDP implies identifying the policy that returns a value
function vector v ∗ such that
v ∗ (i) = max vµ̂ (i) for each i ∈ S.
µ̂
The above means that the optimal policy will have a value function
vector that satisfies the following property: each element of the vector
is greater than or equal to the corresponding element of the value
function vector of any other policy. This concept is best explained
with an example.
Consider a 2-state Markov chain with 4 allowable policies denoted
by µ̂1 , µ̂2 , µ̂3 , and µ̂4 . Let the value function vector be defined by
vµ̂1 (1) = 3; vµ̂2 (1) = 8; vµ̂3 (1) = −4; vµ̂4 (1) = 12;
vµ̂1 (2) = 7; vµ̂2 (2) = 15; vµ̂3 (2) = 1; vµ̂4 (2) = 42;
Now, from our definition of an optimal policy, policy µ̂4 should be the
optimal policy since the value function vector assumes the maximum
value for this policy for each state. Now, the following question should
rise in your mind at this stage. What if there is no policy for which
the value function is maximized for each state? For instance consider
the following scenario:
vµ̂1 (1) = 3; vµ̂2 (1) = 8; vµ̂3 (1) = −4; vµ̂4 (1) = 12;
vµ̂1 (2) = 7; vµ̂2 (2) = 15; vµ̂3 (2) = 1; vµ̂4 (2) = −5;
In the above setting, there is no one policy for which the value function
is maximized for all the states. Fortunately, it has been proved that
under the assumptions we have made above, there exists an optimal
policy; in other words, there exists a policy for which the value function
is maximized in each state. The interested reader is referred to [30,
270], among other sources, for the proof of this.
The important point that we need to address next is: how does
one find the value function of any given policy? Equation (6.19) does
not provide us with any direct mechanism for this purpose. Like in
the average reward case, we will need to turn to the Bellman policy
equation.
5.3. Bellman Policy Equation

The Bellman policy equation is a system of linear equations in which
the unknowns are the elements of the value function associated with
the policy. The Bellman policy equation (or the Bellman equation for
a given policy) for the discounted reward MDP is
|S|
hµ̂ (i) = r̄(i, µ(i)) + λ p(i, µ(i), j)hµ̂ (j) for each i ∈ S. (6.20)
j=1
The number of equations in this system is equal to |S|, the number

of elements in the set S. The unknowns in the equation are the hµ̂
terms. They are the elements of the value function vector associated
with the policy µ̂. The other terms are defined below.
µ(i) denotes the action selected in state i under the policy µ̂ (since
the policy is known, each µ(i) is also known).
r̄(i, µ(i)) denotes the expected immediate reward in state i under
policy µ(i).
p(i, µ(i), j) denotes the one-step transition probability of jumping
from state i to state j under the policy µ̂.
By solving the Bellman equation, one can obtain the value func-
tion vector associated with a given policy. Clearly, the value function
vectors associated with each policy can be evaluated by solving the
respective Bellman equations. Then, from the value function vectors
obtained, it is possible to determine the optimal policy. This method
is called the method of exhaustive enumeration.
Like in the average reward case, the method of exhaustive enu-
meration is not a very efficient method to solve the MDP, since its
computational burden is enormous. For a problem of 10 states with
two allowable actions in each, one would need to evaluate 210 policies.
The method of policy iteration is considerably more efficient.
5.4. Policy Iteration

Like in the average reward case, the basic principle of policy iter-
ation is to start from any arbitrarily selected policy and then move
to a better policy in every iteration. This continues until no further
improvement in policy is possible.
When a policy is selected, the Bellman equation for a policy
(Eq. (6.20)) is used to find the value function vector for that policy.
This is known as the policy evaluation stage because in this stage
we evaluate the value function vector associated with a policy. Then
the value function vector is used to find a better policy. This step
is called the policy improvement step. The value function vector of
the new policy is then found. In other words, one returns to the
policy evaluation step. This continues until the value function vector
obtained from solving the Bellman equation cannot produce a better
policy.
Steps in policy iteration for MDPs.
number of states be |S|. Select any policy in an arbitrary manner.
Let us denote the policy selected in the kth iteration by µ̂k . Let µ̂∗
denote the optimal policy.

equations for the unknown hk terms. For i = 1, 2, . . . , |S|
|S|
k
h (i) = r̄(i, µk (i)) + λ p(i, µk (i), j)hk (j). (6.21)
j=1

 
|S|
µk+1 (i) ∈ arg max r̄(i, a) + λ p(i, a, j)hk (j) for all i ∈ S.
a∈A(i) j=1
If possible, one should set µ̂k+1 = µ̂k .

Step 4. If the new policy is identical to the old one, i.e., if µk+1 (i) =
µk (i) for each i, then stop and set µ∗ (i) = µk (i) for every i. Oth-
erwise, increment k by 1, and return to Step 2.
Policy iteration on Example A. We used policy iteration on

Example A from Sect. 3.3.2. The results are shown in Table 6.4. The
optimal policy is (2, 1).
Table 6.4. Calculations in policy iteration for discounted MDPs
k Policy selected (µ̂k ) Values

1 (1, 1) h1 (1) = 25.026
h1 (2) = 34.631
2 (2, 1) h2 (1) = 53.03
h2 (2) = 51.87
Like in the average reward case, we will next discuss the value
iteration method. The value iteration method is also called the method
of successive approximations (in the discounted reward case). This is
because the successive application of the Bellman operator in the dis-
counted case does lead one to the optimal value function. Recall that
in the average reward case, the value iteration operator may not keep
the iterates bounded. Fortunately this is not the case in the discounted
problem.
5.5. Value Iteration

Like in the average reward case, the major difference between value
iteration and policy iteration is that unlike in policy iteration, in value
iteration, one does not have to solve any equations. This is a big plus—
not so much in solving small MDPs, but in the context of simulation-
based DP (reinforcement learning). We will discuss this issue in more
detail in Chap. 7.
We begin by presenting the Bellman optimality equation for
discounted reward.
 
|S|
J ∗ (i) = max r̄(i, a) + λ p(i, a, j)J ∗ (j) for each i ∈ S. (6.22)
a∈A(i)
j=1
The notation is similar to that defined for the average reward case.
Equation (6.22), i.e., the Bellman optimality equation for discounted
reward contains the max operator; hence, it cannot be solved using
linear algebra techniques, e.g., Gaussian elimination. However, the
value iteration method forms a convenient solution method. In value
iteration, one starts with some arbitrary values for the value function
vector. Then a transformation, derived from the Bellman optimality
equation, is applied on the vector successively until the vector starts
approaching a fixed value. The fixed value is also called a fixed point.
We will discuss issues such as convergence to fixed points in Chap. 11
in a more mathematically rigorous framework. However, at this stage,
it is important to get an intuitive feel for a fixed point.
If a transformation has a unique fixed point, then no matter what
vector you start with, if you keep applying the transformation repeat-
edly, you will eventually reach the fixed point. Several operations
research algorithms are based on such transformations.
We will now present step-by-step details of the value iteration algo-
rithm. In Step 3, we will need to calculate the max norm of a vector.
See Chap. 1 for a definition of max norm. We will use the notation ||.||
to denote the max norm.
Steps in value iteration for MDPs.
Step 1: Set k = 1. Select arbitrary values for the elements of a vector
of size |S|, and call the vector J 1 . Specify > 0.
 
|S|
J k+1 (i) ← max r̄(i, a) + λ p(i, a, j)J k (j) .
a∈A(i)
j=1
Step 3: If
||(J k+1 − J k )|| < (1 − λ)/2λ,
go to Step 4. Otherwise increase k by 1 and go back to Step 2.
Step 4: For each i ∈ S choose
 
|S|
d(i) ∈ arg max r̄(i, a) + λ p(i, a, j)J k (j) and stop.
a∈A(i) j=1
The reader should note that in Step 2, we have used a transformation

derived from the Bellman equation. Also, to be noted is the fact that
the max-norm of the difference vector (J k+1 − J k ) decreases with
every iteration. The reason for the use of the expression (1 − λ)/2λ
in Step 3 will be explained in more detail in Chap. 11. Essentially, the
condition in Step 3 ensures that when the algorithm terminates, the
max norm of the difference between the value function vector returned
by the algorithm and the optimal value function vector is . Since the
max norm is in some sense the “length” of the vector, the implication
here is that the vector generated by the algorithm is different from
the optimal value function vector by a length of . Thus by making
small, one can obtain a vector that is as close to the optimal vector as
one desires. Finally, unlike the average reward case, the values do not
become unbounded in this algorithm.
Example A and value iteration. Table 6.5 shows the calculations
with using value iteration on Example A from the Sect. 3.3.2. The
discounting factor is 0.8.
5.6. Gauss-Siedel Value Iteration

