Markov Decision Processes

D.P. Heyman, M.J. Sobel, Eds., Handbooks in OR & MS, Vol.
2
(~) Elsevier Science Publishers B.V. (North-Holland) 1990
Chapter 8
Markov Decision Processes
Martin L. Puterman*
Faculty of Commerce, The University of British Columbia, Vancouver, B.C., Canada
V6T 1Y8
1. Introduction
This chapter presents theory, applications and computational methods for

Markov Decision Processes (MDP's). MDP's are a class of stochastic sequen-
tial decision processes in which the cost and transition functions depend only
on the current state of the system and the current action. These models have
been applied in a wide range of subject areas, most notably in queueing and
inventory control.
A sequential decision process is a model for dynamic system under the
control of a decision maker. At each point of time at which a decision can be
made, the decision maker, who will be referred to as he with no sexist
connotations intended, observes the state of the system. Based on the informa-
tion from this observation, he chooses an action from a set of available
alternatives. The consequences of choosing this action are twofold; the decision
maker receives an immediate reward, and specifies a probability distribution on
the subsequent system state. If the probability distribution is degenerate, the
problem is deterministic. T h e decision maker's objective is to choose a
sequence of actions called a policy, that will optimize the performance of the
system over the decision making horizon. Since the action selected at present
affects the future evolution of the system, the decision maker cannot choose his
action without taking into account future consequences.
Sequential decision processes are classified according to the times (epochs) at
which decisions are made, the length of the decision making horizon, the
mathematical properties of the state and action spaces and the optimality
criteria. The primary focus of this chapter is problems in which decisions are
made periodically at discrete time points (e.g. minutes, hours, days, weeks,
months or years). The state and action sets are either finite, countable,
compact or Borel; their characteristics determine the form of the reward and
* This research has been supported by Natural Sciences and Engineering Research Council
(Canada) Grant A-5527.
331
332 M . L . Puterman
transition probability functions. The optimality criteria considered include

finite and infinite horizon expected total reward, infinite horizon expected total
discounted reward and average expected reward.
The main objectives in analyzing sequential decision processes in general and
MDP's in particular include:
(a) providing an optimality equation which characterizes the supremal value
of the objective function,
(b) characterizing the form of an optimal policy if it exists,
(c) developing efficient computational procedures for finding policies which
are optimal or close to optimal.
The optimality or Bellman equation is the basic entity in MDP theory and
almost all existence, characterization and computational results are based on
analysis of it. The form of this equation depends on the optimality criteria and
the nature of the states, actions, rewards and transition probability functions.
Because of this, the analysis in this chapter is divided according to optimality
criterion. Section 4 treats the finite horizon total expected reward case, Section
6 the expected total discounted reward criteria, Section 7 the expected total
reward criteria and Section 8 the average expected and sensitive optimality
criteria.
In almost any problem, the class of policies can be made sufficiently general
so that an optimal or nearly optimal policy exists. An important theoretical and
practical consideration is under what conditions does there exist a policy in a
specified class that is optimal or nearly optimal among the class of all policies.
The class of stationary policies, each of which uses a decision rule depending
only on the current state of the system, independent of its history or stage,
receives special attention in this chapter. In many applications, an important
further consideration is whether an optimal policy has some special form or
structure. If this is the case, optimal policies can be found by restricting search
to policies of this form.
For each optimality criterion, results are presented in as much generality as
possible. Vector space notation is used to simplify presentation and unify
results. Numerical results are necessarily for finite state and action problems
and presented almost exclusively for the finite horizon expected total reward
and infinite horizon expected total discounted reward criteria. In the finite
horizon case, calculations are based on backward induction which is often
referred to as dynamic programming, while for infinite horizon problems,
calculations are based on value iteration, policy iteration and their variants~
Awareness of the importance and potential applicability of Markov decision
processes began with the appearance of the books of Bellman (1957) and
Howard (1960). Many precursors to their work abound; the interested reader
is referred to Heyman and Sobel (1984) and Denardo (1982) for historical
synopses. From a theoretical prospective, important early work includes
Blackwell (1962), Dubins and Savage (1965), Strauch (t966), Denardo (1967)
and Veinott (1967). In addition to those referred to above, noteworthy books
include Hinderer (1970), Derman (1970), Whittle (1983), Ross (1985), Ber-
tsekas (1987), Hernfindez-Lerma (1989) and Puterman (1991).
Ch. 8. Markov Decision Processes 333
This chapter is broad in scope. In Section 2, Markov decision processes are

introduced and formal notation is presented. Section 3 gives some simple
examples of MDP's. Section 4 is concerned with finite horizon problems while
section 5 provides an introduction to the infinite horizon setting which is
presented in considerable detail in Sections 6-8. Semi-Markov decision prob-
lems are treated in Section 9. An extensive bibliography is included.
Related topics not covered in this chapter include partially observable
Markov decision processes (Monahan, 1982 and Bertsekas, 1987) and Markov
decision processes with estimated parameters (Mandl, 1974 and Hubner,
1988).
For a quick overview of the subject the reader should focus on Sections 2} 3,
4.4, 6.1-6.6 and 8.1-8.8.
2. Problem formulation
This section defines the basic elements of a Markov decision process and
presents notation.
2.1. Problem definition and notation

A system under control of a decision maker is observed as it evolves through
time. The set T consists of all time points at which the system is observed and
decisions can be made. These time points are denoted by t and are referred to
as decision epochs or stages. T h e set T can be classified in two ways; T is either
finite or infinite and either a discrete set or a continuum. The primary focus of
the chapter is when T is discrete; a special case of the continuous time model
will be discussed in Section 9 of this chapter; the continuous time model will be
the focus of Chapter 9.
Discrete time problems are classified as either finite horizon or infinite
horizon according to whether the set T is finite or infinite. The problem
formulation in these two cases will be almost identical; however, theory and
computational methods will differ considerably. For finite horizon problems,
T = { 1, 2 , . . . , N} and for infinite horizon problems, T = { 1, 2 , . . . }.
The set of possible states of the system at time t is denoted by S t. In finite
horizon problems, S t is defined for t = 1, 2 , . . . , N + 1 although decisions are
only made at times t = 1, 2 , . . . , N. This is because the decision at time N
often has future consequences which can be summarized by evaluating the state
of the system at time N + 1.
If at time t, the decision maker observes the system in state s C St, he must
choose an action, a, from the set of allowable actions at time t, As, t. In most of
the literature, S t and A,,, are either finite, countably infinite or compact subsets
of the real line. The most general setting is when S, and As, t are non-empty
Borel subsets of complete, seperable metric spaces. To avoid measurability and
integrability issues, all of the analysis in this chapter will assume that S t is
discrete.
i
334 M.L. Puterman
There are two consequences of choosing action a when the system is in state
s at time t; the decision maker receives an immediate reward and the
probability distribution for the state of the system at the next stage is
determined. The reward is denoted by the real valued function r,(s, a); when it
is positive it can be thought of as income and when negative as cost. In some
applications, it is convenient to think of rt(s, a) as the expected reward received
at time t. This will be the case if the reward for the current period depends on
the state of the system at the next decision epoch. In such situations rt(s , a, j )
is the reward received in period t if the state of the system at time t is s, action
a @ A s is selected and the system is in state j at time t + 1. Then the expected
reward in period t is
r,(s, a) = ~ r,(s, a, j)pt(jls, a)

j~St+l
where Pt(J Is, a) is defined below. The example in Section 3.2 illustrates this
alternative.
The function p,(jls, a) denotes the probability that the system is in state
j E St+ 1 if action a E As, , is chosen in state s at time t; p,(jls, a) is called the
transition probability function. When S t is not discrete, Pt( J ]s, a) is a density, if
it exists; otherwise the problem formulation is in terms of a distribution
function. In most applications it is convenient to assume that
pt(jls, a ) = l . (2.1)
jESt+ 1
The collection of objects (T, St, As,t, pt(j[s,a), rt(s,a)) is a Markov

decision process. Its distinguishing feature is that the transition probability
function and reward function depend only on the current state of the system
and the current action selected. Some authors refer to the above collection of
objects as a Markov decision problem; here that terminology is reserved for a
Markov decision process together with an optimality criterion. Generalizations
allow the rewards and transition probability functions to depend on some or all
previous states and actions (cf. Hinderer, 1970).
A decision rule is a function dr: S t ~ A s , t that specifies the action the decision
maker chooses when the system is in state s at time t, that is, dt(s ) ~ As. , for
each s E S,. A decision rule of this type is said to be Markovian because it
depends only on the current state and stage of the system and not on its past. It
is referred to as deterministic because it selects an action with certainty. T h e set
of allowable decision rules at time t is denoted by D t and is called the decision
set.
Generalizations of deterministic Markovian decision rules play an important
role in M D P theory. A decision rule is said to be history dependent if it is a
function of the entire past history of the system as summarized in the sequence
of previous states and actions, (sl, a ~ , . . . , s,_l, at_ 1, st). A decision rule is
said to be randomized if in each state it specifies a probability distribution on
the set of allowable actions. A deterministic decision rule can be regarded as a

special case of a randomized rule in which the probability distribution is
degenerate. Some authors distinguish deterministic rules from randomized
rules by referring to them as pure. If a decision maker uses a randomized
decision rule, the action that is actually implemented is a random event. An
important theoretical issue is when is it optimal to use a deterministic Marko-
vian decision rule at each stage.
A policy specifies the decision rule to be used by the decision maker at each
t ~ T. It tells the decision maker which decision to choose under any possible
future state the system might occupy. A policy 7r is a sequence of decision
rules, i.e. 7r = {dl, d 2 , . . . , dN} where d t E D t for t = 1, 2 , . . . , N for N ~< ~. It
will be referred to as randomized, history remembering, deterministic or pure
if each decision rule it contains has that property. Let H denote the set of all
history remembering randomized policies; H = D 1 D 2 . . . DN, N ~ . A
policy is said to be stationary if it uses the identical decision rule in each period
i.e. ~r = (d, d , . . . ) . Stationary policies play an important role in the theory of
infinite horizon Markov decision processes and are discussed in more detail in
the next subsection.
The following subsets of H are distinguished; firM, the set of all Markovian
policies, rip, the set of all non-randomized (pure) history dependent policies,
ris, the set of all stationary (possibly randomized) Markovian policies and HD,
the set of all deterministic stationary Markovian policies. Observe that ri o c
Hs C HM C II and Hp C HD C H.
Specifying a policy 1r = ( d l , d 2 , . . . , dN} induces a stochastic process,
(X~; t E T}, with time varying state space S t. It represents the state of the
system at each decision epoch. When the policy 7r is Markov, that is each
decision rule is Markov, then the stochastic process ( X t ; t E T ) is a discrete
time Markov chain. If the policy is Markov and stationary, then the Markov
chain is also stationary. This Markov chain together with the sequence of
rewards (rt(s, dr(s)), s E St, t E T ) is often referred to as a Markov reward
process. The stochastic process { r t ( X t , d t ( X t ) ) ; t E T ) is the stream of
rewards received by the decision maker if he adopts policy ~r.
2.2. Stationary Markov decision problems

Frequently in infinite horizon problems, the data is stationary. This means
that the set of states, the set of allowable actions in each state, the rewards, the
transition or transfer functions and the decision sets are the same at every
stage. When this is the case, the time subscript t is deleted and notation S, As,
r(s, a), p ( j l s , a) and D is used. Under most optimality criteria, stationary
policies are optimal when the data of the problem is stationary. The stationary
policy which uses the decision rule d at each stage will be often denoted by d.
When all the data is stationary, the following additional notation will be
useful. Let d ~ D be an arbitrary decision rule and define rd(S ) = r(s, d(s)) and
p d ( j l s ) = p ( j l s , d(s)). These quantities are the one period reward and transi-
336 M.L. Puterman
tion probability function if the system is in state s and the action corresponding
to decision rule system is in state s and the action corresponding to decision
rule d(s) is used. Note that if d is a randomized decision rule, then
ra(s ) = ~ r(s, a)P{d(s) = a}

a@A s
and
pe(jls) = ~ p(jls, a)P{d(s) : a}.

aEA s
It is convenient to use vector space notation for analysis and presentation of

results. Define V to be family of bounded real valued functions on S, that is,
v E V if v : S---~ R. For each v ~ V define the n o r m of v by ]]vii = sup,Es Iv(s)[.
When S is finite, 'sup' can be replaced by 'max' in this definition. The vector
space V together with the norm 11" I[ is a complete, normed linear space or
Banach space (cf. Liusternik and Sobolev, 1961).
Let r e be the m-vector, with sth c o m p o n e n t re(s ) and let Pe be the m x m
matrix with it (s, j)th entry given by p e ( j ] s ) , where m is the n u m b e r of
elements in S. Assume that r e ~ V. This means that r e is bounded. (In Section
6.8 the case of unbounded rewards will be considered.) Since Pa is a probability
matrix, it follows that Per E V for all v @ V. The quantity r e is referred to as
the reward vector and Pd as the transition probability matrix corresponding to
decision rule d. When policy 7r = (d~, d 2 . . . . , d N ) is used, the (s, j ) t h compo-
nent of the n-step transition probability P~ is given by
P(s, j) = PalPal2"'" Pan(s, j) := P ( X , ~, : j Ix7 = s). (2.2)
3. Examples
This section presents two very simple examples of Markov decision pro -~
cesses. The reader is referred to White (1985b) for a recent survey of
applications of MDP's.
3.1. A two state M a r k o v decision process

The following simple example will be useful for illustrating the basic
concepts of a Markov decision process and serve as an example of several
results in latter sections.
At each t the system can be in either of two states s I and s 2. Whenever the
system is in state 1 there are two possible actions at, 1 and al. 2 and in state 2
there is one action a2.1. As a consequence of choosing action aa, 1 in sl, the
decision maker receives an immediate reward of 5 units and at the next

decision epoch the system is in state sl with probability 0.5 and state s 2 with
probability 0.5. If instead action a~, 2 is chosen in state s~, the decision maker
receives an immediate reward of 10 units and at the next decision epoch the
system is in state s 2 with probability 1. If the system is in state s2, the decision
maker must choose action a2.1 . As a consequence of this decision, the decision
maker incurs a cost of 1 unit and at the next decision epoch, the system is in
state s2 with certainty. Figure 3.1 gives a convenient symbolic representation of
this problem.
al, 1 //~{s,.s; ..... ~ a2,1
110,1}
Fig. 3.1. Symbolic representation of two state Markov process.
The above problem is stationary, i.e., the set of states, the sets of actions,
the rewards and transition probabilities do not depend on the stage in which
the decision is made. Thus, the t subscript on these quantities is unnecessary
and will be deleted in subsequent references to the problem. A formal
description follows.
Decision epochs:
T = {1,2 . . . . , N } , N~o
States:
S,={sl,s2}, t~T.
Actions:
Asl,t={a~,l,al,2}, As2,t={a2.1} ~ t627'.
Rewards:
r,(si,al.l)--5 , r,(s~,al,2)=lO , r,(s2, a z , l ) = - l . , t62T.