The regular value iteration algorithm discussed above is slow and
can take many iterations to converge. We now discuss its variant
thatmay work faster: the Gauss-Siedel version. It differs from the
Table 6.5. Calculations in value iteration for discounted reward MDPs: The value
of is 0.001. The norm is checked with 0.5 (1 − λ)/λ = 0.000125. When k = 53,
the -optimal policy is found; we start with J(1) = J (2) = 0
k J(1) J (2) Norm

2 10.700 10.000 10.7
3 19.204 18.224 8.504
4 25.984 24.892 6.781
. ... ... ...
. ... ... ...
53 53.032 51.866 0.000121
regular value iteration algorithm in that its updating mechanism

uses current versions of the value function’s elements; this is called
asynchronous updating. This needs to be explained in more detail.
Review value iteration presented above. In Step 2, J k+1 (i) is
obtained using values of the vector J k . Now, consider a two-state
example. When J k+1 (2) is calculated, one already has the value
of J k+1 (1). But in Step 2 in the algorithm description we use J k (1).

This is where the Gauss-Seidel version differs from the regular version.
In Step 2 of Gauss Seidel, we will use J k+1 (1).
In general, we will use J k+1 (i) of every i for which the (k + 1)th
estimate is available. In other words, in the updating of the values,
we will use the latest estimate of the value. The main idea is that the
algorithm is based on the following update:
 
|S|
J(i) ← max r̄(i, a) + λ p(i, a, j)J(j) .
a∈A(i)
j=1
Note that we avoid using the superscript k in the above, essentially

implying that during any update, one should use the latest estimate
of the value function of any state.
Example A and Gauss-Siedel value iteration. Table 6.6 shows
the calculations with Gauss-Seidel value iteration on Example A from
the Sect. 3.3.2. The discounting factor is 0.8. Note that it takes fewer
iterations to converge in comparison to regular value iteration.
6. Bellman Equation Revisited

The validity of the Bellman equation will be proved using math-
ematical arguments in Chap. 11. However, in this section, we will
motivate the Bellman equation in an intuitive manner. The intuitive
explanation will throw considerable insight on the physical meaning of
the value function.
We consider the Bellman policy equation (i.e., Bellman equation
for a given policy) in the discounted case. The value function of a
given policy for a given state is the expected total discounted reward
earned over an infinitely long trajectory, if the given policy is adhered
Table 6.6. Gauss-Siedel value iteration for discounted reward MDPs: Here =
0.001; the norm is checked with 0.5 (1 − λ)/λ = 0.000125; the -optimal is found
at k = 33; we start with J 1 (1) = J 1 (2) = 0
k J(1) J (2) Norm

2 12.644 18.610 18.612500
3 22.447 27.906 9.803309
. ... ... ...
. ... ... ...
33 39.674 43.645 0.000102
to in every state visited. Keeping this in mind, let us strive to obtain

an expression for the value function of a given state i for a policy d.ˆ
Denoting this element by vdˆ(i), we have that the vdˆ(i) must equal
the immediate reward earned in a transition from i plus the value
function of the state to which it jumps at the end of the transition.
For instance, if you were measuring the total reward along a trajectory
and you encounter two states i and j along the trajectory—i first and
then j—then the reward earned from i to the end of the trajectory
would be the sum of (1) the immediate reward earned in going from
i to j and (2) the total discounted reward earned from j to the end
of the trajectory. See also Fig. 6.13. This is the basic idea underlying
the Bellman equation.
r(i,d(i),j)
i j
v(j)
v(i)
Figure 6.13. Total discounted reward calculation on an infinitely long trajectory
Let us next discuss what happens when the state transitions are
probabilistic and there is a discounting factor λ. When the system
is in a state i, it may jump to any one of the states in the system.
Consider Fig. 6.14.
r(i,d(i),j)
i j
Figure 6.14. Immediate reward in one transition
ˆ the expected total discounted reward earned

So when the policy is d,
from state i, which is vdˆ(i), can now be expressed as the sum of
a. The immediate reward earned in going to state j and
b. λ times the value function of the state j.
This is because the value function of state j denotes the expected
total discounted reward from j onwards. But this quantity (that is
the value function of state j) must be multiplied by λ, since it is

earned one time step after leaving i. We assume that the immediate re-
ward is earned immediately after state i is left. Hence the immediate
reward is not multiplied by λ. However, j can actually be any one of
the states in the Markov chain. Hence, in our calculation, we have to
use an expectation over j. Thus, we have that
|S|
vdˆ(i) = p(i, d(i), j) r(i, d(i), j) + λvdˆ(j) ,
j=1
which turns out to be the Bellman equation for a policy (d)ˆ for state i.
We hope that this discussion has served as an intuitive basis for the
Bellman policy equation.
The Bellman optimality equation has a similar intuitive explanation:
In each transition, to obtain the optimal value function at the current
state, i, one seeks to add the maximum over the sum of immediate
reward to the next state j and the “best” (optimal) value function from
state j. Of course, like in the policy equation, we must compute an
expectation over all values of j. We now discuss semi-Markov decision
problems.
7. Semi-Markov Decision Problems

The semi-Markov decision problem or process (SMDP) is a
more general version of the Markov decision problem (or process) in
which the time spent in each state is an important part of the model.
In the SMDP, we assume that the time spent in any transition of the
Markov chain is not necessarily unity; in fact the time spent could
be different for each transition. Furthermore, the time could be a
deterministic quantity or a random variable. It should be clear now
that the MDP is a special case of the SMDP in which the time of
transition is always one.
One may think of an SMDP as a problem very like an MDP with
the difference that every transition can take an amount of time that is
not necessarily equal to 1 unit. Our discussions on the SMDP will be
based on this argument. The Markov chain associated to any policy
(or any action) in the SMDP is called the embedded Markov chain.
Since the transition time from one state to another depends on the
state pair in question, it is common practice to store the expected tran-
sition times in matrices. We will first treat the case of deterministic
transition times.
In the case of deterministic transition times, the transition time is

fixed for a given state-state pair and for a given action. The tran-
sition time of going from state i to state j under the influence of
action a can then be denoted by t̄(i, a, j) in a manner similar to the
notation of the immediate reward (r(i, a, j)) and the transition prob-
ability (p(i, a, j)). Also, each value of t̄(i, a, j) can then be stored in a
matrix—the transition time matrix (TTM). This matrix would be very
like the transition reward matrix (TRM) or the transition probability
matrix (TPM). An SMDP in which the time spent in any transition
is a deterministic quantity will be referred to as a deterministic time
semi-Markov decision problem (DeTSMDP).
When the transition times are random variables from any general
(given) distribution, we have what we will call the generalized SMDP.
The DeTSMDP is, of course, a special case of the generalized SMDP.
When the distribution is exponential, we have a so-called continuous
time Markov decision problem (CTMDP). Thus, we have three types of
SMDPs: (i) generalized SMDPs, (ii) CTMDPs, and (iii) DeTSMDPs.
For the generalized SMDP, we will treat the term t̄(i, a, j) to be the
expected transition time from i to j under action a. These times can
also be stored within a matrix, which will be referred to as the TTM.
For solving the SMDP, via dynamic programming, in addition to the
TPMs and TRMs, we need the TTMs.
The exponential distribution (the distribution for transition times
in CTMDPs) is said to have a memoryless property. This memoryless
property is sometimes also called the Markov property, but the reader
should not confuse this with the Markov property of the MDPs that we
have discussed before. In fact, the Markov property of the exponential
distributions for the transition times of a CTMDP does not allow us to
treat the CTMDP as an MDP. Via a process called uniformization,
however, one can usually convert a CTMDP into an MDP. The “uni-
formized” CTMDP can be treated as an MDP; but note that it has
transition probabilities that are different from those of the original
CTMDP. Thus, the CTMDP in its raw form is in fact an SMDP and
not an MDP. We will not discuss the process of uniformization here.
For this topic, the reader is referred to other texts, e.g., [30].
By a process called discretization, which is different than the uni-
formization mentioned above, we can also convert a generalized SMDP
into an MDP. We will discuss discretization later in the context of
average reward SMDPs. We now illustrate the notion of transition
times with some simple examples.
Assume that the SMDP has two states numbered 1 and 2. Also,
assume that the time spent in a transition from state 1 is uniformly
distributed with a minimum value of 1 and a maximum of 2 (Unif(1,2)),
while the same from state 2 is exponentially distributed with a mean
of 3 (EXPO(3)); these times are the same for every action. Then, for
generating the TTMs, we need to use the following values. For all
values of a,
t̄(1, a, 1) = (1 + 2)/2 = 1.5; t̄(1, a, 2) = (1 + 2)/2 = 1.5;
t̄(2, a, 1) = 3; t̄(2, a, 2) = 3.
Obviously, the time could follow any distribution. If the distributions
are not available, we must have access to the expected values of the
transition time, so that we have values for each t̄(i, a, j) term in the
model. These values are needed for solving the problem via dynamic
programming.
It could also be that the time spent depends on the action. Thus
for instance, we could also represent the distributions within the TTM
matrix. We will use Ta to denote the TTM for action a. For a 2-state,
2-action problem, consider the following data:
Unif (5, 6) 12 Unif (45, 60) Unif (32, 64)

T1 = ; T2 = .
Unif (14, 16) Unif (5, 12) Unif (14, 16) Unif (12, 15)
In the above, the random values denoted by the distributions should

be replaced by their mean values when performing calculations. We
need to clarify that the distributions in a given TTM need not belong
to the same family, and some (or all) entries in the TTM can also be
constants.
7.1. Natural and Decision-Making Processes

In many MDPs and most SMDPs, we have two stochastic processes
associated with every Markov chain. They are:
1. The natural process (NP).
2. The decision-making process (DMP).
Any stochastic process keeps track of the changes in the state of

the associated system. The natural process keeps track of every state
change that occurs. In other words, whenever the state changes, it gets
recorded in the natural process. Along the natural process, the system
does not return to itself after one transition. This implies that the
natural process remains in a state for a certain amount of time and
then jumps to a different state.
The decision process has a different nature. It records only those
states in which an action needs to be selected by the decision-maker.
Thus, the decision process may come back to itself after one transition.
A decision-making state is one in which the decision-maker makes a
decision. All states in a Markov chain may not be decision-making
states; there may be several states in which no decision is made. Thus
typically a subset of the states in the Markov chain tends to be the set
of decision-making states. Clearly, as the name suggests, the decision-
making process records only the decision-making states.
For example, consider a Markov chain with three states numbered
1, 2, 3, and 4. States 1 and 2 are decision-making states while 3 and
4 are not. Now consider the following trajectory:
1, 3, 4, 3, 2, 3, 2, 4, 2, 3, 4, 3, 4, 3, 4, 3, 4, 1.
In this trajectory, the NP will look identical to what we see above.
The DMP however will be:
1, 2, 2, 2, 1.
This example also explains why the NP may change several times
between one change of the DMP. It should also be clear that the DMP
and NP coincide on the decision-making states (1 and 2).
We need to calculate the value functions of only the decision-making
states. In our discussions on MDPs, when we said “state,” we meant
a decision-making state. Technically, for the MDP, the non-decision-
making states enter the analysis only when we calculate the imme-
diate rewards earned in a transition from a decision-making state to
another decision-making state. In the SMDP, calculation of the tran-
sition rewards and the transition times needs taking into account the
non-decision-making states visited. This is because in the transition
from one decision-making state to another, the system may have vis-
ited non-decision-making states multiple times, which can dictate (1)
the value of the immediate reward earned in the transition and (2) the
transition time.
In simulation-based DP (reinforcement learning), the issue of iden-
tifying non-decision-making states becomes less critical because the
simulator calculates the transition reward and transition time in tran-
sitions between decision-making states; as such we need not worry
about the existence of non-decision-making states. However, if one
wished to set up the model, i.e., the TRM and the TPM, careful
attention must be paid to this issue.
In SMDPs, the time spent in the different transitions of the Markov

chain is not necessarily the same. In the MDP, the time is irrelevant,
and as such, in the calculation of the performance metric, we as-
sume the time spent to be unity. This means that the average reward
per unit time for an MDP is actually average reward per transition
(jump) of the underlying Markov chain. For the SMDP, however, one
must be careful in this respect. In the SMDP, average reward per
unit time cannot be substituted by average reward per unit transition
because the two quantities are different. In the SMDP, the average
reward per transition will not mean the same thing.
In the next two subsections, we will discuss the average reward and
the discounted reward SMDPs. Our discussions will be centered on
algorithms that can be used to solve these problems.
7.2. Average Reward SMDPs

We first need to define the average reward of an SMDP. The average
reward of an SMDP is a quantity that we want to maximize. Recall
that t̄(i, a, j) is the mean transition time from state i to state j un-
der the influence of action a. Now, the average reward, using a unit
time basis, starting from a state i and following a policy µ̂, can be
mathematically expressed as:
k
E[ s=1 r(xs , µ(xs ), xs+1 )|x1 = i]
ρµ̂ (i) ≡ lim inf k
k→∞ E[ s=1 t̄(xs , µ(xs ), xs+1 )|x1 = i]
where xs is the state from where the sth jump (or state transition)
occurs. The expectation is over the different trajectories that may be
followed under the conditions within the square brackets.
The notation inf denotes the infimum (and sup denotes the supre-
mum). An intuitive meaning of inf is minimum and that of sup is
maximum. Technically, the infimum (supremum) is not equivalent to
the minimum (maximum); however at this stage you can use the two
interchangeably. The use of the infimum here implies that the average
reward of a policy is the ratio of the minimum value of the total reward
divided by the total time in the trajectory. Thus, it provides us with
lowest possible value for the average reward.
It can be shown that the average reward is not affected by the state
from which the trajectory of the system starts. Therefore, one can get
rid of i in the definition of average reward. The average reward on
the other hand depends on the policy used. Solving the SMDP means
finding the policy that returns the highest average reward.
7.2.1 Exhaustive Enumeration

Like in the MDP case, for the SMDP, exhaustive enumeration is
not an attractive proposition from the computational angle. How-
ever, from the conceptual angle, it is an important method to be first
understood. Hence we now discuss it briefly.
We begin by defining the average time spent in a transition from
state i under the influence of action a as follows:
|S|
t̄(i, a) = p(i, a, j)t̄(i, a, j),
j=1
where t̄(i, a, j) is the expected time spent in one transition from state
i to state j under the influence of action a. Now, the average reward
of an SMDP can also be defined as:
|S|
i=1 Πµ̂ (i)r̄(i, µ(i))
ρµ̂ = |S|
(6.23)
i=1 Πµ̂ (i)t̄(i, µ(i))
where
r̄(i, µ(i)) and t̄(i, µ(i)) denote the expected immediate reward
earned and the expected time spent, respectively, in a transition
from state i under policy µ̂ and
Πµ̂ (i) denotes the limiting probability of the underlying Markov
chain for state i when policy µ̂ is followed.
The numerator in the above denotes the expected immediate reward
in any given transition, while the denominator denotes the expected
time spent in any transition. The above formulation (see e.g., [30])
is based on the renewal reward theorem (see Johns and Miller [155]),
which essentially states that
expected reward earned in a cycle
ρ = average reward per unit time = . (6.24)
expected time spent in a cycle
Example B. The example discussed next is similar to the Example A

in Sect. 3.3.2 as far as the TPMs and the TRMs are considered. What
is new here is the TTM (transition time matrix). There are two states
numbered 1 and 2 in an MDP and two actions, also numbered 1 and
2, are allowed in each state. The TTM for action a is represented by
Ta . Similarly, the TPM for action a is denoted by Pa and the TRM
for action a by Ra .
1 5 50 75
T1 = ; T2 = .
120 60 7 2
0.7 0.3 0.9 0.1
P1 = ; P2 = .
0.4 0.6 0.2 0.8
6 −5 10 17
R1 = ; R2 = .
7 12 −14 13
There are 4 possible policies that can be used to control the system:
µ̂1 = (1, 1), µ̂2 = (1, 2), µ̂3 = (2, 1), and µ̂4 = (2, 2).
The TTMs of these policies are constructed from the individual