338 M.L. Puterman
Transition probabilities:
p,(jls~, a1.1) = 0.5, j~Sl,S2~

pt(sl I Sl, al,2) = O, pt(s 21sl, a1.2) = 1,
p,(s I Is2, a2,1) = O, p,(szlsx, a~,~) = 1, tET.
3.2. A stochastic inventory control problem

This section presents a simplified version of an inventory control problem.
Some non-standard assumptions have been made to facilitate calculations. The
numerical example below will be used to illustrate the algorithmic methods in
the subsequent sections. For more detail on inventory theory, the reader is
referred to Chapter 12.
The problem is as follows. Each month, the manager of a warehouse
determines current inventory of a product. Based on this information, he
decides whether or not to order additional product. In doing so, he is faced
with a tradeoff between the costs associated with keeping inventory on hand
and the lost sales or penalties associated with being unable to satisfy customer
demand for the product. The manager's objective is to maximize some measure
of profit (sales revenue less inventory holding and ordering costs) over the
decision making horizon. The demand for the product throughout the month is
random with a known probability distribution.
Several simplifying assumptions enable a concise formulation. They are:
(a) the decision to order additional stock is made at the beginning of each
month and delivery occurs instantaneously,
(b) demand for the product arives throughout the month but all orders are
filled on the last day of the month.
(c) if demand exceeds the stock on hand, the customer goes elsewhere to
purchase the product (excess demand is lost),
(d) the revenues and costs, and the demand distribution are stationary
(identical each month)
(e) the product can be sold only in whole units, and
(f) the warehouse capacity is M units.
Let s, denote the inventory on hand at the beginning of month t, a, the
additional amount of product ordered in month t and D t the random demand in
month t. The demand has a known probability distribution given by pj =
P{D, = j}, j = 0, 1, 2 . . . . . The cost of ordering u units in any month is O(u)
and the cost of storing u units for 1 month is h(u). The ordering cost is given
by:
g+c(u) if u > 0 , (3.1)

O(u)= 0 if u = 0 .
The functions c(u) and h(u) are increasing in u. For finite horizon problems,
the inventory on hand after the last decision epoch has value g(u). Finally, if j
Ch. 8. Markov DecMon Processes 339
units of product are demanded in a month and the inventory u exceeds

demand, the manager receives f ( j ) , if demand exceeds inventory, then he
receives f(u). Let F(u) be the expected revenue in a month if the inventory
prior to receipt of customer orders is u units. It is given in period t by
U--1
F(u) = ~ f ( j ) p j + f(u)P{O,>~ u ) . (3.2)

j-O
This problem is formulated as a Markov decision process as follows.

Decision epochs:
T={1,2,...,N}, N ~<0o.
States (the amount of inventory on hand at the start of a month):
St={O, 1,2,...,M}, t=l,2,...,N+l.
Actions (the amount of additional stock to order in m o n t h t):
As,,={0,1,2,...,M-s}, t=l,2,...,N.
Expected rewards (expected revenue less ordering and holding costs):
G(s, a) = F(s + a) - O(a) - h(s + a) , t=1,2,... ,N;
(the value of terminal inventory);
rN+I(S , a) = g ( s ) , t= N + 1.
Transition probabilities (see explanation below):
ifM>-j>s+a,
I
p t ( j l s , a) = p,+~,-j
lqs+,~
if M > ~ s + a > ~ j > O ,
if j = O , s + a < ~ M a n d s + a ~ < D , ,
where
q,,~<, = P { D , >>-s -~ a} = Z
d=s+a
Pd"
A brief explanation of the derivation of the transition probabilities follows.

If the inventory on hand at the beginning of period t is s units and an order is
placed for a units, the inventory prior to external demand is s + a units. For
the inventory on hand at the start of period t + 1 to be j > 0 units, the demand
in period t had to have been s + a - j units. This occurs with probability p,~+,, j.
If the demand exceeds s + a units, then the inventory at the start of period
t + 1 is 0 units. This occurs with probability qs+a. Finally the probability that
the inventory level ever exceeds s + a units is 0, since demand is non-negative.
As a consequence of assumption (b) above, the inventory on hand through-
out the month is s + a so that the total monthly holding cost is h(s + a). If
instead, the demand is assumed to arrive at the beginning of a month h(s + a)
is the expected holding cost.
The decision sets consist of all rules which assign the quantity of inventory to
be ordered each month to each possible starting inventory position in a month.
A policy is a sequence of such ordering rules. An example of a decision rule is:
order only if the inventory level is below 3 units at the start of the month and
order the quantity which raises the stock level to 10 units. In month t this
decision rule is given by:
10-- s , s<3,
d,(s)= 0, s1>3.
Such a policy is called in (s, S) policy (See Chapter 12 for more details).
A numerical example is now provided. It will be solved in subsequent
sections using dynamic programming methods. The data for the problem are as
follows: K = 4, c(u) = 2u, g(u) = O, h(u) = u, M = 3, N = 3, f ( u ) = 8u and
P'~
= {
~ if d = 0 ,
~1 i f d = l ,
~ if d = 2 .
The inventory is constrained to be 3 or fewer units and the manager wishes

to consider the effects over three months. All costs and revenues are linear.
This means that for each unit ordered the per unit cost is 2, for each unit held
in inventory for 1 month, the per unit cost is 1 and for each unit sold the per
unit revenue is 8. The expected revenue when u units of stock are on hand
prior to receipt of an order is given in Table 3.1.
Table 3.1
u ~u)
0 0
1 0x~+8x~=6
2 0x+8x+16x~=8
3 0x~+8+16xl=8
Combining the expected revenue with the ordering, and holding costs gives
the expected profit in period t if the inventory level is s at the start of the
period and an order for a units is placed. If a = O, the ordering and holding cost
equals s and if a is positive, it equals 4 + s + 3a. It is summarized in the table
below where an x corresponds to an infeasible action. Transition probabilities

only depend on the total inventory on hand prior to receipt of orders. They are
the same for any s and a which have the same value for s + a. T o reduce
redundant information, transition probabilities are presented as functions of
s + a only. The information in Table 3.2 defines this problem completely.
Table 3.2
r,(s, a) p,(j[s, a)
a=0 a=l a=2 a=3 j=0 j=l j=2 j=3
s=0 0 -1 -2 -5 s+a=0 1 0 0 0
s=l 5 0 -3 x s+a=l 3 ~ 0 0
s=2 6 -1 x x s+a=2 1 ~ 4l 0
s=3 5 x x x s+a=3 0 ~ ~
4. T h e finite h o r i z o n case
This section presents and analyzes finite horizon, discrete time Markov
decision problems. It introduces a concept of optimality and discusses the
structure of optimal policies and their computation. The Principle of Optimali-
ty which underlies the backward induction procedure is shown to be the basis
for analysis. The section concludes with a numerical example.
4.1. O p t i m a l i t y criteria
Each policy yields a stream of random rewards over the decision making
horizon. In order to determine which policy is best, a method of comparing
these reward streams is necessary. Most of the dynamic programming literature
assumes that the decision maker has a linear, additive and risk neutral utility
function over time and uses expected utility as an evaluation function. Con-
sequently, the expected total reward over the decision making horizon is used
for reward stream evaluation and comparison.
The results in this section requires a formal definition of the history of a
Markov decision process. Let /4, denote the history up to epoch t, t =
1, 2 . . . . . N + 1. Define A, = Xses, A s . ,. Then
H 1 = {$1}, (4.1a)
H, = {S1, A t , S 2. . . . . A,__ 1, S,}
={H,_,,A,_t,S,}, t=2 ..... N+I. (4.1b)
Equation (4.1b) shows t h a t / 4 , can be defined inductively in terms of kt,_~.

This means that for h, E H , , h, = (s t, a t, s 2 . . . . . a,_ t, s,) = (h, 1, a,_~, s,) with
h,_t E H,_ t. The history contains the sequence of states and realized actions of
342 M.L. Puterman
the process up to decision epoch t. Clearly, requiring policies to depend on

entire histories will be computationally inhibitive and impractical.
Let 7r = (dl, d 2 . . . . . tiN) be a history dependent policy. That is, for each
t = 1 . . . . . N, d t : Ht---~ A t. When a policy ~r is selected and a history realized,
denote the corresponding history by H 7. Let v~(s) equal the expected total
reward over the planning horizon if policy 7r is used and the system is in state s
at the first decision epoch. It is given by
'rr r 7r
VN(S ) : E=,, r,(X, , dt(H t )) + rN+x(XN+ 1 (4.2)
where E~, s denotes expectation with respect to the joint probability distribution
of the stochastic process determined by 7r conditional on the state of the system
prior to the first decision being s. If the policy is randomized, this distribution
also takes into account the realization of the action selection process at each
decision epoch.
U n d e r the assumption that r,(s, a) is b o u n d e d for (s, t ) ~ S, x As,,, VN(S )
exists and is bounded for each ~ r E / / and each N < ~ . If rewards are
discounted, that is a reward received in a subsequent period is worth less than a
reward received in the current period, a discount factor At-l, 0 < A < 1, is
included inside the summation in (4.2). This will not alter any results in this
section but will be important in the infinite horizon case.
The decision maker's objective is to specify (at decision epoch 1) a policy
7r @ H with the largest expected total reward. When both S t and A~,, are finite
there are only finitely many policies so such a policy is guaranteed to exist and
can be found by enumeration. In this case, the decision maker's problem is that
of finding a 7r* with the property that
7r* 7r
v N (s) = max oN(s ) =- v*N(s), s ~ S1. (4.3)
r~H
The policy ~-* is called an optimal policy and V~v(S) is the optimal value function
or value of the finite horizon Markov decision problem. Theory in the finite
horizon case is concerned with characterizing 7r* and computing v~(s).
When the problem is such that the maximum in (4.3) is not attained, the
maximum is replaced by a supremum and the value of the problem is given by
v[v(s ) = sup v~(s), s ~ S 1 . (4.4)

7rEH
In such cases, the decision maker's objective is to find an e-optimal policy, that
is, for any e > 0, a policy ~r* with the property that
~r* .
v~ (s) + ~ > v?~(s), s ~ s~.
By the definition of the supremum, such a policy is guaranteed to exist. The

problem defined above is a discrete time, finite horizon Markov decision
problem with expected total reward criterion.
4.2. Policy evaluation

In this subsection the basic recursion of dynamic programming is introduced
in the context of computing the expected total reward of a fixed policy. Let
,rr = ( d I , d 2 . . . . , t i N ) be a history remembering policy. For each t define the
expected reward received in periods t, t + 1 . . . . , N + 1 if the history at epoch t
is h, E H~ by
u =(hi)
t = E~,h, r~ X2, H, . (4.5)
--t
The expectation in (4.5) is with respect to the process determined by policy

conditional on the history up to epoch t being h t. Note that u ,rr 7r
1 = v u. The
difference between these quantities is that v N is defined in terms of the entire
future, while u~ is defined in terms of a portion of the future beginning in
period t.
The following algorithm gives an inductive procedure for evaluating the
return of a fixed history dependent policy. It is a basis for several of the results
below. The policy evaluation algorithm for evaluating the return of policy
7r = (d~, d 2 , . . . ,dN) is as follows. For ease of exposition, it is assumed that 1r
is deterministic but it is not required to be Markov or stationary.
The Finite Horizon Policy Evaluation Algorithm.

1. Set t = N + 1 and
UN+I(hN+I) = rN+I(SN#I) for all h N + 1 = ( h N , a N , SN+I) ~ HN+ 1 .
2. Substitute t - 1 for t and compute u[(h,) for each h, ~ H, by
uT(h,) = r,(st, dt(ht)) + ~ P,(jlst, d,(h,))u,\~(h,, dr(h,), j)

jESt+ 1
(4.6)
noting that (h,, d,(ht), j) = ht+ ~ C H,+ 1.
3. If t = 1, stop, otherwise return to step 2.
The idea leading to equation (4.6) for general t is as follows. The expected
value of policy 7r over periods t, t + 1 . . . . , N + 1 if the history at epoch t is h,
is equal to the immediate reward received if action dr(h,) is selected plus the
expected reward over the remaining periods. The second term contains the
product of the probability of being in state ]" at epoch t + 1 if action d,(h,) is
used, and the expected reward obtained using policy ~r over periods t + 1,
. . , N + 1 if the history at epoch t + 1 is h,+ 1 = (h,, d,(h~), j). Summing over
all possible j gives the desired expectation expressed in terms of u ~ t instead of
in terms of the reward functions and conditional probabilities required to
explicitly write out (4.5).
344 M.L. Puterman
This inductive scheme reduces the problem of computing expected rewards

over N + 1 periods to a sequence of N similar 1 period problems having
ir
immediate reward r, and terminal reward u,+ 1. This reduction is the essence of
dynamic programming; multistage problems are reduced to a sequence of
simpler inductively defined single stage problems. The procedure to find optimal
policies that is described below, is quite similar.
That u7 agrees with that defined in equation (4.5) is based on the following
argument which can be made formal through induction. First this algorithm
fixes the value of UN+1 to be the terminal reward that would be obtained if the
history at epoch N + 1 was hN+ 1. Clearly this is the correct value. It next
evaluates U~v for all possible histories h N using equation (4.6) in step 2. This
equation is the basic recurrence. By writing out u N from (4.5) explicitly and
~r
substituting us+ 1 for rN+a, these expressions are seen to agree.
Fundamental to justifying the inductive calculation above is the additivity of
the policy evaluation equation (4.2). This additivity arises from the assumption
of linear utility. Other utility functions have been considered by Howard and
Matheson (1972), Jacquette (1973), Eagle (1975) and Rothblum (1984). White
(1988) surveys the use of variance and other probabilistic criteria.
4.3. The optimality equation and the principle of optimality

This section introduces the optimality equation and investigates its prop-
erties. It shows that solutions correspond to optimal value functions and that
these value functions can be used to determine optimal policies. Proofs of
results appear in Heyman and Sobel (1984, p. 112-124) and Derman (1970, p.
11-17). Hinderer (1970) considers the problem in more generality by allowing
the set of feasible decisions at each epoch to be history dependent.
Let
u*(h,) = sup u~(ht). (4.7)

rEH
The quantity u* is the supremal return over the remainder of the decision
horizon when the history up to time t is h,. When minimizing costs instead of
maximizing rewards this is sometimes called a cost-to-go function (Bertsekas,
1987).
The optimality equations of dynamic programming are the fundamental
entities in the theory of Markov decision problems. They are often referred to
as functional equations or Bellman equations and are the basis for the backward
induction algorithm. They are given by
u,(h,) = sup
aEAst,t
{r,(s,,a)+JESt+t
Z p,(j[s,,a)u,+,(h,,a,j)} (4.8)
for t = 1 , . . . , N and h t U= t t t. When t = N + 1, the boundary condition uN+ 1 =

rN+ I is imposed. In many applications, as well as in Section 7, rN+ 1 is
identically zero. These equations reduce to the policy evaluation equations
(4.6) by replacing supremum over all actions by the action corresponding to a

specific policy. When the supremum in (4.8) is attained, for instance when each
A,t.t is finite, 'max' is used instead.
A solution to the system of equations (4.8) is a sequence of functions
ut : I-I~--->A t, t = 1 . . . . , N , with the property that u N satisfies the Nth equa-
tion, Uu_~ satisfies the ( N - 1)th equation with the u N which satisfies the Nth
equation substituted into the right hand side of the ( N - 1)th equation, etc.
These equations have several important and useful properties:
(a) Solutions to the optimality equations are the optimal returns from period
t onward for each t.
(b) They provide sufficient conditions to determine whether a policy is
optimal.
(c) They yield an efficient procedure for computing optimal return functions
and policies.
(d) They can be used to determine theoretical properties of policies and
return functions.
The following theorem summarizes the optimality properties.
Theorem 4.1. S u p p o s e u t is a solution o f (4.8) f o r t = 1 . . . . , N a n d UN+1 =
EN+ 1. Then
(a) ut(ht) = u * ( h t ) f o r all h, E H t , t = 1, . . . , N + 1, a n d
(b) u l ( s l ) = v~(sl) f o r all s I E S 1.
Result (a) means that solutions of the optimality equation are the optimal
value functions from period t onward for each t and result (b) means that the
solution to the first equation is the value function for the MDP. Note that no
assumptions have been imposed on the state space and the result is valid
whenever the summation in (4.8) is defined. In particular, the results hold for
finite and countable state problems.
Result (b) is the statement that the optimal value from epoch 1 onward is the
optimal value function for the N period problem. It is an immediate con-
sequence of (a). The proof of (a) is based on the backward induction argument;
it appears in the references above.
The next theorem shows how the optimality equation can be used to find
optimal policies when the maximum is attained on the right hand side of the
optimality equation. Theorem 4.3 considers the case of a supremum.
Theorem 4.2. S u p p o s e u*, t = 1 , . . . , N , are solutions o f the optimality equa-

tions (4.8) a n d UN+ 1 = ru+ ~. Define the p o l i c y ~r* = (d~, d ~ , . . . , d~) f o r
t:l,... ,Uby
r,(s,, d * ( h t ) ) + ~ p,+~(jist, d*(ht))u,*~(h,, d*(h,), j)

j~St+ 1
= max IG(s~, a ) + ~ p t ( j [ s , , a)ut*+l(h ~, a, j ) } . (4.9)

aEAst,t [. j~St+l
346 M.L. Puterman
Then:
(a) 7r* is an optimal policy and
17"*
v u (s) = v?~(s), s ~ s,. (4.10)
(b) For each t = l , 2 , . . . , N + l ,
U,7r* (ht)=u*(h,), h, E H t, (4.11)
Equation (4.9) is often expressed as
d*(h,)=argmax{r,(st, a)+ ~, pt(jlst, a)ut*Fl(ht, a, j)}.

aEAs'" J<S'+I (4.12)
The operation 'arg max' corresponds to choosing an action which attains the
maximum on the right hand side of (4.12). It is not necessarily unique.
The theorem means that an optimal policy is found by first solving the
optimality equations and then for each history choosing a decision rule which
selects any action which attains the maximum on the right hand side of (4.9).
When using these equations in computation, the right hand side is evaluated
for all a @ As,,t and the maximizing actions are recorded. An optimal policy is
one which for each history selects any of these maximizing actions.
Part (b) of this theorem is known as 'The Principle of Optimality', and is
considered to be the basic paradigm of dynamic programming. It first appeared
formally in Bellman (1957, p. 83) as:
"An optimal policy has the property that whatever the initial state and initial
decision are, the remaining decisions must constitute an optimal policy with
regard to the state resulting from the first decision."
An equivalent statement that yields further insight appears in Denardo
(1982, p. 15). It can be paraphrased in the language of this chapter as:
There exists at least one policy that is optimal for the remainder of the decision
making horizon for each state at every stage.
The policy rr* has these properties.
In case the supremum in (4.8) is not attained, the decision maker can use an
e-optimal policy which is found as follows.
Theorem 4.3. Let e > 0 be arbitrary and suppose u*, t = 1 . . . . . N + 1, are

solutions o f the optimality equations (4.8) and u~v+l = rN+ 1. Define the policy
~r = (dl, d2, . . . , d N ) to be any policy satisfying
r,(s,, d;(h,)) + p,(jls,, d,(h,))u,+,(h,, d,(n,), j ) + --N

]ESt+I
( )
a@Ast,t J~St+l
Then:
(a) 7r ~ is an e-optimal policy with
~re
O N ( S ) -1- e ~ O ~ r ( S ) , S e S 1 . (4.14)
(b) For each t = l , 2 , . .. , N ,

~E
u, (ht) + ( U - t + 1)(e/U)>1 u*(ht), h, E 11,. (4.15)
The following corollary is an immediate consequence of Theorems 4.2 and

4.3.
Corollary 4.4. There exists an (e-) optimal deterministic policy.
The result follows Since the policy obtained in these two theorems is
deterministic. When the maximum is attained, the policy defined by (4.12) is
optimal, otherwise the policy defined by (4.13) is e-optimah
The following important theorem states that the optimal value functions and
policies depend only on the state of the system at decision epochs and not on
the past, that is, when rewards and transition probabilities are Markov, the
optimal policy depends on the past only throught the current state.
Theorem 4.5. Let ut, t-- 1, . . , N, with UN+ t rN+l ' be solutions of (4.8).
=z
Then if UN+I(hN+I) depends on hN+ 1 only through SN+I:

(a) for each t = 1 , . . . , N, ut(h) depends on h, only through s,, and
(b) there exists (e-) optimal deterministic Markov policies.
A formal proof is based on induction. The key idea in establishing this result
is that if for some n, u,+ 1 depends on the history only through the current
state, then the maximizing action and u n depend on the history only through
the current state.
4.4. The backward induction algorithm

This section presents the backward induction or dynamic programming
algorithm for finding optimal policies and value functions for finite horizon
MDP's. The maximum in (4.8) is assumed to be attained. The results of
Theorem 4.5 allow restriction to deterministic Markov policies.
The Backward Induction Algorithm.

1. Set t = N + l and
u,(s,) = rt(s,) for all s t ~ S,.

2. Substitute t -- 1 for t and compute ut(s,) for each s t C S t by
ut(st)= max {rt(st, a ) + Z Pt(j[st, a)u,+l(s,)}o (4.t6)

aEAst't JESt ~ I
348 M.L. Puterman
D e n o t e by A's,.,, the set of actions as,., satisfying
as,.,=argmax{rt(st, a)+ ~ P,(j[s,,a)u,+l(s~)}. (4.17)

aEAst,t JESt+ 1
3. If t = 1, stop. Otherwise return to step 2.
Comparison of this algorithm to the policy evaluation algorithm of Section

4.2, shows that it accomplishes the following:
(a) It finds sets of actions As*t which contain all actions in As,.t which attain
the maximum in (4.17).
(b) It evaluates any policy which selects an action in As*, for each s, ~ S, for
all t E T.
(c) It computes the expected total reward for the entire decision making
horizon and from each period to the end of the horizon for any policy
described in (b).
An immediate consequence of T h e o r e m 4.1 is that u 1 = v N* and u t = u *t for
all t. Thus this algorithm finds the optimal reward functions and all optimal
policies. A single optimal policy can be found by restricting step 2 to obtain
only one action (instead of all actions) which attains the arg max in (4.17).
4.5. Computationalresults
The backward induction algorithm is used to solve the numerical version of
the stochastic inventory example of Section 3.2. (Since the data are stationary,
the time index may be deleted.) Define u,(s, a) by
ut(s, a) = r(s, a) + ~ p ( j l s , a)u,+l(j) . (4.18)

i~S
Computations proceed as follows.

1. Set t = 4 and u4(s) = r4(s ) = 0, s = 0, 1, 2, 3.
2. Since t 1 , continue. Set t = 3 and for s = 0 , 1 , 2 , 3 ,
u3(s)=maxtr(s,a)+
aEA s L
~ P(j[s,a)u4(j)}=max
jES aEAs
{r(s,a)l
In each state the maximizing action is 0. Thus for all s, Ast3 = {0} and
u3(0 ) = 0, u3(1 ) = 6, u3(2 ) = 6 and U3(3 ) = 5.
3. Since t 1, continue. Set t = 2 and
Uz(S ) = max {Uz(S, a ) } .

aGA s
The quantities U2(S, a), U2(S) and A,* 2 are summarized in Table 4.1 where x~s
denote non-existent actions.
Table 4.1
u2(s, a) u2(s) A*s,2
a=0 a=l a=2 a=3
s = 0 0 1/4 2 1/2 2 2
s = 1 6 4 2 x 61 0
s,= 2 10 4~ x x 10 0
s = 3 10 x x x 10 0
4. Since t ~ 1, continue. Set t = 1 and
Ua(S) = max {ul(s, a)}.

aEA s
The quantities Ul(S, a), ul(s ) and As*,1 are summarized in Table 4.2~
Table 4 . 2
ul(s, a) u,(s) A*s,1
a=0 a=l a=2 a=3
s = 0 2 33/16 66/16 67/16 67/16 3

s = 1 129/16 98/16 99/16 129/16 0
s = 2 194/16 131/16 x x 194/16 0
s = 3 227/16 x x x 227/16 0
5. Since t = 1. Stop.
This procedure has produced the optimal expected total reward function
v~(s) and optimal policy 7r*= (d~(s), d~(s), d~(s)) which are reproduced in
Table 4.3.
Table 4 . 3
s d~(s) d~(s) d~(s) v](s)
0 3 2 0 67/16
1 0 0 0 129/16
2 0 0 0 194/16
3 0 0 0 227/16
The quantity v ~ (s) gives the expected total reward obtained using this policy
when the inventory at the start of month 1 is s units.
This policy has a particularly simple form; if at the start of month 1 the
inventory is 0 units, order 3 units, otherwise do not order, if at the start of
month 2 the inventory is 0 units, order 2 units, otherwise do not order; and do
not order in month 3. This is an example of an (s, S) policy.
350 M.L. Puterman
5. Foundations of infinite horizon models
This section introduces several optimality criteria for infinite horizon M D P ' s
and discusses the relationship between them. It provides an introduction to the
material in Sections 6-8. We assume the data of the problem are stationary and
S is either finite or countable.
5.1. Policy valuation

In a stationary infinite horizon M a r k o v decision process, each policy 7r--
(dl, d 2 , . . . ) induces a bivariate discrete time stochastic process;
{[X t , r(Xt, d~(X, ))], t = 1, 2,
qr "rr ~ . . .
.}. The first c o m p o n e n t X7 is the state of the
system at time t and the second c o m p o n e n t is the reward received if the system
is in X~ and decision rule d,(Xt) is used. W h e n w is randomized, the marginal
distribution of r depends on the distribution on the action space induced by the
decision rule. When 7r is Markov, the above stochastic process is called a
M a r k o v reward process.
Unless otherwise noted rewards are bounded, i.e.,
sup sup lr(s, a)] = M < (5.1)

sES aEA s
To evaluate an infinite sequence of rewards or expected rewards requires a

notion of convergence. Since S is assumed discrete, limits are always t a k e n
pointwise. When S is finite this is equivalent to the stronger notion of
convergence in s u p r e m u m norm however for m o r e general S, this equivalence
does not hold. Fortunately, convergence in n o r m is appropriate for computa-
tional results which are most widely applicable in the finite state case.
Several methods have been proposed to provide a value for a fixed policy in
the infinite horizon case. They are now described.
(a) The expected total reward of policy 7r is given by
v~(s) = E~,s{~_ r(X~, dt(X;~))}. (5.2)

In most applications, the series on the right hand side of (5.2) is divergent o~
not even well defined. W h e n the limit exists, v ~ satisfies
'rr
v~(s) = lim VN(S)
where v N is defined by (4.2) with rN~ , = 0. Several special cases are dis-
tinguished for which this criterion is appropriate. These are introduced in the
next subsection and are the subject of Section 7.
(b) The expected discounted reward of policy ,r is defined by
o;(s) = .a- 1 rl.et, , d,(X,

. . . . . ))} (5.3)
for 0 ~< A < 1. Note that v r = lima,a v~ when the limit exists. Condition (5.1)
ensures that ]v~(s)] < (1 - A)-IM for all s E S and 7r ~ H.
In (5.3), the present value of the reward received in the first period has value
r. This is equivalent to assuming that rewards are received at the beginning of
the period immediately after the decision rule is applied.
(c) the average reward or gain of policy 7r is given by
g (s) = o (s) (5.4)
where VN(S ) is defined by (4.2). In this case, it is not necessary that rN+ 1 equals
zero since averaging removes the effect of the terminal reward. When the limit
in (5.4) does not exist, the limit is replaced by the limit inferior to account for
the worst possible limiting behavior u n d e r policy ~.
The quantity g~ is the average expected reward per period obtained using
policy ~r. Justification for calling it the gain is provided in Section 8.
5.2. The expected total and discounted reward criteria

A plausible objective in infinite horizon problems is to find a policy that
maximizes vL However, even when r is bounded, v ~ may be infinite or not well
defined. To deal with this unsatisfactory situation, two approaches have been
followed; to distinguish cases when this quantity is finite and to choose
alternative optimality criteria that account for the rate at which v u diverges.
When v'~(s) is used to evaluate policies, investigators have distinguished the
following cases;
(a) The positive bounded case (Blackwell, 1967): For each s E S and a C
As, r(s, a) >I 0 and sup,~ n v~(s) < +~.
(b) The negative case (Strauch, 1966): For each s E S and a A ~, r(s, a) ~ 0
and there exists a ~" E / / w i t h v ~ ( s ) > - ~ .
(c) The convergent case (Hordijk, 1974): For each s E S,
sup E~,, [r(X,, d,(X t < +~. (5.5)

~'EH
(d) The discounted case (Howard, 1960): In (5.3), 0 ~< A < 1

These cases are related as follows. The positive case is a special case of the
convergent case. If non-stationary rewards are considered and a transformed
reward function is defined by rt(s, a) = A'-lr(s, a) or restrictions are placed on
the transition probabilities, the discounted c~ise can also be shown to be a
special case of the convergent case. The negative case is distinct because it
allows the quantity in (5.2) to be infinite.
Positive dynamic programming has been applied to optimal stopping prob-
lems and gambling problems (Ross 1983, pp. 76-83). Mathematically, the
positive case is convenient because it ensures that v~(s) is well defined. It cat~
arise in several different ways and is intimately related to the chain structure of
the underlying Markov chains. For instance, if S is finite and under every
policy the chain ends up in a recurrent class in which the rewards are zero, then
v=(s) is bounded.
Negative problems arise in the context of minimization of expected total
costs when immediate costs are non-negative. Changing signs converts all costs
to negative rewards and minimization to maximization. The condition that at
least one policy has v~(s) > - ~ is equivalent to the existence of a policy with
finite total expected cost. Such problems also arise in the context of minimizing
the probability of reaching an undesirable state, minimizing the expected time
to reach a desirable state (Demko and Hill, 1981) and optimal stopping with
minimum expected total cost criterion.
Restricting rewards to be negative ensures that v~(s) is well defined,
however it may be infinite for many policies. The restriction that at least one
policy has v~(s) finite ensures that the expected total reward criteria is useful.
Theoretically this problem is more challenging than the positive case because it
permits policies with infinite rewards.
The discounted case is the most important in economic applications and the
best understood theoretically and computationally. It will be studied in detail
in Section 6. Discounting arises naturally in an economic context when the time
values of the rewards are taken into account. The discount factor A is the
present value of one unit of currency received in the subsequent period so that
v~ is the expected total present value of the income stream obtained using
policy 7r. Allowing A to be non-constant leads to non-stationary problems.
Derman (1970, pp. 31-32) shows that discounting is equivalent to a problem
with expected total reward criteria and a random termination time, 7, that is
independent of the actions of the decision maker and geometrically distributed
with parameter A.
Generalizations of the discounted case include the transient case (VeinotL
1969, Hordijk, 1974, Pliska, 1978 and Whittle, 1983) and problems in which
there is a single absorbing state and the expected time until absorption is
bounded for all policies (Blackwell, 1962, Mine and Osaki, 1968 and van
Dawen, 1986a).
5.3. Optimality criteria

The valuation expressions defined in Section 5.1 lead to natural notions of
optimality. These are sometimes unsatisfactory and additional optimality
criteria have been considered. Several are described below.
A policy 7r* is said to be total reward optimal if
v~(s) >I v~(s) for each s @ S and all zr ~ H o
This concept is applicable in the cases distinguished in Section 5.2. In such

cases the value of the MDP is given by
v*(s) = sup v~T(s). (5.6)