TTMs of each action. The TTMs are:
1 5 1 5 50 75 50 75
Tµ̂1 = ; Tµ̂2 = ; Tµ̂3 = ; Tµ̂4 = .
120 60 7 2 120 60 7 2
The TPMs and TRMs were calculated in Sect. 3.3.2. The value of
each t̄(i, µ(i)) term can be calculated from the TTMs in a manner
similar to that used for calculation of r̄(i, µ(i)). The values are:
t̄(1, µ1 (1)) = p(1, µ1 (1), 1)t̄(1, µ1 (1), 1) + p(1, µ1 (1), 2)t̄(1, µ1 (1), 2)
= 0.7(1) + 0.3(5) = 2.20;

t̄(2, µ1 (2))=84; t̄(1, µ2 (1))=2.20; t̄(2, µ2 (2))=3.00; t̄(1, µ3 (1))=52.5;
t̄(2, µ3 (2)) = 84.0; t̄(1, µ4 (1)) = 52.5; and t̄(2, µ4 (2)) = 3.00.
The corresponding r̄ terms, needed for computing ρ, were calculated
before; see Sect. 3.3.2. Then, using Eq. (6.23), the average rewards of
all the policies are as follows:
ρµ̂1 = 0.1564, ρµ̂2 = 2.1045, ρµ̂3 = 0.1796, and ρµ̂4 = 0.2685.
Clearly, policy µ̂2 is an optimal policy. The optimal policy here is

different from the optimal policy for the MDP (in Sect. 3.3.2), which
had the same TPMs and TRMs. Thus, although it is perhaps obvious,
we note that the time spent in each state makes a difference to the
average reward per unit time. Hence, it is the SMDP model rather
than the MDP model that will yield the optimal solution, when time is
used for measuring the average reward. We will now discuss the more
efficient DP algorithms for solving SMDPs.
7.2.2 Policy Iteration

Policy iteration of average reward MDPs has a clean and nice
extension to SMDPs. This is unlike value iteration which for average
reward does not extend easily from MDPs to SMDPs. We will begin
by presenting the Bellman equation for a given policy in the context
of average reward SMDPs.
The Bellman policy equation in the average reward context for
SMDPs is:
|S|
hµ̂ (i) = r̄(i, µ(i)) − ρµ̂ t̄(i, µ(i)) + p(i, µ(i), j)hµ̂ (j) for each i ∈ S.
j=1
(6.25)
In the above, notation is as usual; ρµ̂ denotes the average reward per
unit time generated by policy µ̂. The only difference between the above
equation and the Bellman policy equation for average reward MDPs
lies in the time term, i.e., the t̄(i, µ(i)) term. The MDP version can be
obtained by setting the time term to 1 in the above. Like in the MDP
case, the above is a system of linear equations in which the number
of equations is equal to the number of elements in the set S. The
unknowns in the equation are the hµ̂ terms. They are the elements of
the value function vector associated with the policy µ̂.
Steps in policy iteration for SMDPs.
number of states be |S|. Select any policy in an arbitrary manner.
Let us denote the policy selected by µ̂k . Let µ̂∗ denote the optimal
policy.
equations.
|S|
k k
h (i) = r̄(i, µk (i)) − ρ t̄(i, µk (i)) + p(i, µ(i), j)hk (j). (6.26)
j=1
Here one linear equation is associated to each value of i. In this

system, the unknowns are the hk terms and ρk . Any one of the hk
terms should be set to 0.
Step 3. (Policy Improvement) Choose a new policy µ̂k+1 so that

for all i ∈ S
 
|S|
µk+1 (i) ∈ arg max r̄(i, a) − ρ t̄(i, a) +
k
p(i, a, j)hk (j) .
a∈A(i) j=1
If possible, one should set µ̂k+1 = µ̂k .

µk (i) for each i, then stop, and set µ∗ (i) = µk (i) for every i. Oth-
erwise, increment k by 1, and go back to the second step.
Policy iteration on Example B. We will show the use of policy

iteration on generic example from Sect. 7.2.1. The results are shown
in Table 6.7. The optimal policy is (2, 1).
7.2.3 Value Iteration

Recall that value iteration for MDPs in the average reward context
poses a few problems. For instance, the value of the average reward
(ρ∗ ) of the optimal policy is seldom known beforehand, and hence
the Bellman optimality equation cannot be used directly. Let us first
analyze the Bellman optimality equation for average reward SMDPs.
 
|S|
J ∗ (i) = max r̄(i, a) − ρ∗ t̄(i, a) + p(i, a, j)J ∗ (j) for each i ∈ S.
a∈A(i)
j=1
(6.27)
The following remarks will explain the notation.
The J ∗ terms are the unknowns. They are the components of the
optimal value function vector J ∗ . The number of elements in the
vector J ∗ equals the number of states in the SMDP.
The term t̄(i, a) denotes the expected time of transition from state
i when action a is selected in state i.
The term ρ∗ denotes the average reward associated with the optimal
policy.
Now in an MDP, although ρ∗ is unknown, it is acceptable to replace
ρ∗ by 0 (which is the practice in regular value iteration for MDPs),
or to replace it by the value function associated with some state of
the Markov chain (which is the practice in relative value iteration
for MDPs). In the MDP, this replacement is acceptable, because it

does not affect the policy to which value iteration converges (although
it does affect the actual values of the value function to which the
algorithm converges). However, in the SMDP, such a replacement
violates the Bellman equation. This is the reason we have difficulties
in developing a regular value iteration algorithm for the average reward
SMDP. We now illustrate this difficulty with an example.
7.2.4 Regular Value Iteration

Our example will show that naive application of value and relative
value iteration will not work in the SMDP. Let us define W (i, a) as
follows:
|S|
∗
W (i, a) ≡ r̄(i, a) − ρ t̄(i, a) + p(i, a, j)J k (j)
j=1
for all (i, a) pairs, where J k (i) denotes the estimate of the value func-
tion element for the ith state in the kth iteration of the value iteration
algorithm.Let us define W (i, a), which deletes the ρ∗ and also the time
Table 6.7. Calculations in policy iteration for average reward SMDPs (Example B)
Iteration (k) Policy selected (µ̂k ) Values ρ

1 (1, 1) h1 (1) = 0 0.156
h1 (2) = −7.852
2 (2, 2) h2 (1) = 0 0.2685
h2 (2) = 33.972
3 (1, 2) h3 (1) = 0 2.1044
h3 (2) = 6.432
term (t̄(., .)), as follows:

|S|
W (i, a) ≡ r̄(i, a) + p(i, a, j)J k (j)
j=1
for all (i, a) pairs.

Provided the value of ρ∗ is known in advance (this is not the case
in practice, but let us s assume this to be the case for a minute), a
potential value iteration update with the Bellman equation will be:
 
|S|
∗
J k+1
(i) ← max r̄(i, a) − ρ t̄(i, a) + p(i, a, j)J k (j) . (6.28)
a∈A(i)
j=1
Now, consider an SMDP with two actions in each state, where t̄(i, 1) =
t̄(i, 2). For this case, the above value iteration update can be writ-
ten as:
J k+1 (i) ← max{W (i, 1), W (i, 2)}. (6.29)
If regular value iteration, as defined for the MDP, is used here, one
must not only ignore the ρ∗ term but also the time term. Then, an
update based on a regular value iteration for the SMDP will be (we
will show below that the following equation is meaningless):
 
|S|
J k+1
(i) ← max r̄(i, a) + p(i, a, j)J k (j) , (6.30)
a∈A(i)
j=1
which can be written as:

J k+1 (i) ← max{W (i, 1), W (i, 2)}. (6.31)
Now, since t̄(i, 1) = t̄(i, 2), it is entirely possible that using (6.29), one
obtains J k+1 (i) = W (i, 1), while using (6.31), one obtains J k+1 (i) =
W (i, 2).
In other words, the update in (6.31) will not yield the same maxi-
mizing action as (6.29), where (6.29) is based on the Bellman equation,
i.e., Eq. (6.28). (Note that both will yield the same maximizing action
if all the time terms are equal to 1, i.e., in the MDP). Thus regu-
lar value iteration, i.e., (6.31), which is based on Eq. (6.30), is not
an acceptable update for the SMDP in the manner shown above. It
should be clear thus that Eq. (6.28) cannot be modified to eliminate
ρ∗ without eliminating t̄(., .) at the same time; i.e., Eq. (6.30) has no
validity for SMDPs! And herein lies the difficulty with value iteration
for SMDPs. There is, however, a way around this difficulty, which we
now discuss.
7.2.5 Discretization for Generalized SMDPs

For circumventing the difficulty described above, one recommended
procedure is to “discretize” the SMDP and thereby convert it into
an MDP. This process transforms the SMDP into an MDP by con-
verting the immediate rewards to rewards measured on a unit time
basis. The “discretized” Bellman optimality equation is:
 
|S|
∗
J (i) = max r̄ (i, a) −
ϑ
ρϑ∗ + pϑ (i, a, j)J ∗ (j) , for all i ∈ S,
a∈A(i)
j=1
(6.32)
where the replacements for r̄(i, a) and p(i, a, j) are denoted by r̄ϑ (i, a)
and pϑ (i, a, j), respectively, and are defined as:
r̄ ϑ (i, a) = r̄(i, a)/t̄(i, a);
ϑp(i, a, j)/t̄(i, a) if i = j
pϑ (i, a, j) =
1 + ϑ[p(i, a, j) − 1]/t̄(i, a) if i = j
In the above, ϑ is chosen such that:
0 ≤ ϑ ≤ t̄(i, a)/(1 − p(i, a, j)) for all a, i and j.
That the above Bellman optimality equation in Eq. (6.32) is perfectly

valid for the average reward SMDP and is equivalent to using Eq. (6.27)
for value iteration purposes is proved in [30], amongst other sources.
The advantage of Eq. (6.32) for value iteration is that unlike Eq. (6.27)
it contains the term ρϑ∗ without time as a coefficient.
It is now possible to use natural value iteration using this trans-
formed Bellman equation to solve the SMDP after setting ρϑ∗ to 0
like in the MDP (note that the resulting equation will be different
than Eq. (6.30) because of the differences in the transition rewards
and probabilities). It is also possible to use the relative value iteration
algorithm just as in the MDP. Before moving on to the next topic, we
note that the discretization procedure changes the transition proba-
bilities of the Markov chain. Also, very importantly, the transformed
Bellman equation has transition time encoded within the transformed
transition probabilities and rewards. We do not present additional de-
tails of this topic, however, because in the simulation-based context
(as we will see in the next chapter), a discretization of this nature can
be avoided.
7.3. Discounted Reward SMDPs

Our attention in this subsection will be focussed on the DeTSMDP
under an additional structural assumption on how the rewards are
earned. This is done to keep the analysis simple. In the next chap-
ter that considers a simulation-based approach, which is after all the
focus of this book, we will cover a more general model (the generalized
SMDP) for discounting without these assumptions.
We begin with the definition of an appropriate discounting factor
for the SMDP in discounting. From standard engineering economics,
we know that the present value and future value share a relationship
due to the rate of interest (or inflation). If a fixed rate of interest is
used that is compounded after a fixed time τ (e.g., typically, annually

in our engineering economics problems), the relationship is:
1
Present Value = Future Value,
(1 + γ)τ
where γ is the rate of return or rate of interest. In the above formula, γ
should be expressed as a fraction (not in percent, although we usually
think of rate of interest in percentages). If τ = 1, i.e., transition time
in the process is equal to 1, we have an MDP. Then λ, which we used
to denote the discounting factor, is:
1
λ= .
1+γ
If γ > 0, we have that 0 < λ < 1. (Note that in average reward
problems, we assume γ = 0.) Now, let us consider the case of contin-
uously compounded interest, where the time period can be small
(as opposed to fixed time period above). The discount factor λ will
now be raised to τ , which leads to:
τ
1 τ
τ
λ = ≈ e−γ = e−γτ .
1+γ
The above exponential approximation based on the exponential series,
which is true when γ is small, has been well-known in the operations
research community because the principle is widely used in the engi-
neering economics community, where it is used to relate the present to
the future value:
Present Value = e−γτ Future Value.
In other words, to obtain the present value (worth) the future value is
multiplied by e−γτ , which, in a rough sense, becomes the discounting
factor here. Here, we have continuously compounded rates, since the
rewards are earned over time continuously. The Bellman equations
must hence account for this new discounting factor. We now make
two assumptions about the SMDP that we first study.
Assumption 6.4 The discounted SMDP is a DeTSMDP.
Assumption 6.5 The immediate reward is earned in a lump sum im-
mediately after the transition starts.
Assumption 6.5 implies that the immediate reward term, r(i, a, j),
will denote the reward earned as soon as action a is selected in state
i, i.e., as soon as the transition begins. Hence, we assume that no

reward is earned after that instant during that transition, and that
all the reward associated with a transition is earned as a lump sum
at the start of the transition. This assumption is not needed in the
average reward problem. The reasons are as follows. In the average
reward problem, since there is no discounting, the immediate reward
earned at any time during the transition can simply be lumped to-
gether into r(i, a, j). However, in the discounted SMDP, if a part of
the immediate reward is earned during the transition, then since we as-
sume continuous compounding, that portion of the immediate reward
earned during the transition must be properly discounted. This can
be done by a suitable modification of the Bellman equation that we
will discuss at the end of this subsection. For the time being, in order
to keep the analysis simple, we assume that the immediate reward is
earned as soon as the transition starts.
Under Assumptions 6.4 and 6.5, the Bellman equation for a policy
is given by:
|S|
hµ̂ (i) = r̄(i, µ(i)) + e−γ t̄(i,µ(i),j) p(i, µ(i), j)hµ̂ (j)
j=1
for each i ∈ S. The role of the discounting factor here is played by

e−γ t̄(i,µ(i),j) .
7.3.1 Policy Iteration

In what follows, we present a discussion on the use of policy iteration
in solving discounted reward SMDPs under Assumptions 6.4 and 6.5.
The policy iteration algorithm is very similar to that used for MDPs.
The discounting factor accommodates the time element.
Step 1. Set k = 1. Here k will denote the iteration number. Let

the set of states be denoted by S. Select a policy in an arbitrary
manner. Let us denote the policy selected by µ̂k . Let µ̂∗ denote
the optimal policy.

equations. For i = 1, 2, . . . , |S|,
|S|
k
h (i) = r̄(i, µk (i)) + e−γ t̄(i,µk (i),j) p(i, µ(i), j)hk (j). (6.33)
j=1

 
|S|
µk+1 (i) = arg max r̄(i, a) + e−γ t̄(i,a,j) p(i, a, j)hk (j) .
a∈A(i) j=1
If possible one should set µ̂k+1 = µ̂k .

µk (i) for each i, then stop, and set µ∗ (i) = µk (i) for every i. Oth-
erwise, increment k by 1, and go back to the second step.
7.3.2 Value Iteration

Value iteration can be carried out on discounted reward SMDPs
under Assumptions 6.4 and 6.5.
Step 1: Set k = 1. Select arbitrary values for the elements of a vector
of size |S|, and call the vector J 1 . Specify > 0.
 
|S|
J k+1 (i) ← max r̄(i, a) + e−γ t̄(i,a,j) p(i, a, j)J k (j) .
a∈A(i)
j=1
Step 3: If
sp(J k+1 − J k ) < ,
go to Step 4. Otherwise increase k by 1, and go back to Step 2.
 
|S|
d(i) ∈ arg max r̄(i, a) + e−γ t̄(i,a,j) p(i, a, j)J k (j) ,
a∈A(i) j
and stop.
7.3.3 Generalized SMDPs Without Assumptions 6.4

and 6.5
We now develop the Bellman equations for the generalized SMDP
where Assumptions 6.4 and 6.5 are not needed, i.e., the time spent in
a transition need not be deterministic and the reward can be earned
any time during the transition. This is the most general case that
requires a suitable modification of the model presented above for the
discounted SMDP.
Since we will assume that the transition time can be random, we

need to introduce its distribution function. Also, we now need to
distinguish between the immediate reward earned at start, which we
call the lump sum reward, and the reward earned continuously over
the transition. We present some notation needed to set up the Bellman
equations. Let fi,a,j (.) denote the pdf of the transition time from i to
j under the influence of action a, rC (i, a, j) denote the continuous rate
of reward (e.g., in dollars per hour) from state i to j under action a,
and rL (i, a, j) denote the immediate reward (e.g., in dollars) earned
from i immediately after action a is taken and the system goes to j
(lump sum reward). The Bellman equation for a policy µ̂ is: For every
i ∈ S, Jµ (i) =
∞
p(i, µ(i), j)rL (i, µ(i), j) + R(i, µ(i)) + e−γτ fi,µ(i),j (τ )Jµ (j)dτ ,
j∈S j∈S 0
∞
1 − e−γτ
where R(i, a) = rC (i, a, j) fi,a,j (τ )dτ
0 γ
j∈S
and lim fi,a,j (τ ) = p(i, a, j).