~GH
An optimal policy ~'* exists when
v~*(s) = v*(s) for all s E S .
An equivalent criterion is available when the expected discounted reward is

used to evaluate policies. A policy ~* is said to be discount-optimal if the fixed
A, 0 ~ < A < I ,
v~*(s) >! v~(s) for each s ~ S and all 7r E / / .
In such cases the value of the M D P is
~r
v;(s) = sup v~(s). (5.7)
A discount-optimal policy 7r* exists whenever
v. ( s ) - v ~ ( s ) for a l l s E S .
A policy ~'* is said to be gain optimal or average optimal if
g~*(s) >1g'~(s) for each s ~ S and all 7r ~ 1I.
The gain of the M D P is
g*(s) = sup g~(s). (5.8)

~rE//
When the limits defining g~(s) do not exist, two notions of gain optimality have
been considered (Flynn, 1976, Federgruen, Hordijk and Tijms, 1979, and
Federgruen, Schweitzer and Tijms, 1983). A policy ~-* is said to be average
optimal in the strong sense if its smallest limit point is as least as great as any
limit point of any other policy. T h a t is for each s ~ S,
g'~*(s) = lim inf 1 vN,~. (s)

1
~> lim sup o~(s) for all ~r E IJ
A policy 7r* is said to be average optimal in the weak sense if its largest limit
point is as least as great as any limit point of any other policy. That is, for each
sES,
g"*(s) = lira sup X/vTJ

1 (s)
1
/> lim sup ON(S) for all "o"E 11
N--*~ N
The following simple example motivates the optimality criteria described

below.
Example 5.1. Let S = {1,2,3} and suppose the action sets, rewards and
transition probabilities are as follows. For s = 1, A , = {a, b}, r(s, a) = 1,
r ( s , b ) = O and p ( 2 ] s , a ) = l , p ( 3 [ s , b ) = l . For s = 2 , A s = { a } , r(s,a)=O,
p(1 ts, a) = 1 and for s = 3, A, = {b}, r(s, b) = 1 and p ( l l s , b) = 1. Clearly the
stationary policies which always use action a or b yield average rewards of . It
is easy to see that these policies are average optimal.
This example shows that the average reward criterion does not distinguish
between policies which might have different appeal to the decision maker.
Starting in state 1, policy a with reward stream (1,0, 1 , 0 , . . . ) is clearly
superior to b with reward stream (0, 1, 0, 1 , . . . ) because it provides 1 unit in
the first period which can be put to alternative use. Denardo and Miller (1968)
called a criteria such as the average reward unselective because it depends only
on the tail behavior of the sequence of rewards and does not distinguish
policies with returns which differ only in a finite number of perods. Several
more selective criteria have been proposed. They are based on either
(a) the comparative finite horizon expected total reward as the number of
periods becomes large, or
(b) the comparative expected discounted reward as the discount factor ,~
increases to 1.
Those based on v 7rN are discussed first. Denardo and Miller (1968) called a
policy 7r* overtaking optimality if for each s @ S,
qr * 7r
lim mf 0 N (S) -- ON(S) ~ O for all 7r ~ H , (5.9)
N~
and showed with an example that this criterion is over, elective, that is, there
need not exist an optimal policy with respect to this criterion. The following
criterion (Veinott, 1966) is less selective A policy 7r* is said to be average
overtaking optimal if for each s E S,
lira inf 1 .- for all (5.1o)

n=a
Note that neither of these criteria requires that limN__~ Vu exists.

Sladky (1974) generalized the average overtaking optimal criterion as fol.-
lows. Let v~v.a be the expected total return for policy 7r up to period N and
define VN,n recursively for n ~> 1 by
ON, n = Oj,n_ l o
j=l
A policy 7r* C H is said to be n-average optimal for n = - 1 , 0, 1 , . . . if

--1 "rr* 'rr
limlnfN {VN,,+ 2 - vN,.+2}~>O
for all 7r ~ H. Observe that 0-average optimality corresponds to average

overtaking optimality.
Several optimality criteria have been based on the asymptotic behavior of
the total discounted reward. The following important concept of optimality was
first proposed by BlackweU (1962). A policy 7r* is said to be 1-optimal if for
each s E S there exists a A*(s) such that
v~*(s) - v~(s) >~0 for all 7r C H for A*(s) ~< A < 1 . (5.11)
Such policies are now referred to as Blackwell optimal. Blackwell proposed this
criterion in the context of S finite in which case A* = sup,es A*(s) is attained. In
countable state problems this supremum might equal 1. Dekker (1985) dis-
tinguishes cases when A* < 1 as strongly Blackwell optimal.
Veinott (1969) generalized Blackwell optimality by proposing the following
family of sensitive optimality criteria. A policy 7r* is said to be n-discount
optimal if for each s E S,
lim inf (1 --- h) "[v~*(s) - v~(s)] i> 0 for all 7r E H . (5,12)

AI'1
This criterion unified several optimality criteria based on the expected dis-
counted reward including average or gain optimality, bias and Blackwell
optimality. It has been shown that (-1)-discount optimality is equivalent to
average optimality, 0-discount optimality is equivalent to bias optimality and
~-discount optimality is equivalent to Blackwell optimality. These equivalences
are discussed in more detail in Section 8.9 where the Laurent series expansion
on which these are based is presented.
Blackwell optimality is the most selective of the n-discount optimality
criteria as it implies n-discount optimality for all finite n. It implies gain and
bias optimality. In general, n-discount optimality implies m-discount optimality
for all m < n so that bias optimality (n = 0) is more selective than gain
optimality (n = - 1).
Optimality criteria have also been based on the asymptotic behavior of
policies for finite horizon problems as the horizon gets large. Morton (1978)
calls a policy forecast horizon optimal if it is the pointwise limit of optimal
policies for finite horizon problems. Since the limit need not exist, Hopp, Bean
and Smith (1988) have introduced a weaker criteria, periodic forecast horizon
optimality. A policy is said to be periodic forecast horizon optimal if it is the
limit of a subsequence of optimal policies for finite problems in an appropriate
metric. These two criteria are of particular importance in nonstationary
problems.
356 M.L. Puterman
When the assumption that Ejc s p ( j l s , a)~< 1 is not satisfied the above
criteria are inappropriate. Rothblum (1984) showed that problems for which
Z j c s P ( j l s , a ) > l include Markov decision processes with multiplicative
utilities (Howard and Matheson, 1972) and controlled branching processes
(Mandl, 1967, Pliska, 1976). Optimality criteria in this case are based on
choosing policies which maximize the spectral radius (Bellman, 1957, p. 329).
More general optimality criteria have been proposed by Rothblum and Veinott
(1975).
6. Discounted Markov decision problems
This section analyzes infinite horizon MDP's under the expected total
discounted cost optimality criterion. The optimality equation is introduced and
its fundamental role in Markov decision process theory and computation is
demonstrated. Several algorithms for solving the optimality equation are
presented and discussed; the section concludes with a numerical example and a
discussion of discounted Markov decision problems with unbounded rewards.
Throughout this section it is assumed that the problem is stationary and as
before, S is assumed to be discrete.
6.1. The optimality equation

The optimality or Belhnan equation
v(s) = sup
a~-A s
{r(s,a)+Z Ap(j[s,a)v(j)}, sES,
jES
(6.1)
plays a key role in the theory of Markov decision problems. In vector notation
it can be written as
v = sup {ra + APav } . (6.2)

dED
The supremum in (6.2) is understood to be taken componentwise so that (6.2)

is shorthand notation for (6.1). A formulation based on using (6.2) as the
optimality equation would allow decision sets which incorporate constraints
across states, however, the consequences of such a formulation will not be
explored in this chapter.
When the supremum on the right hand side of (6.1) or (6.2) is attained, for
example when A s is finite, 'sup' will be replaced by 'max'.
Define the operator T : V---~V by
Tv ~- sup {r a + APdV) (6.3)

dED
C h . 8. M a r k o v Decision Processes 357
and for each d E D define the operator T d : V--~ V by
Tar =- r d + h P d v . (6.4)
Comparing (6.2) and (6.3) shows that the optimality equation can be expressed
as v = Tv. Thus, a solution v of the optimality equation is a f i x e d p o i n t of T.
This observation will be fundamental to Section 6.2.
The main properties of the optimality equation to be presented below
include:
(1) If a solution of the optimality equation exists, it equals the value of the
discounted M D P (Theorem 6.3).
(2) The value of the discounted MDP satisfies the optimality equation
(Corollary 6.7).
(3) The solution of the optimality equation is unique (Corollary 6.3).
(4) The optimality equation characterizes optimal policies (Theorem 6.4).
A recursive equation to compute the expected total discounted reward of a
fixed policy r is now developed. Let r = (d~, d e . . . . ) be an arbitrary policy.
Its expected total discounted reward v~(s) was given by (5.3). The expectation
in (5.3) can be expressed in terms of transition probabilities as follows
v 7r
A := ~ ~t " - ~ P "Ir- ' Fd n (6.5)
n=l
2
= r d l -1- A P d l r d 2 -[- A P d l P d 2 r d 3 -}- " "
= rdx + APda(rd 2 + APdfd3 + A2pd2Pafd4 + . .)
= r d l "4- APdiv ~' (6.6)
where 7# = (d2, d3, . . .) and the limit implicit in (6.5) is componentwise. The
relationship in (6.6) holds for arbitrary policies, however if 7r is stationary,
r' = 7r. Denote the stationary policy 7r = (d, d . . . . ) by d. Rewriting the above
relationship yields
d d d
V a = r d + APdO A ~ TdO A . (6.7)
This equation shows that v d is a solution of o = r d + APdtJ. This is extended in

the following.
Proposition 6.1o F o r any stationary policy d, v~ is the u n i q u e solution o f
v = r a + APdv = Tdv , (6.8)

Further,
d
0;~ . (!. . A. P d. ) l r d L A n--1P dn-1 r d (6.9)
n=l
358 M.L. Puterman
The following is the fundamental result about the optimality equation and its
solutions. A proof appears in Blackwell (1962).
Theorem 6.2. Suppose v E V satisfies

v/> ( < ) s u p {r e + APev } . (6.10)
dED
Then v >i (~<)v 2,
Since any solution of the optimality equation must satisfy both of the above
inequalities, it is equal to the optimal value function and consequently is
unique. Alternatively, uniqueness can be established using the contraction
mapping methods of Section 6.3.
Theorem 6.3. If the equation v = Tv has a solution, it is unique and equals v~.
Thus, the value function is a solution of the optimality equation and if

v >~(<-)Tv, v is an upper (lower) bound for v,.
* These inequalities are fun-
damental to demonstrating the convergence of algorithms for solving the
optimality equation and developing computable bounds for v~.
6.2. The existence of optimal policies

In this section it is shown that existence of a solution to the optimality
equation, together with attainment of the supremum in (6.2) is sufficient for
the existence of a deterministic stationary optimal policy. Assume first that the
maximum is attained in (6.2) so the optimality equation is given by
v = max {r a + hPev } = Tv . (6.11)

dED
The following terminology will be used frequently throughout this chapter.

For v E V, a decision rule d o is said to be v-improving if
d o = arg max (r e + hPev } (6.12)

dED
or equivalently,
Taw = rdv -~ hPeov -- max {r e + APev } = Tv (6.13)

dED
Using component notation, d o is v-improving if for all s E S,
r(s, do(s))+ )5 h p ( j l s , do(s))v(s )

j~s
= m a x { r ( s , a ) + ~, Z p ( j l s , a ) v ( j ) } .
a@As jCS
Theorem 6.4. Suppose v * E V satisfies (6.11) and d * E D is v*-improving.

Then
d* 7r
vA = max uA
~rE//
A v*-improving decision rule is often said to be conserving (Dubins and

Savage, 1965). In other words, d* is conserving if
d* = a r g m a x {r e + APdV* } (6.14)
dED
or
Te.v* = Tv*. (6.15)
The following important corollary is a restatement of T h e o r e m 6.4 in this

terminology.
Corollary 6.5, Suppose a conserving decision rule d* exists. Then the de-
terministic Markov stationary policy which uses d* every period is optimal in the
class of all policies.
Conserving decision rules exist whenever the optimality equation has a

solution and the maximum in (6.11) is attained. It will be shown in Section 6.3
that the optimality equation always has a unique solution when 0 ~< A < 1, so
that in the discounted case, attainment of the maximum in (6.11) is equivalent
to the existence of optimal policies in H D. In such cases there is no need to
consider history dependent or randomized policies; it is sufficient to optimize
over the class of stationary deterministic Markov policies.
In terms of the data of the problem, sufficient conditions for the attainment
of the maximum and the existence of optimal policies in H D include:
(a) A s is finite for each s ~ S,
(b) for each s E S, A , is compact, r(s, a) is upper semi-continuous in a and
for each j E S, p ( j I s, a) is continuous in a (Maitra, 1968).
Blackwell (1965) provided an example with S finite (a singleton) and A
countable in which there does not exist an optimal policy, and an example in
which there does not exist an e-optimal policy for sufficiently small e > 0. That
is, there is no policy such that
vA(s )>v*(s)-- e for a l l s C S .
tn that paper he provided conditions under which e-optimal stationary policies

exist.
360 M.L. Puterman
6.3. Value iteration and its variants

This section shows how the theory of contraction mappings on Banach
spaces is used to demonstrate the existence of a unique solution to the
optimality equation, and to analyze the convergence of the value iteration
method for solving the optimality equation. Shapley (1953) introduced the
basic ideas of value iteration in the context of stochastic games; the use of
contraction mappings for MDP's is usually attributed to Denardo (1967); a
good summary and some extensions appear in Federgruen and Schweitzer
(1978).
6.3.1. Theoretical considerations

The results below are based on Denardo's (1967) observations that the
operator T defined in (6.3) is a contraction mapping on V, the space of
bounded real valued functions on S with supremum norm, and that a solution
of the optimality equation is a fixed point of T. The operator T : V---~V is a
contraction mapping because
IITu--T II Allu-- /[ for all u , v ~ V . (6.16)
where A satisfies 0 ~< A < 1.

Since V is a complete normed linear space or Banach Space (L=(S)) the
Banach Fixed Point Theorem (Liusternik and Sobolev, 1961) can be applied to
obtain the following important result.
Theorem 6.6. The operator T has a unique fixed point v* E V and for every
v ~ V, the sequence {v n} defined by v n+l = Tv n converges in norm to v*.
The convergence in Theorem 6.6 is in the norm sense, that is
lira Iio - o*11:0
where II" II is defined in Section 2.2. When S is finite this is equivalent to

pointwise convergence but in the countable case, it is a considerably stronger
result. The following corollaries are the main applications of this theorem in
the discounted case. The first is an application of the above theorem when the
supremum is attained in (6.2) so that
Tv = max {r d + APdV } .
dED
Corollary 6.7. Suppose the supremum in (6.2) /s attained. Then

(a) v~ is the unique solution to the optimality equation,
(b) there exist conserving decision rules, and
(c) the stationary deterministic Markov policy that uses any conserving
decision rule is optimal among the class of all policies'
When the supremum is not attained, only e-optimal policies are possible
This result is summarized as follows.
Corollary 6.8. Suppose the optimality equation is given by (6.2). Then

(a) v~ is the unique solution of (6.2) and
(b) for e > O, the deterministic, stationary policy which uses the decision rule
d ~ defined by
r e, +APe, v~ + e ( 1 - A)~>sup {r e + APav~} (6.17)

dED
is e-optimal among the class of all policies.
6.3.2. The value iteration algorithm

Another important consequence of Theorem 6.6 is the convergence of the
value iteration algorithm for solving the optimality equation. The value
iteration algorithm finds a stationary policy that is e-optimal. In general,
e-optimality occurs if either
(i) the supremum is (6.2) is not attained, or
(ii) the algorithm is terminated in a finite number of iterations.
It will be assumed throughout this section that the supremum in (6.2) is
attained so that the source of e-optimality is the finite termination of the
algorithm. Under this assumption Tv --- maxaE o {r d + APdv }. Otherwise the
stopping criterion in step 3 below requires modification to account for the two
sources of e-optimality.
The Value Iteration Algorithm.