τ →∞
The Bellman optimality equation is given by: For every i ∈ S, J(i) =
∞
max p(i, a, j)rL (i, a, j) + R(i, a) + e−γτ fi,a,j (τ )J (j)dτ .
a∈A(i) 0
j∈S j∈S
These equations involve integrals, and in order to use them within

policy and value iteration, one must perform these integrations, which
is possible when the pdfs are available. In a simulation-based context,
however, these integrations can be avoided; we will see this in Chap. 7.
Since our focus is on simulation-based solutions in this book, we bypass
this topic. See [242, 30] for additional details.
8. Modified Policy Iteration

An approach outside of value and policy iteration goes by the name
modified policy iteration. It combines advantages of value iteration
and policy iteration and in some sense is devoid of the disadvantages
of both. At this point, a discussion on the relative merits and demerits
of both approaches is in order.
Policy iteration converges in fewer iterations than value iteration,
but forces us to solve a system of linear equations in every iteration.
Value iteration is slower (computer time) than policy iteration, but

does not require the solution of any system of equations.
Solving linear equations is not difficult if one has a small system of
linear equations. But in this book, we are interested in models which
can work without transition probabilities, and as such formulating a
linear system is to be avoided. The real use of this algorithm will
therefore become obvious in simulation-based approaches to DP. We
will discuss this algorithm for MDPs only.
Modified policy iteration uses value iteration in the policy evalua-
tion stage of policy iteration; thereby it avoids having to solve linear
equations. Instead of using Gauss elimination or some other linear
equation solver, the algorithm uses value iteration to solve the system.
However, it uses the scheme of policy iteration of searching in policy
space. In other words, it goes from one policy to another in search of
the best doing value iteration in the interim.
An interesting aspect of the modified policy iteration algorithm is
that the policy evaluation stage can actually be an incomplete value
iteration. We now present details for the discounted MDP.
8.1. Steps for Discounted Reward MDPs

set of states be denoted by S. Assign arbitrary values to a vector J k
of size |S|. Choose a sequence {mk }∞k=0 where the elements of this
sequence take non-decreasing integer values, e.g., {5, 6, 7, 8, . . .}.
Specify > 0.
Step 2. (Policy Improvement) Choose a policy µ̂k+1 such that

 
|S|
µk+1 (i) ∈ arg max r̄(i, a) + λ p(i, a, j)J k (j) .
a∈A(i) j=1
If possible one should set µ̂k+1 = µ̂k when k > 0.
Step 3. (Partial Policy Evaluation)
Step 3a. Set q = 0 and for each i ∈ S, compute:

 
W q (i) ← max r̄(i, a) + λ p(i, a, j)J k (j) .

a∈A(i)
j
Step 3b. If
||(W q − J k )|| < (1 − λ)/2λ,
go to Step 4. Otherwise go to Step 3c.
Step 3c. If q = mk , go to Step 3e. Otherwise, for each i ∈ S, com-
pute:
 
W q+1 (i) ← r̄(i, µk+1 (i)) + λ p(i, µk+1 (i), j)W q (j) .
j∈S
Step 3d. Increment q by 1, and return to Step 3c.

Step 3e. For each i ∈ S, set:
k
J k+1 (i) ← W m (i),
increment k by 1 and return to Step 2.
Step 4. The policy µ̂k+1 is the -optimal policy.
Remark 1: A Gauss-Seidel version of the value iteration performed in

Step 3a and Step 3c can be easily developed. This can further enhance
the rate of convergence.
Remark 2: The algorithm generates an -optimal policy, unlike

regular policy iteration.
Remark 3: The algorithm can be shown to converge for any values

of the sequence {mk }∞ k=0 . However, the values of the elements can
affect the convergence rate. The choice of the elements in the sequence
{mk }∞
k=0 naturally affects how complete the policy evaluation is. An
increasing sequence, such as {5, 6, 7, . . .}, ensures that the evaluation
becomes more and more complete as we approach the optimal policy.
In the first few iterations, a very incomplete policy evaluation should
not cause much trouble.
8.2. Steps for Average Reward MDPs

set of states be denoted by S. Assign arbitrary values to a vector J k
of size |S|. Choose a sequence {mk }∞k=0 where the elements of this
sequence take non-decreasing integer values, e.g., {5, 6, 7, 8, . . .}.
Specify > 0.
Step 2. (Policy Improvement) Choose a policy µ̂k+1 such that

 
|S|
µk+1 (i) ∈ arg max r̄(i, a) + p(i, a, j)J k (j) .
a∈A(i) j=1
If possible, one should set µ̂k+1 = µ̂k , when k > 0.

Step 3. (Partial Policy Evaluation)
Step 3a. Set q = 0, and for each i ∈ S, compute:
 
W q (i) = max r̄(i, a) + p(i, a, j)J k (j) .

a∈A(i)
j∈S
Step 3b. If
sp(W q − J k ) ≤ ,
go to Step 4. Otherwise go to Step 3c.
Step 3c. If q = mk , go to Step 3e. Otherwise, for each i ∈ S, com-
pute:
 
W q+1 (i) = r̄(i, µk+1 (i)) + p(i, µk+1 (i), j)W q (j) .
j∈S
Step 3d. Increment q by 1 and return to Step 3c.

k
Step 3e. For each i ∈ S, set: J k+1 (i) = W m (i), increment k by 1
and return to Step 2.
Step 4. The policy µ̂k+1 is the -optimal policy.
Some elements of the value function W can become very large or

very small, because value iteration for average reward can make the it-
erates unbounded. As a safeguard, one can use relative value iteration
in the policy evaluation step of modified policy iteration.
9. The MDP and Mathematical Programming

It is possible to set up both the average reward and the discounted
reward MDPs as Linear Programs (LPs). The LP is a heavily re-
searched topic in operations research, and efficient methods exist for
solving relatively large LPs. (For an elaborate discussion on LP, see
any undergraduate text, e.g., [293].) However, when casting even a

small MDP as an LP, one creates large LPs. Hence, it is unclear
whether this is the best approach for solving the MDP. However, the
LP provides us with some interesting insight on the nature of the dy-
namic program. It should be noted that it is a parametric optimization
(static) approach to solving what is essentially a control (dynamic)
problem. We begin with the average reward case.
Average reward MDPs: An LP formulation. We present with-
out proof an LP formulation that exploits the Bellman equation:
Minimize ρ subject to
|S|
ρ + v(i) − p(i, µ(i), j)v(j) ≥ r̄(i, µ(i)) for i = 1, 2, . . . , |S| and all µ(i) ∈ A(i).
j=1
Here all values of v(j) for j = 1, 2, . . . , |S| and ρ are unrestricted in

sign (URS). The decision variables in the LP are the v(.) terms and ρ.
The optimal policy can be determined from the optimal value of the
vector v. (See the last step of any value iteration algorithm.)
An alternative LP formulation, which is more popular (although it
is difficult to see its relationship with the Bellman equation), is the
dual of the LP above:
Maximize r̄(i, a)x(i, a) such that

i∈S a∈A(i)
for all j ∈ S, x(j, a) − p(i, a, j)x(i, a) = 0, (6.34)

a∈A(j) i∈S a∈A(i)
x(i, a) = 1, (6.35)
i∈S a∈A(i)
and x(i, a) ≥ 0 ∀(i, a). (6.36)