1. Select v E V, specify e > 0 and set n = 0.
2. For each s E S, compute vn+l(s) by
v"+'(s)=max{r(s,a) + Z )tP(jls, a)v"(j)}~

a~As jES
(6.18)
3. If IIv "+' - v"ll < e(1 - A ) / 2 A go to step 4. Otherwise increment n by 1

and return to step 2.
4. For each s E S, set
d~(s)=argmaxtr(s,a)+Z ,~p(j]s,a)v"+'(j)}.
a~As ~ jES
(6.19)
and stop. If the arg max in (6.19) is not unique, any action achieving this
maximum can be selected.
The main step in the algorithm is 2 which gives the recursion v "+~ = Tv ~ in
component notation. Theorem 6.6 guarantees the convergence of the algorithm
to the optimal value ftmction and that the stopping criterion is satisfied in
362 M.L. Puterman
finitely many iterations. When the stopping criterion in step 3 is met, the
stationary policy corresponding to a vn+l-improving decision rule is e-optimal.
Improved stopping rules are discussed in Section 6.7.
Theorem 6.6 ensures convergence of value iteration for arbitrary state spaces
provided that the appropriate norm is selected so that the value functions and
norm are a Banach space. This means that value iteration will converge in
norm if S is finite, countable, compact or Borel. Unfortunately direct im-
plementation of the maximization in (6.20) is only practical when S is finite.
For more general state spaces, the maximization can only be carried out by
using special structure of the rewards, transition probabilities and value
functions to determine the structure of maximizing decision rules. If it can be
established that a property of v n is preserved by induction, for example
unimodality, and that this property ensures that the optimizing decision rule is
of a certain form, i.e., control limit, then if the property of v n holds in the
limit, as a consequence of Corollary 6.7 there exists an optimal stationary
policy with the special structure. This idea has been used extensively in
inventory theory (Chapter 12), replacement theory and queueing control to
determine the structure of optimal policies for infinite horizon problems.
The value iteration algorithm as defined above terminates in a finite number
of iterations when the stopping criteria in step 3 is satisfied. Consequently,
there is no guarantee that the resulting policy is optimal. In special cases,
action elimination procedures discussed in Section 6.7.3 can be used to ensure
termination with an optimal policy. Note that, the policy determined in step 4
is e-optimal in the norm sense, that is
I 1 ~ ~ -- ~211 < ~
The above algorithm is also called successive approximation, backward

induction or dynamic programming.
Results about the convergence of value iteration are summarized in the
following theorem.
Theorem 6.9. Let v ~ V be arbitrary. Then

(a) the iteratives o f value iteration converge in norm to v*2 and
(b) the algorithm terminates in a finite number o f iterates with an e-optimal
policy determined by (6.19).
Some further properties of the iterates of the algorithm are that
It~ ~+' - v~ll = IITo n 7'~Yll ~ Ally" - ~211- (6.20)

This inequality means that the convergence rate of tile algorithm is linear. This
is often referred to as geometric convergence because iterating (6.19) yields
tto" - o211 ~ a~llv - v~l[. (6.21)

When A is close to one, the above bounds suggest that the convergence of this
algorithm will be quite slow. The subsequent subsections discuss other more
efficient methods for solving discounted MDP's.
Using standard arguments, the following error bound for the iterates of
value iteration can be obtained;
Iio n - o211 ~ - U ~ Ilvl - 11 (6.22)
In practice, the error bound below is more useful:
2A"
llv~ ~ - o211 ~ V ~ - X I t ~ 1 - 11 (6.23)
By specifying e a priori and performing one value iteration step, (6.23) can be
used to estimate the number of additional iterations required to obtain the
desired precision.
6.33. Variants o f value iteration

One of the major disadvantages of using value iteration is that it converges
geometrically at rate A. If A is close to 1, solution by this method would require
a large number of iterations. Several authors, including Morton (1971) and
Morton and Wecker (1977), have suggested instead that value iteration be
normalized by either
(a) subtracting an appropriate vector, or
(b) using relative differences at sucessive iterates.
In either case the normalized iterates { w n } satisfy
Iiw "+' - w~ll < A~nll~ - w;ll
where w~ is the normalized optimal value function and a n is the modulus of the
subdominant eigenvalue (the second largest eigenvalue in modulus) of the
transition matrix of the on-improving decision rule. The advantage of this
approach is that if the subdominant eigenvalues for most policies (especially
the optimal one) are considerably smaller than 1, the rate of convergence of
this normalized or relative v a l u e iteration will be considerably faster than that
for value iteration.
When the transition matrices of all policies are irreducible, relative value
iteration can be implemented by selecting an arbitrary state s 0, defining w for
each s E S by
w(s) = vO(s)- v(s0)
and iterating according to
w'+'(s) = Twn(s) - Twn(so) for s C S.

364 M.L. Puterman
When the policies have more general chain structure, a different normalization
which requires identification of recurrent classes can achieve this improved rate
of convergence.
Another modification that will accelerate computations is to use the Gauss-
Seidel variant of value iteration (Hastings, 1969). In it, updated values of
vn+l(s) are substituted into the recursive equation as soon as they are
evaluated. Suppose that the states are labelled sa, s 2 , . . . , s u and are evaluated
in order of their subscripts. Then the Gauss-Seidel iterative recursion is
v"+ltsj) max tr(s,a)+ A [ ~ p(s ilsj,a)v'+l(si)+ ~ p(silsj,a)vn(si)]}.

aEAsj k i<j i~j
As a consequence of a result in Senata (1981, pp. 41-44), the rate of

convergence of Gauss-Seidel iteration will be greater than that of value
iteration. This has been observed empirically by several authors (cf. Porteus
and Totten, 1978).
6.3.4. Convergence of policies and turnpike theory

Results in the preceding sections are concerned with the behavior of the
sequence of values { v"} obtained from the value iteration algorithm. Even if v n
is close to v~, there is no guarantee that the decision rule that attains the
maximum in (6.14) is similar to the optimal decision rule. Brown (1965),
Shapiro (1968) and Federgruen and Schweitzer (1978) investigate the asymp~
totic properties of the sets of v"-improving decision rules.
In practice, infinite horizon models are usually approximations to finite
horizon models with many stages. A question of practical concern is, "Under
what conditions is the optimal policy for the infinite-horizon model optimal for
the finite horizon model?" An answer to this question is provided through
turnpike or planning horizon theory. One such result is the following (Shapiro,
1968).
Theorem 6.10. There e~cists an N* such that for any n >I N*, the optimal
decision in a finite horizon problem when there are n periods' remaining is in
D*, the set of optimal stationary policies for the infinite horizon problem.
A consequence of this result is that if there is a unique optimal stationary

policy for the infinite horizon problem, then it is optimal to use the correspond-
ing decision rule in the first n - N* periods of a problem with finite horizon
n > N*. The optimal policy in the remaining N* periods must be determined
by backward induction. The optimal infinite-horizon strategy is referred to as
the turnpike and it is reached after travelling N* periods on the nonstationary
'side roads'.
Ch. 8. Markov Dec&ion Processes 365
Another interpretation of this result is that it is optimal to use any d C D*

for the first decision in a finite horizon problem in which the horizon is known
to exceed N*. Thus it is not necessary to know the horizon specifically but only
that it exceeds N*. For this reason, N* is often called a planning horizon. A
bound on N* is given in Denardo (1982, p. 176). The concept of planning
horizons has been extended to non-stationary models by Hopp, Bean and
Smith (1988).
6.4. Policy iteration

Policy iteration or approximation in policy space was introduced by Bellman
(1957) and independently by Howard (1960). It is a highly efficient procedure
for solving Markov decision problems. This section discusses this algorithm for
finite state problems with finite and compact action sets. The maximum in (6.2)
is assumed to be attained.
6.4. i. The policy iteration algorithm

The algorithm is as follows.
The Policy Iteration Algorithm (Howard, 1960).

1. Set n = 0 and select an arbitrary decision rule d o E D.
2. (Policy evaluation) obtain yd, by solving
( 1 - aPd,)v = ra. . (6.24)
3. (Policy improvement) Choose d,,+~ to satisfy
rd,+
,
1 Jr- A P d n + l Od
n = md~D
ax {r d + hPdOd. } (6.25)
and set d n~ 1 = dn if possible.

4. If d n + 1 = d n, stop and set d* = d~. Otherwise increment n by 1 and return
to 2.
The above algorithm yields a sequence of policies {dn} and value functions
{re,). It terminates when the maximixing policy in step 3 repeats. This occurs
with certainty in a finite number of iterations in finite state and action problems
but not in compact action problems for which the number of stationary policies
is infinite.
Step 2 is called the policy evaluation step because in it, (6.25) is solved to
obtain the expected discounted reward of stationary policy dn. This equation is
usually solved by Gauss elimination. In step 3, a v e -improving decision rule is
selected. Since the decision rule is not necessarily unique, the condition that
d~+ 1 = dn is included to avoid cycling and ensure termination.
To carry out step 3, the set of all v~ -improving decision rules is required
before selecting a particular decision rule. An alternative specification of the
366 M.L. Puterman
algorithm would retain the entire set of vd-improving decision rules and
terminate when it repeats. This modification is unnecessary since at termina-
tion, v n = v~, so that all conserving decision rules are available.
Alternatively, one might implement step 3 by just finding a decision rule
dn+ ~ with the property that
rdn+l 3- APdn+lUdn ~ rd. 3- A P d o d .
with strict inequality for at least one component. If this specification is used,
the algorithm will still converge in the finite action case, but at a much slower
rate than using the implementation in step 3. If the set of actions is compact,
convergence to v A* is not guaranteed.
6.4.2 The finite action case

This section discusses convergence of the policy iteration algorithm in the
finite state and action case. Fundamental is the result that ( I - A P a ) -~ is a
positive matrix, i.e., if u I> O, ( I - }tPd)-lu ~ O. Consequently the values at
successive iterations of policy iteration are monotone non-decreasing.
Proposition 6.11. Suppose dn+ ~ is van-improving. Then
1)d~+l ~ Odn .
Since there are only finitely many deterministic stationary policies, the
algorithm must terminate in a finite n u m b e r of iterations. At termination,
dn+l = dn, so that
Od. =~ rd.+l + hPd.+tVd. ~-- m a x { r d 3- APdVdn } .

d~D
Thus vdn solves the optimality equation and d* is conserving. Applying

Theorems 6.3 and 6.4 gives the following important result.
Theorem 6.12. Suppose S is finite and for each s C S, A s is finite. Then the
policy iteration algorithm terminates in a finite number o f iterations and the
policy d* is discount optimal.
6.4.3. The compact action case

When the decision set is not finite, the argument used to prove T h e o r e m
6.12 is no longer valid since there is no guarantee that the stopping criterion in
Step 4 will ever be satisfied. In such cases, an analytic approach can be used to
demonstrate convergence. Drawing a parallel to the analysis in Section 6.2,
other issues of concern are:
(a) What is the consequence of initiating the algorithm at step 3 (instead of
at step 1) with a v that is not the return of some policy?
C h . 8. M a r k o v D e c i s i o n Processes 367
(b) What is the rate of convergence of the algorithm?

The development relies on the theoretical foundation of Puterman and
Brumele (1978, 1979) and Puterman and Shin (1978). Define the operator
B : V---~V by
B y - - - m a x {r d +
dED
(APe- I)v} . (6.26)
Then the optimality equation (6.13) can be expressed as
By = 0 (6.27)
and solving the M D P can be regarded as finding a zero of B instead of a fixed

point of T. The key point is that policy iteration is equivalent to using
Kantorovich's generalization of Newton's method for finding a zero of B
(Kantorovich, 1952).
For v C V define the set of decision rules D o to be all d v ~ D satisfying
d o = a r g m a x {r a + ( A P a - I ) v } . (6.28)
dED
Note that the I is (6.28) does not effect the maximization so that D o is the set
of v-improving decision rules.
Proposition 6.13. For u, v E V and any d o E D o,
B u >! B y + (hPdo - I ) ( u - v) . (6.29)
This result follows easily from the definitions of the quantities in (6.29). It is
called the 'support inequality' and is a vector space generalization of the
gradient inequality which defines convex functions in R n. Thus in a generalized
sense, the operator B is 'convex' and hPdo - I is the 'support' of B at v.
Figure 6.1 illustrates the convexity and construction of B y . In the situation
depicted, there are four policies. For each, the function r i + (Pi - 1 ) v is given.
At each v E V, B y is the maximum of these functions. With the exception of
r4 + (P4 - l ) v , all are supports for some v in the illustrated portion of V. Note
By is convex.
The following proposition provides a closed form representation for the
sequence of values generated by policy iteration and is fundamental to this
analysis.
Proposition 6.14. S u p p o s e the sequence {v n } is' obtained f r o m the policy

iteration algorithm. Then f o r any do, E Do,,
Un
v n+l = - (hPao,,- 1)-~Bv" (6.30)
\
' ~ .............. By
V
Fig. 6.1. Construction of By.
Noting the analogy between the support in v and the derivative in R 1,

expression (6.30) is a vector space version of Newton's method. Note also that
dr, in (6.30) corresponds to the decision rule dn+ 1 obtained in step 3 of the
policy iteration algorithm. In R1 if a function f ( x ) is convex decreasing and has
a zero, then starting Newton's method at a point at which the function is
positive ensures that the iterates converge monotonically to the zero. This
observation is the basis for Theorem 6.15 below. A proof can be based on
comparing the iterates of policy iteration to those of value iteration and
showing that if policy iteration and value iteration begin at the same point,
then the iterates of policy iteration are always bounded below by those of value
iteration and above by v~, the solution of the optimality equation which is
assumed to exist. For a more general approach see Puterman and Brumelle
(1978, 1979).
Theorem 6.15. Suppose By ~ 0 and there exists" a unique v* such that By* :: O.
Then the sequence o f iterates {v'} defined by (6.30) converges monotonically
and in norm to the zero o f B, v*.
Since v I is the expected total discounted reward of a v-improving decision

rule, By 1 >I O, and the above conclusions hold for arbitrary v . In terms of the
policy iteration algorithm, this means that:
Ch. 8. M a r k o v Decision Processes 369
(a) The sequence of values generated by policy iteration converges monoto-

nically and in norm to the solution of the optimality equation.
(b) If the policy iteration algorithm is initiated in step 3 with an arbitrary v ,
the conclusion in (a) holds.
These results are not restricted to the finite state, compact action case. They
require only that the maximum be attained in (6.2). Consequently, they apply
in the important case when S is countable, A, is compact, and p(]is, a) and
r(s, a) are continuous in a for each s E S with V equal to the family of bounded
functions of S with supremum norm.
Implementation of policy iteration when the set of decision rules is not finite
requires a stopping rule to guarantee finite convergence. That is, step 3 of the
value iteration algorithm can be incorporated to ensure fnite convergence to
an e-optimal policy. For more details on this point, see Section 6.5. As in the
case of value iteration, when S is not finite, step 3 cannot be implemented
unless special structure is available.
6.4.4. Rates of convergence

When S is discrete, the norm of a matrix H given by
IlUll : sup if] Ihsjl.

sES jES
If S is finite, this supremum is attained.

The following is the main result on convergence rates (Puterman and
Brumelle, 1979).
Theorem 6.16. Suppose {v n } is generated by policy iteration and there exists a

K, 0 < K < ~, such that
Itrdv- e..ll < KIIv - ull f o r all u, vE V . (6.31)
Then
KA
Hvn+l - v ll IIv" 2112 (6.32)
This theorem says that if (6.31) holds, policy iteration converges quadratical-
ly to the optimal value function. This accounts for the fast convergence of
policy iteration is practice. In contrast, value iteration and its variants converge
linearly.
In terms of the data of the problem, sufficient conditions for (6.31) to hold
are that for each s E S:
(a) A s is compact and convex~
(b) p(j]s, a) is affine in a, and
(c) r(s, a) is strictly concave and twice continuously differentiable in a.
When A s is finite, (6.31) need not hold because Pa~ is not unique at several
V E V; however, if a rule such as that in step 3 of the policy iteration algorithm
is used to break ties, the algorithm provides a unique support at each v. Thus
convergence will be quadratic although K might be large. Other conditions
which imply (6.31) can be derived from selection theorems in Fleming and
Rishel (1975).
Corollary 6.17. Suppose {v n} is" generated by policy iteration and
lira IIP o - P o ll = o. (6.33)
Then
lira Ilv"+l - v lt
~__,o~ i~v~-~v-~ = 0 . (6.34)
If a sequence satisfies (6.34) its convergence is said to be superlinear (cf.

Ortega and Rheinboldt, 1970). This corollary says that if (6.33) holds, the
sequence generated by policy iteration converges superlinearly to v~. This
means that the convergence is asmptotically faster than geometric convergence
with any convergence rate constant. Thus under (6.33), policy iteration will
attain the same degree of precision in fewer iterations than value iteration. In
terms of the data of the problem, if the conditions for quadratic convergence
above are relaxed to require only that r(s, a) be strictly concave in a, then
(6.34) holds.
Puterman and Brumelle (1979) also have developed error bounds for
]Ivn ---V~]I in terms of IIv ~ - v[[, but evaluating them is tedious.
6.5. Modified policy iteration

The evaluation step of the policy iteration algorithm is usually implemented
by solving the linear system
( I - APd,)v = rd, (6.35)
by Gaussian elimination. This requires M 3 multiplications and divisions,

where M is the number of states. For large M, exact solution of (6.35) can be
computationally prohibitive. But, it is not necessary to determine this quantity
precisely to identify an improved policy (see Figure 6.2). An approximate
solution to the above equation can be obtained by using successive approximao
tions with a fixed policy.
Morton (1971) suggested this approach: it was formalized by van Nunen
(1976a) and Puterman and Shin (1978). Van Nunen called this algorithm 'value
oriented successive approximations' and regarded it as a variant of value
iteration in which the same decision rule is used for several evaluations.
\
Bv
V
v=0 vO U~ uI V 1 uo U~ V 2 Vx ~ - - ~
V
Fig. 6.2. Illustration of modified policy iteration of order 2.
Puterman and Shin called it 'modified policy iteration' and viewed it as a

variant of policy iteration in which policy evaluation is implemented iteratively.
Their approach is the basis for this section.
The Modified Policy Iteration Algorithm (MPI) of Order m.

1. Select v ~ V, specify e > 0 and set n = 0.
2. (Policy improvement) Choose dn+ 1 to be any vn-improving decision rule.
3. (Policy evaluation)
(a) Set k = 0 and define Uo(S) by
uo(s)=maxlr(s,a)
aEAs ~"
+ j@S
Z Ap(jls, a)v"(j)}. (6.36)
(b) If Iluo - v"ll < e(a .... A)/2A go to 4. Otherwise go to (c).

(c) If k = m, go to (e). Otherwise, compute u~+ 1 by
U~+l(S ) = r(s, d.~ ,(s)) + ~ Ap(jls, d.+l(S))u~(j) . (6.37)
(d) Increment k by 1 and return to (c).

(e) Set v "+1 = u~ and go to step 2.
4. Set d ~ = d,+ 1 and stop.
This algorithm combines features of both policy iteration and value iteration
Like value iteration, it is an iterative algorithm. The stopping criterion used in
step 3b is identical to that of value iteration; when it is satisfied, the resulting
policy is e-optimal. The computation of u 0 in step 3a requires no additional
work because it already has been determined in step 2 when obtaining a

vn-improving decision rule.
Like policy iteration, the algorithm contains an improvement step, step 2,
and an evaluation step, step 3; however, the evaluation is not done exactly.
Instead it is carried out iteratively in step 3c, which is repeated m times. In
vector notation this corresponds to
v "+1 = ( r dn+l
)"+%"
The quantity m can be selected in advance or adaptively. For instance, m can

be chosen so that Ilu~+l - u~H is less than some prespecified tolerance which
can vary with n. Ohno and Ichiki (1987) investigate alternative specifications
for this tolerance and show numerically that fixed low orders of m work well,
adaptive choice is better and considerable reduction in effort is obtained by
using Gauss-Seidel methods in both the improvement and evaluation steps.
The algorithm is based on the following policy iteration representation from
Proposition 6.14:
v ~+1 = v ~ + ( 1 - APa+,)-IBv" . (6.38)
Expanding ( I - APan+I)--1 in its Neumann series representation (Yosida, 1968),

truncating it at m and substituting into (6.38) gives the following representation
for the iterates of modified policy iteration:
v"+l = v ~ + ~ ( AP,~.+)#Bv" . (639~._.__.

k=O
Equation (6.39) shows that modified policy iteration includes value iteration
and policy iteration as extreme cases; modified policy iteration of order 0 is
value iteration and of infinite order is policy iteration. The modified policy
iteration algorithm corresponds to performing one value iteration step in which
the maximum in (6.36) is computed and then m successive approximation steps
with the fixed decision rule dn+ 1. Figure 6.2 illustrates this for modified policy
iteration of order 2.
The quantity v "+1 is the expected total discounted reward obtained by using
the stationary policy dn+ ~ in a problem with finite horizon m and terminal
reward v'. Alternatively, v n+~ is the expected total discounted reward of the
policy which used d,+ 1 for the first m periods, d, for the next m periods and so
forth, in an (n + 1)m period problem with terminal reward v .
The convergence of the algorithm has been demonstrated by Puterman and
Shin (1978) and Rothblum (1979) and can be summarized as follows.
Theorem 6.18. Suppose Bv~ ~ 0 . Then

(i) the sequence o f iterates o f m o d i f i e d policy iteration converge m o n o t o n F
cally and in n o r m to v~,
* and
(ii) the algorithm terminates in a finite number of iterations with an e-optimal

policy.
One might conjecture that the iterates of modified policy iteration of order
m + k (k 1>0) always dominate those for MPI order m when started at the
same initial value. An example of van der Wal and van Nunen (1977) which
appears in Puterman and Shin (1978) indicates that this conjecture is false.
Puterman and Shin (1978) provide the following result regarding the rate of
convergence.
Theorem 6.19. I f
(6.40)
then
Iio n+' -vYll __ / ~ r n + l

limsup iin_ Yll
This result demonstrates the appeal of this algorithm. When the policy is
close to optimal, the convergence rate of the algorithm is close to that of m + 1
steps of value iteration. Computationally this represents a major savings over
value iteration because MPI avoids the maximization at each pass through the
algorithm. Conditions which imply (6.40) were given in the previous section. It
always holds for finite state and action problems in which a rule is used to
uniquely choose the vn-improving policy is step 2.
The MP! algorithm will converge in fewer iterations than value iteration and
at least as many iterations as policy iteration; however the computational effort
per iteration exceeds that for value iteration and is less than that for policy
iteration. Computational results in Puterman and Shin (1978) suggest that it is
a more computationally efficient method for solution of practical Markov
decision problems then either value iteration or policy iteration. Determining
an efficient procedure for selecting m is still an open problem although results
of Dembo and Haviv (1984) provide insight.
6.6, Linear programming

The discounted infinite horizon MDP can be formulated as a linear program--
ming problem (d'Epenoux, 1960). The following discussion gives results based
on that formulation. The development follows Derman (1970) and Kallenberg
(1983).
Theorem 6,2 showed that if'
V~rd+A~V
374 M.L. Puterman
for all d ~ D, then v is an upper bound for the value of the MDP, v h. * Since va*
also satisfies this inequality, it must be the smallest such solution. This is the
basis for the following linear program.
Primal Linear Program.
Minimize ~ aiv(j )
jcs
subject to
v(s) >~r(s, a) + ~hp(jls, a)v(j),

j~s
a ~ A s and s @ S,
and v(s) unconstrained.
The constants aj are arbitrary positive quantities which are assumed without
loss of generality to satisfy ZjEs aj = 1. Its dual is:
Dual Linear Program.
Maximize ~.] E r(s, a)x(s, a)

sES a~A s
subject to
x(j,a)- ~ ~ hp(jls, a)x(s,a)=aj, j~S,

aU-A s s~S a~A s
and x(j, a) >t 0 for all a ~ A j, j E S.
Numerical results adopted from Koehler (1976) which appear in Puterman

and Slain (1978) demonstrate that modified policy iteration is considerably
more efficient than simplex method based linear programming codes for
solving discounted Markov decision problems. Other computational disadvan-
tages of linear programming include the additional effort required to generate
the linear programming tableau and the inability of linear programming
methods to take advantage of the easily available initial basis feasible solution
described in Theorem 6.21 below. However, recent theoretical and computa-
tional advances in solution methods for linear programming could alleviate
these shortcomings. Two clear advantages of the linear programming approach
are that it allows easy inclusion of constraints (Kallenberg, 1983, pp. 72-77)
and it facilitates sensitivity analysis.
The interest in the linear programming formulation is partly theoretical and
partly due to excellent software. Most important results are based on the dual
formulation. They are as follows.
C h . 8. M a r k o v D e c i s i o n Processes 375
Theorem 6.20. The dual problem is always feasible and bounded. For a
randomized stationary policy d, the quantities
x(s,a)= Z otj 2 ;~"P(Xa.=s,d(Xd.)=alX~=j), a E A s , sES,

jEs ,=0 (6.41)
are a feasible solution to the dual problem.

Conversely, if x(s, a) is a solution to the dual problem, then the randomized
stationary policy d defined by
x(s, a~
P(d(s)=a)= , aEA,,sES,
Z x(s,a')
a'~A s
satisfies (6.41).
The quantity x(s, a) defined in (6.41) is the discounted joint probability that
the system is in state s and action a is selected, averaged over initial
distribution { o~/}.
Corollary 6.21. Any basic feasible solution has the property that for each s E S,
x(s, a) > 0 for only one a E A s. If x* is an optimal basic feasible solution, an
optimal deterministic stationary policy is obtained by setting d*(s) = a whenever
x*(s, a) > O.
The matrix defining tile constraints in the dual problem is a Leontief matrix
(Veinott, 1968), that is, each column has exactly one positive entry and for any
non-negative right hand side the linear system has a non-negative solution. A
consequence of this observation is that for any non-negative vector a the dual
linear program has the same optimal basic feasible solution.
The relationship between the simplex algorithm and the dynamic program
ming algorithms is as follows. When the dual problem is solved by the simplex
algorithm with block pivotting, it is equivalent to policy iteration. When policy
iteration is implemented by changing only the action which gives the maximum
improvement over all states, it is equivalent to solving the dual problem by the
usual simplex method. Modified policy iteration is equivalent to a variant of
linear programming in which the basic feasible solution is evaluated by
relaxation instead of direct solution of the linear system.
6. 7. Bounds and action elimination

Tile methods presented in this section can be used to improve the efficiency
of value iteration, policy iteration and modified policy iteration. Fundamental
are bounds on the optimal value function.
376 M , L . Puterman
6. 7.1. Bounds for discounted Markov decision processes

This section presents iteratively determined upper and lower bounds for the
optimal value function. They are of considerable importance computationally
because they can be used to:
(a) provide stopping criteria for the non-finite iterative algorithms,
(b) provide improved terminal value functions when the algorithms have
been stopped and
(c) eliminate suboptimal actions throughout the iterative process.
They are based on the result in Theorem 6.2 that if By/> (~<)0, then
v
References on this topic include MacQueen (1967), Porteus (1971), White

(1978) and Puterman and Shin (1982). Define the real valued functions L and
U f o r v ~ V by
L(v) = rain v(s) and U(v) = max v(s). (6.42)

sES s~S
Let 1 E V denote the function that is 1 for all s ~ S. It is assumed throughout
this section that S is finite so that the extrerna in (6.42) are attained.
The following result is fundamental. A proof is based on applying Proposi-
tion 6.13 and Theorem 6.2. The decision rule d o in the tightest lower bound is
any v-improving decision rule.
Proposition 6.22. For any v ~ V and m >~O,

v + (1 - A)-~L(Bv)I <~v + By + A(1 -- A)-~L(Bv)I
<~ v + ~ ( A P d y B v + Am+l(1 - A)-~L(Bv)I <~ v~

k=O
<~ v + Bv + A ( 1 - A)-~U(Bv)I ~ v + ( 1 - A) -IU(Bv)I . (6.43)
These bounds can be applied during any iterative procedure by replacing v in

(6.43) by the current value v". Results for value iteration and policy iteration
are obtained by setting m = 0 and letting m-->~ respectively in (6.43). For
value iteration, the two tightest lower bounds are identical. For modified policy
iteration and policy iteration, the tightest lower bound becomes vn+l~ < v2; in
the case of policy iteration v n+l Vdn+l.
6. 7.2. Stopping criteria and extrapolations

Value iteration and modified policy iteration are non-finite algorithms and
consequently require conditions to ensure that they stop in a finite number of
iterations with an e-optimal solution. The stopping criterion used in the
algorithms of Sections 6.2-6.4 is too conservative. Empirical evidence indicates
that these criteria require many unnecessary iterations to confirm that a policy
identified early in the iterative process is optimal (Section 6.8). The bounds in
Proposition 6.22 are the basis for more efficient stopping criteria.
The span of v, denoted by sp(v) is defined by
sp(v) = U(v) - L(v) .
The result below is an immediate consequence of Proposition 6.22.
Proposition 6.23. Suppose for v @ V and e > 0 that
(1 - A)
sp(Bv) = U(Bv) - L ( B v ) < - - e. (6.44)
Then
I]v + By + A ( 1 - A ) - I L ( B v ) I - v*~lt < e (6.45)
and for any v-improving decision rule do,
tl%-v ll< . (6.46)
When e is small and (6.44) is satisfied, Bo is nearly constant so that the value
of a v-improving decision rule differs from v by nearly a constant amount. Thus
at the subsequent value, By will again be nearly constant. It can be shown that
this constant must be close to 0. When this constant has been added to v, the
resulting value is closer to v~ as shown by (6.45).
As a consequence of Proposition 6.23, (6.44) provides an alternative stop-
ping rule for value iteration and modified policy iteration. It replaces the
stopping rule in step 3 of the value iteration algorithm or step 3b of the
modified policy iteration algorithm. Note that in both of these algorithms,
By n= Tv n - v" is available prior to testing whether the stopping criteria are
met. Determining U(Bv ~) and L ( B v ' ) requires little additional effort. A l
though policy iteration is a finite algorithm, the above stopping criteria can be
included after step 3, if all that is required is an e-optimal policy.
The quantity v + By + (1 - A) ~AL(Bv)I in (6.48) is called a lower bound
extrapolation by Porteus and Totten (1978) and gives an improved approxima
tion to v~ upon termination of the algorithm. One might conjecture that
convergence of the algorithms would be faster if such extrapolations could be
incorporated at each iteration. Unfortunately this is not the case because the
set of maximizing actions is unaffected by addition of a scalar. Such extrapola-
tions are more useful in semi-Markov decision problems in which transition
matrices do not have equal row sums. Other extrapolations are also available
(Porteus, 1980a).
6.7.3. Action elimination