In the above, the x(i, a) terms are the decision variables. It can be
proved [251] that a deterministic policy, i.e., a policy that prescribes
a unique action for each state, results from the solution of the above.
The deterministic policy can be obtained as follows from the optimal
solution of the LP. For all i ∈ S and a ∈ A(i), compute:
x∗ (i, a)
d(i, a) = ∗
b∈A(i) x (i, b)
where x∗ (i, a) denotes the optimal value of x(i, a) obtained from solv-
ing the LP above, and d(i, a) will contain the optimal policy. Here
d(i, a) must be interpreted as the probability of selecting action a in

state i. Since it can be proved that a deterministic optimal policy will
be contained in the d(i, a) terms, it will be the case that d(i, a) will
in fact equal 1 if action a is to be chosen in state i and 0 otherwise.
Thus, the optimal policy can be readily obtained from the values of
d(i, a).
SMDP: For the SMDP, the LP described above can be used after
replacing constraint (6.35) with:
x(i, a)t̄(i, a) = 1.
i∈S a∈A(i)
Discounted reward MDPs: An LP formulation. It can be shown

that if
|S|
v(i) ≥ r̄(i, µ(i)) + λ p(i, µ(i), j)v(j)
j=1
for all policies µ̂, then v(i) is an upper bound for the optimal value
v ∗ (i). This paves the way for an LP. The formulation, using the x and
v terms as decision variables, is:
|S|
Minimize j=1 x(j)v(j) subject to
|S| |S|
j=1 x(j) = 1 and v(i) − λ j=1 p(i, µ(i), j)v(j) ≥ r̄(i, µ(i)) for i =
1, 2, . . . , |S| and all µ(i) ∈ A(i);
x(j) > 0 for j = 1, 2, . . . , |S|, and v(j) is URS for j = 1, 2, . . . , |S|.
10. Finite Horizon MDPs

In the finite horizon MDP, the objective function is calculated over
a finite time horizon. This is different than the infinite time horizon
we have assumed in all the problems above.
With the finite time horizon, we can think of two objective functions:
1. Total expected reward and
2. Total expected discounted reward.
Note that in the infinite horizon setting, the total expected reward is
usually infinite, but that is the not the case here. As such the total
expected reward is a useful metric in the finite horizon MDP.
In this setting, every time the Markov chain jumps, we will
assume that the number of stages (or time) elapsed since the start
increases by 1. In each stage, the system can be in any one of the

states in the state space. As a result, the value function depends not
only on the state but also on the stage.
In the infinite horizon MDP, when we visit a state, we do not concern
ourselves with whether it is the first jump of the Markov chain or the
nth jump. Thus, for instance in the infinite horizon problem, state 15
visited in stage 1 is the same thing as state 15 visited in stage 119.
This is because the value function is associated with only the state.
In other words, in the infinite horizon problem, for every state, we
associate a unique element of the value function. In the finite horizon
MDP, on the other hand, we need to associate a unique element of the
value function with a given state-stage pair. Thus state 15 visited in
stage 1 and the same state 15 visited in stage 119 will have different
value functions. Further, in the infinite horizon MDP, we have a unique
transition probability matrix and a unique transition reward matrix for
a given action. In the finite horizon MDP, we have a unique transition
probability matrix and a unique transition reward matrix for a given
stage-action pair.
It should be clear then that the finite horizon problem can become
more difficult than the infinite horizon problem with the same number
of states, if there is a large number of stages. For a small number
of stages, however, the finite horizon problem can be solved easily.
The popular method used for solving a finite horizon problem is called
backward recursion dynamic programming. As we will see below, is
has the flavor of value iteration.
.......
.......
.......
.......
.......
1 2 T T+1
Terminal
Stage
(Non-
Decision Making Stages decision
-making)
Figure 6.15. A finite horizon MDP
See Fig. 6.15 to obtain a pictorial idea of a finite horizon MDP.

In each of the stages, numbered 1 through T , the system can visit
a subset of the states in the system. It is entirely possible that in

each stage the set of states visited is a proper subset of the total state
space; that is in each stage, any state in the system may be visited.
Thus in general, we can construct a transition probability matrix and a
transition reward matrix for each action after replacing the state by a
state-stage combination. In each stage, the system has a finite number
of actions to choose from, and the system proceeds from one stage to
the next until the horizon T is met. The goal is to maximize the total
expected reward (or the total expected discounted reward) in the T
transitions.
Our approach is based on the value iteration idea. We will endeavor
to compute the optimal value function for each state-stage pair. We
will need to make the following assumption:
Since there is no decision making involved in the states visited in
the (T + 1)th stage, (where the trajectory ends), the value function
in every state in the (T + 1)th stage will have the same value. For
numerical convenience, the value will be 0.
The Bellman optimality equation for the finite horizon problem for
expected total discounted reward is given by:
 
|S|
∗
J (i, s) ← max r̄(i, s, a) + λ p(i, s, a, j, s + 1)J ∗ (j, s + 1)
a∈A(i,s)
j=1
(6.37)
for i = 1, 2, . . . , |S|, and s = 1, 2, . . . , T with some new notation:
J ∗ (i, s): the value function for the ith state when visited in the sth
stage of the trajectory
A(i, s): the set of actions allowed in state i when visited in the sth
stage of the trajectory
r̄(i, s, a): the expected immediate reward earned when in state i,
in the sth stage, action a is selected
p(i, s, a, j, s + 1): the transition probability of going from state i
in the sth stage to state j in the (s + 1)th stage when action a is
selected in state i in the sth stage
The optimality equation for expected total reward can be obtained by
setting λ, the discount factor, to 1 in Eq. (6.37). The algorithm that
will be used to solve the finite horizon problem is not an iterative
algorithm. It just makes one sweep through the stages and produces
the optimal solution. We will now discuss the main idea underlying
the backward recursion technique for solving the problem.
The backward recursion technique starts with finding the values of
the states in the T th stage (the final decision-making stage). For this,
it uses (6.37). In the latter, one needs the values of the states in
the next stage. We assume the values in the (T + 1)th stage to be
known (they will all be zero by our convention). Having determined
the values in the T th stage, we will move one stage backwards, and
then determine the values in the (T − 1)th stage.
The values in the (T − 1)th stage will be determined by using the
values in the T th stage. In this way, we will proceed backward one
stage at a time and find the values of all the stages. During the evalua-
tion of the values, the optimal actions in each of the states will also be
identified using the Bellman equation. We now present a step-by-step
description of the backward recursion algorithm in the context of dis-
counted reward. The expected total reward algorithm will use λ = 1
in the discounted reward algorithm.
A Backward Recursion. Review notation provided for Eq. (6.37).
Step 1. Let T (a positive integer) be the number of stages in the

finite horizon problem. Decision-making will not be made in the
(T + 1)th stage of the finite horizon problem. The notation u∗ (i, s)
will denote the optimal action in state i when the state is visited
in the sth stage. The stages are numbered as 1, 2, . . . , T + 1. For
i = 1, 2, . . . , |S|, set
J ∗ (i, T + 1) ← 0.
Step 2a. Set s ← T .