The above bounds can be used to identify suboptimal actions at each
iteration of any algorithm for solving discounted MDP's. The suboptimal
actions are eliminated from the action set at subsequent iterations and
r(s, a) + ~ Xp(jls, a ) v " ( j )

j~s
must be evaluated for fewer actions. Also, the identification and elimination of
suboptimal actions is the only way of determining an optimal (as opposed to an
e-optimal) policy when using a non-finite iterative algorithm such as value
iteration or modified policy iteration. When all but one action is eliminated in
each state, the stationary policy which uses the remaining decision rule is
necessarily optimal.
Action elimination procedures are based on the following observation of
MacQueen (1967).
Proposition 6.24. I f
r(s, a) + ~ h p ( j l s , a)v*~(y) - oF(s) < 0 , (6.47)
then any stationary policy which uses action a in state s cannot be optimal.
Referring to Figure 6.1 shows why this result holds. In its policy 3 is optimal
and the one-step improvement functions of all other policies are bounded
above by 0 at v h.
Since o F is unknown, the result in Proposition 6.24 cannot be used directly to
identify suboptimal actions. Instead, Proposition 6.24 permits upper and lower
bounds for v~ to be substituted into (6.47) to obtain an implementable
elimination rule.
Theorem 6.25. Suppose v + ~ v F >1 v-. Then if
r(s, a) + ~ hp(j[s, a)v~(j) < v-(s), (6.48)

jcs
any stationary policy which uses action a in state s is suboptimal
Implementations of these bounds for MPI appear in Ohno (1980) and

Puterman and Shin (1982). Algorithms which identify actions that cannot be
chosen in the improvement step at the subsequent iteration have been deo
veloped by Hastings and van Nunen (1977), Hubner (1977) and Puterman and
Shin (1982). These are based on similar principles but their implementation is
considerably more tedious.
6.8. Computational results

In this section, an infinite horizon version of the stochastic inventory mode~
of Section 3.2 is solved using value iteration, policy iteration and modified
p o l i c y i t e r a t i o n a n d t h e r e s u l t s a r e discussed. T h e d a t a a r e t h e s a m e as t h o s e
u s e d in the finite h o r i z o n c a s e in S e c t i o n 4.5, the d i s c o u n t r a t e A is 0.9. T h e
o b j e c t i v e is to d e t e r m i n e t h e s t a t i o n a r y p o l i c y t h a t m a x i m i z e s t h e e x p e c t e d
t o t a l infinite h o r i z o n d i s c o u n t e d r e w a r d . A l l c a l c u l a t i o n s w e r e c a r r i e d o u t using
M D P L A B ( L a m o n d , 1984).
6.8.1. Value iteration

T o initiate t h e a l g o r i t h m , v = 0 a n d e = 0.1. T h e a l g o r i t h m will t e r m i n a t e
with a s t a t i o n a r y p o l i c y t h a t h a s e x p e c t e d t o t a l d i s c o u n t e d r e w a r d w i t h i n 0.1 o f
o p t i m a l . C a l c u l a t i o n s p r o c e e d as in t h e finite h o r i z o n b a c k w a r d i n d u c t i o n
a l g o r i t h m until t h e s t o p p i n g c r i t e r i o n of
e(1 -- A) 0.1 x 0.1

2A - 2x0.9 =0.0056
is satisfied. T h e v a l u e f u n c t i o n s v" a n d t h e m a x i m i z i n g a c t i o n s o b t a i n e d in S t e p
2 at s e v e r a l i t e r a t i o n s a r e p r o v i d e d in T a b l e 6.1.
Table 6.1
Value iteration results
n v"(s) d'(s) A"
s=0 s=l s=2 s=3 s=0 s=l s=2 s=3
0 0 0 0 0 0 0 0 0
1 0 5.0 6.0 5.0 2 0 0 0 6.0000
2 1.6 6.125 9.6 9.95 2 0 0 0 2.8250
3 3.2762 7.4581 11.2762 12.9368 3 0 0 0 1.6537
4 4.6632 8.8895 12.6305 14.6632 3 0 0 0 0.3721
5 5.9831 10.1478 13.8914 15.9831 3 0 0 0 0.0616
6 7.1306 11.3218 15.0383 17.1306 3 0 0 0 0.0271
7 8.1690 12.3605 16.0828 18.1690 3 0 0 0 0.0061
10 10.7071 14.8966 18.6194 20.7071 3 0 0 0
15 13.5019 17.6913 21.4142 23.0542 3 0 0 0
30 16.6099 20.7994 24.5222 26.6099 3 0 0 0
50 17.4197 21.6092 25.3321 27.4197 3 0 0 0
56 17.4722 21.6617 25.3845 27.4722 3 0 0 0
57 17.4782 21.6676 25.3905 27.4782 3 0 0 0
58 17.4836 21.6730 25.3959 27.4836 3 0 0 0
E s t i m a t e s b a s e d o n t h e e r r o r b o u n d in (6.23) i n d i c a t e t h a t 68 i t e r a t i o n s a r e
r e q u i r e d to o b t a i n a 0 . 1 - o p t i m a l policy. I n fact, using t h e s t o p p i n g c r i t e r i o n
a b o v e l e a d s to t e r m i n a t i o n a f t e r 58 i t e r a t i o n s w h e n II v 58 - v 57[[ = 0.0054. T h e
0 . 1 - o p t i m a l s t a t i o n a r y p o l i c y is d E = (3, 0, 0, 0). This is t h e p o l i c y w h i c h o r d e r s
o n l y w h e n t h e s t o c k level is 0, a n d in t h a t case o r d e r s 3 units. O b s e r v e t h a t t h e
o p t i m a l p o l i c y was first i d e n t i f i e d at i t e r a t i o n 3 b u t t h e a l g o r i t h m d i d n o t
t e r m i n a t e until i t e r a t i o n 58.
The stopping criterion of Section 6.7.2 yields
1-a 0.1
A n =- U(Bv n) - L(Bv n) < ~ e = ~..9 x 0.1 = 0.0111. (6.49)
To apply this stopping rule, note that By n = v n+a - v". Observe from the last
column of Table 6.1 that when using this stopping rule, the algorithm termi-
nates with a 0.1-optimal policy after only 7 iterations.
6.8.2. Policy iteration

To start the policy iteration algorithm choose the myopic policy, i.e., that
which maximizes the immediate one-period reward r(s, a). The algorithm then
proceeds as follows:
1. Set d = ( 0 , 0 , 0 , 0 ) a n d n = 0 .
2. Solve the evaluation equations obtained by substituting the transition
probabilities and rewards corresponding to policy d o into (6.24) to obtain
v = (0, 6.4516, 11.4880, 14.9951).
3. For each s the quantities
r(s, a) + ~ p(jls, a)v(j)

j=O
are computed for a = 0 , . . . , 3 - s and the actions which achieve the maximum
are placed into Ao, s. In this example there is a unique maximizing action in
each state so that
Ao, 0 = {3}, AS.a = {2}, Ao, 2 = (0}, Ao, 3 = {0}.
4. Since d ( 0 ) = 0, it is not contained in Ao,0, so continue.

5. Set d ~ = (3, 2, 0, 0), n = 1 and return to the evaluation step.
The detailed step by step calculations for the remainder of the algorithm are
omitted. The value functions and corresponding maximizing actions are pre-
sented below. Since there is a unique maximizing action in the improvement
step at each iteration, A*,s is equivalent to dn(s) and only the latter is
displayed, in 'Fable 6.2.
The algorithm terminates in three iterations with the optimal policy d * =
(3, 0, 0, 0). Observe that an evaluation was unnecessary at iteration 3 since
d 2= d 3 terminated the algorithm prior to the evaluation step. Unlike value
iteration, the algorithm has produced an optimal policy as well as its expected
total discounted reward v 3 =v~.* Note in this example that the e-optimal policy
found using value iteration is optimal but it could not be recognized as such
without using action elimination.
Table 6.2
Policy iteration results
n v(s) d"(s)
s=0 s=l s=2 s=3 s=0 s=l s=2 s=3
0 0 6.4516 11.4880 14.9951 0 0 0 0

1 10.7955 12.7955 18.3056 20.7955 3 2 0 0
2 17.5312 21.7215 25.4442 27.5318 3 0 0 0
3 x x x x 3 0 0 0
6.8.3. Modified policy iteration

T h e following illustrates the application of modified policy iteration of o r d e r
5. T h e first pass through the algorithm is described in detail; calculations for
the r e m a i n d e r are p r e s e n t e d in tabular form.
1. S e t v = ( 0 , 0 , 0 , 0 ) , n = 0 a n d e = 0 . 1 .
2. O b s e r v e that
3
r(s, a) + ~, Ap(jls, a)v(j) = r(s, a)
j=o
so that for each s, the m a x i m u m value occurs when a = 0. T h u s A * , = {0} for

s = 0 , 1 , 2 , 3 and d" = ( 0 , 0 , 0 , 0 ) .
3.0. Set k = 0 and Uo = (0, 5, 6, 5).
3.1. Since Ilu - vll = 6 > 0 . 0 0 5 6 , continue.
3.2. Since k = 0 < 5, continue. C o m p u t e u by
3
u(s) = r(s, d(s)) + ~ ,~p(jls, d(s))u~(j)
j=o
3
= r(s, O) + ~ Ap(jls, O)u~(j)
j-O
to obtain u = (0, 6.125, 9.60, 10.95).

3.3. Set k = 2 and r e t u r n to 3.2.
T h e loop is r e p e a t e d 4 m o r e times to evaluate u :, u3,

0 u 40 and u 0s . T h e n v i is
set equal to u and the m a x i m i z a t i o n in step 2 is carried out. T h e resulting
iterates a p p e a r in T a b l e 6.3.
In Step 3.1, following iteration 11, the c o m p u t e d value of u is (17.4976,
21.6871, 25.4100, 27.4976), so that Ilu -vl'll--0.0038
and the policy
(3, 0, 0, 0) is e - o p t i m a l with e = 0.1.
If instead the stopping criteria of Section 6.7.2 is applied the algorithm
t e r m i n a t e s after 4 iterations. T h e last c o l u m n of 'Fable 6.3 gives zl n as defined
in (6.50).
Table 6.3
Iterates of modified policy iteration
n v"(s) d(s) A"
s=0 s=l s=2 s=3 s=0 s=l s=2 s=3
0 0 0 0 0 0 0 0 0
1 0 6.4507 11.4765 14.9200 3 2 0 0 6.0000
2 7.1215 9.1215 14.6323 17.1215 3 0 0 0 4.9642
3 11.5709 15.7593 19.4844 21.5709 3 0 0 0 2.6011
4 14.3639 18.5534 22.2763 24.3639 3 0 0 0 0.0022
5 15.8483 20.0377 23.7606 25.8483 3 0 0 0
10 17.4604 21.6499 25.3727 27.4604 3 0 0 0
11 17.4738 21.6833 25.4062 27.4938 3 0 0 0
6.8.4. Discussion of results

Using the norm based stopping criterion, value iteration required 58 itera-
tions to obtain a 0.1-optimal solution while using the span based criterion
(6.49), value iteration required 7 iterations and modified iteration of order 5
required 4 iterations. Clearly the span based stopping rule greatly improved the
efficiency of each of these procedures.
At each pass through the value iteration algorithm, a maximization over the
action set was required in each state. This means that the quantity
r(s, a) + ~ Ap(jls, a)v"(j)

j~s
had to be computed for each action at each iteration. Thus in problems with
large action sets, this step of the algorithm would be time consuming. Modified
policy iteration performs this maximization far less frequently so that one
would expect a considerable improvement in efficiency, especially when the
maximizing actions do not change often. Tables 6.1 and 6.3 show that after the
third iteration there was no change in the on-improving decision rule so that
the value iteration algorithm performed many unnecessary maximizations.
Based on results using the span based stopping rule, value iteration required 8
maximizations while modified policy iteration of order 5 required 5. While not
a dramatic savings in this small problem, it illustrates the potential for
considerable improvement in larger problems.
Note also that the sequence of on-improving decision rules obtained using
value iteration and MPI was different. Comparing Tables 6.1 and 6.2 shows
that those of modified policy iteration and policy iteration were identical. This
is to be expected, since modified policy iteration does an approximate policy
evaluation at each pass through the algorithm which is usually adequate to
identify an improved decision rule. In the example above, when the span based
stopping rule was used with modified policy iteration, MPI required 5 maximi-
zations while policy iteration required 4. Thus because Gaussian elimination
was avoided, MPI probabily required fewer multiplications and divisions. In
this small problem, such improvements are unimportant but in problems with
large state spaces, MPI can be considerably more efficient than policy iteration.
This was demonstrated numerically in Puterman and Shin (1978). An open
question in implementing modified policy iteration is how best to choose the
order which can be varied from iteration to iteration.
Calculations using action elimination are not given here. The reader is
referred to Puterman and Shin (1982) and Ohno and Ichiki (1987) for results
using such methods. In Puterman and Shin, one-step ahead action-elimination
algorithms were shown to be most efficient and are recommended for use
together with a Gauss-Seidel version of modified policy iteration whenever
solving a large discounted MDP. It is expected that increased efficiency can be
attained by incorporating other methods of Section 6.3.3.
6.9. Unbounded rewards and countable state spaces

Countable state spaces are natural settings for applications of MDP's to
queueing control and inventory management. In such applications, rewards or
costs are often assumed to be linear in the state variable and consequently are
unbounded. Almost all of the previous results in this section implicitly assumed
that r and v were bounded. Most importantly, existence of a solution to the
optimality equation was based on the result that Tv defined in (6.2) is a
contraction mapping on the set of bounded real valued functions on S. This
subsection discusses a modified approach when the boundedness requirement is
removed.
Harrison (1972) is the first author to explicitly deal with this issue. Sub
sequent contributions for discounted problems include Lippman (1975), van
Nunen (1976b), Wessels (1978) and van Nunen and Wessels (1978). Most of
these results are based on modifying V so that it includes a sufficient number of
unbounded functions. The state space is assumed to be countable in this
section.
Let w be an arbitrary positive real valued function on S (for example, when
S is a subset of (0, ~), w(s) = s). Define the weighted supremum norm I1" tlw for
real valued functions v on S by
Ilvll w = s u p w(s) llv(s)i (6.50)

sES
and let V w be the Banach space of real valued functions v on S satisfying

IlVllw< .
For a matrix H V w-~ Vw the above norm induces the matrix norm given by
HHH w = sup w(s)-1 ~

s~S j~S
Ih~jlw(j)
where h,j is the (s, ])th component of H.