Step 2b. For i = 1, 2, . . . , |S|, choose
 
|S|
u∗ (i, s) ∈ arg max r̄(i, s, a) + λ p(i, s, a, j, s + 1)J ∗ (j, s + 1) .
a∈A(i,s) j=1
Step 2c. For i = 1, 2, . . . , |S|, set

 
|S|
J ∗ (i, s) ← r̄(i, s, u∗ (i, s)) + λ p(i, s, u∗ (i, s), j, s + 1)J ∗ (j, s + 1) .
j=1
Step 2d. If s > 1, decrement s by 1 and return to Step 2b; otherwise

STOP.
11. Conclusions
This chapter discussed the fundamental ideas underlying MDPs and
SMDPs. The focus was on a finite state and action space within
discrete-event systems. The important methods of value and policy
iteration (DP) were discussed, and the two forms of the Bellman equa-
tion, the optimality equation and the policy equation, were presented
for both average and discounted reward. The modified policy iteration
algorithm was also discussed. Some linear programming for solving
MDPs along with finite horizon control was covered towards the end
briefly. Our goal in this chapter was to provide some of the theory
underlying dynamic programming for solving MDPs and SMDPs that
can also be used in the simulation-based context of the next chapter.
Bibliographic Remarks. Much of the material presented in this chapter is clas-

sical and can be found in [250, 242, 30, 270]. For an in-depth discussion on stochas-
tic processes, the reader is referred to [295]. For highly intuitive and accessible
accounts on MDPs and Markov chains, see [134, 293]. Filar and Vrieze [86] present
the MDP as a special case of a game-theoretic problem, thereby tying MDP theory
to the much broader framework of stochastic games. They also discuss several open
research issues of related interest.
Much of the pioneering work in algorithms and DP theory was done by Bellman
[23], Shapley [271] (value iteration, 1953), Howard [144] (policy iteration, 1960),
White [320] (relative value iteration, 1963), and van Nunen [305] (modified pol-
icy iteration, 1976). Shapley [271] made important parallel contributions in game
theory (which are applicable to MDPs) at a very early date (1953). Finally, we
need to point out that in this book we have focussed on the case of discrete-event
systems with finite state and action spaces governed by Markov chains; the theory
of control optimization is much broader in its scope.
Case study on total productive maintenance. Our case study

on total productive maintenance, drawn from [112, 126], is meant to
illustrate the use of DP on a problem larger than Example A. Many
systems fail with an increasing probability as they age, and such sys-
tems can often benefit from preventive maintenance. Well-known ex-
amples of such systems are manufacturing (production) lines, bridges,
roads, and electric power plants.
We consider a production line that deteriorates with time. We make
the following assumptions about it: (1) The line is needed every day,
and every morning the manager must decide whether to continue with
one more production cycle or shut the line down for preventive main-
tenance. (2) If the line fails during the day, the repair takes the re-
mainder of the day; further the line becomes available for production
after repair only the next morning. (3) The preventive maintenance
takes the entire day. (4) After a repair or a maintenance, the line is
as good as new. (5) Let i denote the number of days elapsed since
the last preventive maintenance or repair (subsequent to a failure);
then the probability of failure during the ith day can be modeled as
1 − ξψ i+2 , where ξ and ψ are scalars in the interval (0, 1), whose values
can be estimated from the data for time between successive failures of
the system.
We will use i to denote the state of the system, since this leads to a
Markov chain. In order to construct a finite Markov chain, we define
for any given positive value of ∈ , ī to be the minimum integer
value of i such that the probability of failure on the īth day is less than
or equal to (1 − ). Since we will set to some pre-fixed value, we can
drop from our notation. In theory, the line will have some probability
of not failing after any given day, making the state space infinite, but
our definition of ī permits truncation of the infinite state space to a
finite one. The resulting state space will be: S = {0, 1, 2, . . . , ī}. This
means that the probability of failure on the īth day (which is very
close to 1) will be assumed to equal 1.
Clearly, when a maintenance or repair is performed, i will be set
to 0. If a successful day of production occurs, i.e., the line does not
fail during the day, the state of the system is incremented by 1. The
action space is: {produce, maintain}. Cm and Cr denote the cost
of one maintenance and one repair respectively. Then, we have the
following transition probabilities for the system.
For action produce: For i = 0, 1, 2, . . . , ī − 1
p(i, produce, i + 1) = ξψi+2 ; p(i, produce, 0) = 1 − ξψi+2 .
For i = ī, p(i, produce, 0) = 1. For all other cases not specified above,
p(., produce, .) = 0. Further, for all values of i,
r(i, produce, 0) = −Cr ; r(i, produce, j) = 0 when j = 0.
For the action maintain: For all values of i, p(i, maintain, 0) = 1

and r(i, maintain, 0) = −Cm . For all other cases not specified above,
p(., maintain, .) = 0; r(., maintain, .) = 0.

Numerical Result: We set ξ = 0.99, ψ = 0.96, Cm = 4, Cr = 2, and
ī = 30. Thus, we have 31 states and 2 actions. Our objective function
is average reward, and the optimal policy, which is determined via
policy iteration, turns out to be one with a threshold nature; i.e.,
the action is to produce for i = 0, 1, . . . , 5 and to maintain for all
values of i from 6 onwards. The policy iteration algorithm took 3

iterations and generated the following values in the final iteration:
v(0)=0; v(1) = −0.4098; v(2)=−0.7280; v(3)=−0.9701; v(4) = −1.1438;
v(5) = −1.2490; and for all i ≥ 6, v(i) = −1.2757. In policy

iteration, the first policy chooses production in each state. Fur-
ther, ρ1 = −0.7566, ρ2 = −0.7425 and ρ3 = ρ∗ = −0.7243. Note that
all the values are negative, because we use rewards (to be consistent
with our notation in this chapter), whereas we only have costs in this
model.

Simulation Based Optimization-1

Uploaded by

Copyright:

Available Formats

Simulation Based Optimization-1

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Simulation Based Optimization-1

Uploaded by

Copyright:

Available Formats

Chapter 6

© Springer Science+Business Media New York 2015 123

Figure 6.1. A single-server queue

although Z(t) is a mathematically perfect example of a stochastic pro-

3. Markov, Semi-Markov, and Decision

An important property of a Markov process is that it jumps regu-

We denote the probability of this transition by P (i, j). This idea is

P{X(t + 1) = j|X(t) = i, X(t − 1) = l} = f (i, l, j), (6.1)

where f (i, l, j) denotes a probability that depends on i, l, and j. This

The Markov process, on the other hand, is governed by the follow-

3.1. Markov Chains

Figure 6.5. Schematic of a Markov chain with three states

Figure 6.6. Schematic of a Markov chain with four states

n-step transition probabilities. The n-step transition probability

transition probabilities. Let us illustrate the meaning of this result

3.1.1 Regular Markov Chains

This is not regular because:

3.1.2 Limiting Probabilities

The elements of a given column in Pn start approaching each other,

Π(j) ≡ lim P n (i, j).

The quantity Π(j) will also be referred to as the limiting or steady-

Since the transitions are assumed to take unit time, 57 % of the

where |S| denotes the number of elements in the set S.

[Π(1), Π(2), . . . , Π(|S|)]P = [Π(1), Π(2), . . . , Π(|S|)],

or in the following abbreviated form: ΠP = Π, where Π is a row vector

For j = 1 : 0.7Π(1) + 0.4Π(2) + 0.6Π(3) = Π(1). (6.6)

− 0.3Π(1) + 0.4Π(2) + 0.6Π(3) = 0. (6.9)

0.2Π(1) − 0.8Π(2) + 0.1Π(3) = 0. (6.10)

0.1Π(1) + 0.4Π(2) − 0.7Π(3) = 0. (6.11)

Now from Eq. (6.5), we have:

Π(1) + Π(2) + Π(3) = 1. (6.12)

Thus we have four Equations: (6.9)–(6.12) and three unknowns:

3.1.3 Ergodic Markov Chains

Figure 6.7. A Markov chain with one transient state

We will now discuss the semi-Markov process.

3.2. Semi-Markov Processes

Figure 6.8. A Markov chain with two transient states

3.3. Markov Decision Problems

Example. Consider a queuing system, such as the one you see in

State = Number waiting for service Action

Figure 6.9. A Markov chain underlying a simple single-server queue

Another possible control mechanism could be:

running this system will be transmitted to the customers through

It naturally makes business sense to use the control mechanism

3.3.1 Elements of an MDP

Decision maker. The decision maker is an entity that selects the

Policies. The control mechanism is usually referred to as a policy.

The word deterministic implies that in any given state, we can

(1, 1), (1, 2), (2, 1), and (2, 2).

An MDP, let us reiterate, revolves around ﬁnding the most suitable

0.7 0.3 0.1 0.9

0.7 0.3 0.1 0.9

0.4 0.6 0.8 0.2

Matrix for action 1 Matrix for action 2

Construction of the TPM for policy (2,1)

See Fig. 6.10 for a pictorial demonstration of the construction pro-

be noted with care. In our previous discussions, we spoke of transition

When policy µ̂ is followed, the immediate reward earned in going from

0.2, 0.3, and 0.5,

respectively. The immediate rewards earned in these three possible