384 M.L. Puterman
The following two conditions are adapted from Lippman (1975):

(1) There exists a constant M such that for all d E D,
Ilrallw M. (6.51)
(2) There exists a finite non-negative constant L such that for all d E D,
Paw ~ w + L . (6.52)
If (6.51) and (6.52) holds with L = 0, then T defined in (6.3) is a contraction

mapping on V w so that the results of Section 6.3 apply. Unfortunately, the
requirement that L = 0 is too restrictive for most applications. Lippman
showed that with L > 0, that T is an n-stage contraction on Vw, so that results
of D e n a r d o (1967) can be applied to obtain the following generalization of
earlier results.
Theorem 6.26. Suppose that (6.51) and (6.52) hold. Then the optimality
equation
Tv = sup {r a + hPdv } (6.53)

dED
has a unique solution v~ in V w which can be found by value iteration

Furthermore, if the supremum in (6.53) is attained, there exists an optimal
deterministic stationary policy.
The condition that Ilrallw~M is equivalent to Jr(s, a)[ <~Mw(s) for all
a ~ A s and s ~ S which implies that r grows at most rate w in s. Based on this,
a suitable choice for w is
w(s) = sup [r(s, a)[ (6.54)

aEA s
in which case M can be chosen to be 1 in (6.52).

The condition that Paw <- w + L is equivalent to
E~,aw(X,) = ~ p ( j [ s , a ) w ( j ) <~ w(s) + L (6.55)

j~S
for all a E A s and s E S where X 1 is the random variable with values in S and
probability distribution given by p ( - I s , a). This means that when w is chosen so
that (6.52) holds, under any decision rule, the expected reward in the next
period cannot exceed w by more than L units. This condition places restrictions
on allowable transitions but does allow transitions to distant states with small
weighted probability.
A countable state version of the inventory model of Section 3.3 can be
shown to satisfy (6.51) and (6.52) under the reasonable assumption that the
Ch. 8o Markov Decision Processes' 385
maximum quantity ordered in any period is bounded. Lippman (1975) shows

that these conditions are satisfied in certain machine replacement, optimal
consumption and queueing control problems.
7. Undiscounted Markov decision problems--I
The next two sections are concerned with MDP's without discounting. In this
section the focus is problems with expected total reward criterion. Assumptions
are imposed so that the expected total reward is well defined for all policies
and finite for some policies. Implicit in these formulations are restrictions on
the reward functions. When the expected total rewards is unbounded or not
well defined, for all policies, the average and sensitive optimality criteria of
Section 8 are of greater practical significance.
In Section 6, a complete theory for discounted problems was presented.
Crucial to this theory was the existence of a discount factor A < 1 which
ensured that
(a) the operator T defined in (6.3) was a contraction,
(b) ( I - A P a ) -1 existed and was non-negative for all d, and
(c) for bounded v,
lirn A nP =n v =
lira n
n ~ n----~ c~ ,
As a result of (a), T has a unique fixed point which could be obtained by

value iteration. Because of (b), v A is well defined, any V~-lmprovmg decision
rule has an expected total discounted reward at least as great as that of 7r, and
bounds on the optimal total discounted reward are available. Condition (c) is
used to show that v~ is a solution of the optimality equation.
The focus of this section will be the expected total reward criterion (5.2)
That is, policies will be compared on the basis of
v~(s) = E, s r ~ , d, ,
which is the expected total discounted reward with A .... 1. Without additionai
assumptions, there is no guarantee that the above limit exists. Also, (a)-(c)
above are not valid. This situation necessitates restricting attention to cases
when v ~ is well defined in addition to using different methods of analysis.
Strauch (1966) and Blackwell (1967) distinguished the positive and negative
cases in which the expected total discounted reward is well defined. In the
positive case all rewards are non-negative and in the negative case all rewards
are non-positive. These two cases are distinct because maximization is done in
each so that one cannot be obtained from the other by sign reversal.
The key mathematical properties that will be used in this section are that v"
is well defined, the optimal return operator is m o n o t o n e and monotone
386 M.L, Puterman
sequences are convergent. The behavior of P~v for large n is crucial and
results below about its properties will indicate why the positive and negative
cases are distinguished. To paraphrase Strauch (1966), analysis in the total
reward cases is based on improving the tails while that in the discounted case is
based on ignoring the tails.
Several authors including Hordijk (1974), Schal (1975), van Hee (1978), van
der Wal (1984) and van Dawen (1986a, 1986b) have analyzed MDP's with total
reward without distinguishing the positive and negative cases; however, in this
section they will be discussed separately. The state space will be assumed to be
general except where noted.
7. l. The positive bounded case

An M D P is said to be positive if r(s, a) >! 0 for all a @ A s and s ~ S and
bounded if there exists an M < ~ such that
c~
UCr = 2 P~t - I r e a M (7.1)
t=l
for all 7r E / L For some results, the uniform boundedness assumption on v ~'
can be relaxed (van H e e , 1978).
When S is finite and Zje s p(jls, a) = 1 for each a E A s and s E S, (7.1)
holds for each ~r if and only if r = 0 on the recurrent classes of 7r. W h e n S is
infinite, r = 0 at each positive recurrent state implies (7.1).
The objectives in analyzing the positive bounded case are to characterize the
value of the MDP,
v*(s) : sup v~(s),
to determine when an optimal or e-optimal policy 7r* exists, and to character-

ize its form. Motivated by Blackwell (1967), several authors have investigated
the positive case including Strauch (1966), Ornstein (1969) and Hinderer
(1970); other references appear above.
7.1.1. The optimality equation

The optimality equation in undiscounted MDP's with expected total reward
criteria is given by
aEA s L jES
It can be expressed in matrix-vector notation by
v = max {r e + Pdv} ~ Tv (7.2)

d~.D
where conventions of previous sections regarding the maximum are assumed to

hold. The operator T will be referred to as the optimal return operator. Since
the returns of all policies are non-negative, it will be expeditious to analyze
(7.2) on V + -- {v E V: v I> 0}. The optimal return operator T maps V + into
V + and the optimality equation corresponds to a fixed point equation for T on
V +"
The result below is analogous to T h e o r e m 6.2 in the discounted case. Its
proof depends explicitly on the assumed positivity of v.
Theorem 7.1. S u p p o s e there exists a v E V + f o r which v >i Tv. Then v >I v*.
In contrast to the discounted case, the converse of T h e o r e m '7.1 that v <~ To

implies v ~< v*, does not hold without further assumptions. The following
simple example demonstrates this and also motivates several other results in
this section.
Example 7.2. Let S : {1,2}, A 1 {a, b}, A 2 {a}, r(1, a) : 0 , r(1, b) : 1,

= =
4 2 , a) = 0, p ( l l 1, a) -- 1, p(211, b) -- 1, p(212, a) = 1 and p ( j l s , a ' ) = 0 other-

wise. There are two available decision rules d and e given by d ( 1 ) = a,
d(2) = a, e(1) = b and e(2) = a. Under d, both states 1 and 2 are recurrent,
while under e, 1 is transient and 2 is recurrent. The optimality equation is
v(1) = max {v(1), 1 + v(2)} and v(2) = v ( 2 ) .
It is easy to see that decision rule e is optimal, r e ( l ) = 1, re(2) = 0 and v e = v*.

Let v ( 1 ) = v ( 2 ) = 2 and observe that T v ( 1 ) = 3, T v ( 2 ) = 2. Thus Tv >! v does
not guarantee that v ~< v* without further conditions.
This example also illustrates another important feature of the positive case,
that the optimality equation does not possess a unique solution, since v* + c
satisifies Tv* = v* for any constant vector c.
The following result adapted from Schal (1975) and van Dawen (1986a, b)
provides sufficient conditions for the above implication to be valid.
Theorem 7.3 S u p p o s e there exists a v @ V ~ such that v ~< To. Then v <~ v* if
l i m i n f E , ~, { v ( X'tr
N)} = l i m i n f P ~Nv = 0 f o r all r r ~ H . (7.3)
N--+~ N--+~
Sufficient conditions for (7.3) to hold in the positive case include

(a) v = 0 or
(b) all policies are aperiodic and v(s) = 0 on all states that are in the positive
recurrent class of some policy.
Combining the above two theorems gives an analogous result to Theorem
6.3 in the discounted case.
388 M.L. Puterman
Theorem 7.4. Suppose there exists a v E V +, such that v = Tv and (7.3) holds.
Then v = v *.
The following theorem (Blackwell, 1967) contains analogous results to those

of Theorems 6.3 and 6.9 in the discounted case. Among other things, it states
that value iteration starting from v = 0 converges to a solution of the
optimality equation and it provides an alternative identification of the optimal
solution to that in Corollary 7.4.
Theorem 7.5. The optimal return v* is the smallest solution o f the optimality
equation To = v in V + and v* = limn_.= Tn0.
Z 1.2. Identification and existence o f optimal policies

In the discounted case, Theorem 6.4, showed that a stationary policy based
on any v*-improving (conserving) decision rule is optimal. That is not true in
the positive case as illustrated by Example 7.2 in which both decision rules are
v*-improving but d is sub-optimal. The following weaker result is available.
Corollary 7.6. The stationary policy based on d' is optimal if and only if v d, is a
fixed point o f T.
Van Dawen (1986b) refers to the decision rule described in the above
corollary as unimprovable, that is, d' is unimprovable if Tv a'= v d'. Corollary
7.6 is equivalent to: d' is optimal if and only if d' is unimprovable.
In the discounted case it was shown that an optimal stationary policy existed
under very weak assumptions. In particular, any condition which guaranteed
the existence of a conserving decision rule was sufficient to ensure the existence
of a stationary optimal policy. Unfortunately, this is not the situation in the
positive case as demonstrated by examples of Strauch (1966) and Blackwell
(1967).
The following theorem is based on a subtle argument allowing generalization
of results in the discounted case.
Theorem 7.7. Suppose that S is finite and for each s ~ S, A , is finite. Then there
exists an optimal stationary policy.
When S is more general, different notions of optimality and classes of

optimal policies have been considered by Ornstein (1969) and Strauch (1966)o
Related results have been obtained by Derman and Strauch (1966), van Hee
(1978), van der Wal (1981, 1984) and van Dawen (1986a,b).
7.I.3o Computational methods in the positive case
Value iteration: Theorem 7.5 showed that a solution of the optimality equation
can in principle be obtained by the value iteration scheme v " = Tv n-1 where
v = 0. Unfortunately, stopping rules, bounds and action elimination criteria
are not available for this procedure so its significance is primarily theoretical.
This convergence result can also be used to characterize the structure of an
optimal policy (Kreps and Porteus, 1977) so that search for optima can be
carried out in a smaller class of policies with special form.
Policy iteration: Strauch (1966) provides a countable state example in which

the decision rule determined in the improvement step of policy iteration has a
strictly smaller return than the initial policy. Consequently for countable state
problems, policy iteration cannot be used to determine an optimal policy.
When S is finite, Example 7.2 shows that choosing any va"-improving policy
in the improvement step does not guarantee a policy with greater return, since
decision rule d is re-improving but va(1) < r e ( l ) . However if the same decision
rule is va-improving at two successive iterations, it is unimprovable and
consequently by Corollary 7.6, it is optimal. In the discounted case, the rule
"set dn+ 1 = d, if possible" was included to avoid cycling, here it is essential to
ensure that policy iteration yields monotone increasing value functions (cf. van
Dawen, 1986b).
Another complication in implementing policy iteration is that that v =
ra + Pay does not necessarily have a unique solution (cf. Example 7.4).
Applying T h e o r e m 7.5 with A s replaced by (d(s)) characterizes v a as the
minimal solution of v = r a + Pay in V +.
Thus in the finite state and action case, policy iteration can be used to obtain
an optimal policy in the positive case provided that in the evaluation step the
minimal positive solution of the evaluation equation is selected, the rule "set
d, +1 = dn if possible" is followed in the improvement step and it is started from
the return of a stationary policy instead of an arbitrary v 0 @ V +.
Linear programming: Linear programming for finite state and action positive
MDP's has been studied by Kallenberg (1983). The formulation is similar to
that in the discounted case; the constraints of the primal problem are identical
to those in the discounted case with A set equal to 1. However, the additional
condition that v t> 0 is added and consequently the equality constraints in the
dual problem are replaced by inequalities. That v* is a solution of the primal
problem follows from T h e o r e m 7.5. The dual problem is feasible because
x(s, a) = 0 satisfies its inequality constraints. If the dual program has a finite
optimum, then the problem can be solved by the simplex method and an
optimal stationary policy is given by
a if x(s, a) > 0 and s E S* ,
d(s) arbitrary if s E S -- S * ,
where S* = {s E S: x(s, as) > 0 for some a ~ As). W h e n the dual is unbounded
a more complicated procedure is provided by KaUenberg.
Z2. The negative case

An M D P is said to be negative if r(s, a) <<-0 for all a ~ A s and s ~ S. Under"
this assumption, v = is well defined but might equal - ~ . Without the further
3913 M.L. Puterman
assumption that v '~ is finite for at least one policy 7r ~ H, v = = --oo for all 7r ~ / /
so that all policies are equivalent under the total expected reward criterion. If
this is the case, the average reward criterion can be used to discriminate
between policies.
The most natural setting for the negative case is cost minimization with
non-negative costs. In such problems the reward function r is interpreted as
negative cost; maximization of expected total reward corresponds to minimiza-
tion of expected total cost. Contributors to theory in the negative case include
Blackwell (1961), Strauch (1966), Kreps and Porteus (1977), Schal (1978),
Whittle (1979), (1980a and b), Hartley (1980), Demko and Hill (1981) and van
Dawen (1985).
Analysis in this section will parallel that in Section 7.1 as closely as possible.
S will be arbitrary except where noted.
7.2.1. The optimality equation

The optimality equation for the negative case is given in (7.2). Since r ~<0,
solutions will be sought in V-, the set of non-positive real valued functions on
S. Unlike V +, elements of V- are not required to be bounded.
The following results are analogous to Theorems 7.1, 7.3 and 7.4 in the
positive case.
Theorem 7.8. S u p p o s e there exists a v E V - satisfying v <~ Tv. Then v <~ v*.
For the implication with the inequalities reversed, additional conditions are
required as illustrated by Example 7.11 below. One such sufficient condition is
given in the following result.
Theorem 7.9. S u p p o s e there exists a o ~ V ~ such that Tv ~ v and
E 7r
lira sup ,~{v(XN) } = lira sup P =N v = O. (7.4)
N---~ ~ N - - - * oo
Then v >I v*.
Sufficient conditions for (7.4) to hold are identical to those which imply (7.3)
in the positive case. Combining Theorems 7.8 and 7.9 gives the following
important result.
Theorem 7.10. S u p p o s e there ex&ts a v C V such that v = Tv and (7.4) holds.

Then v = v*.
The following example of Blackwell (1960) provides a counterexample to

many conjectures in the negative case.
Example 7.11, Let S = {1, 2}~ A 1 = {a, b}, A z = {a}, r(1, a) = 0 , r(1, b) = - 1 ,
r(2, a) = 0, p ( l [ 1 , a) = 1, p(211, b) = 1,p(212, a) = 1 a n d p ( j l s , a') = 0 other~
wise. There are two available decision rules, d and e, given by d ( 1 ) = a,

d(2) = a and e(1) = b, e(2) = a. The optimality equation is
v(1) = max {v(1), - 1 + v(2)} and v(2) = v(2).
Decision rule d is optimal and v a ( 1 ) = 0 , Va(2)=0 and v d = v *. Choosing

v(1) = v(2) = - 1 implies that Tv(1) = - 1 and Tv(2) = - 1 , so that Tv <~v but
v ~< v*. Note that (7.4) is not satisfied in this example. Also, the optimality
equation does not have a unique solution, since any v E V - satisfying v(1)/>
- 1 + v(2) satisfies Tv = v.
The following result is the analog of Theorem 7.5 in the positive case. It
provides an alternative characterization of the optimal return to that of
Theorem 7.10.
Theorem 7.12. The optimal expected total reward v* is the largest solution of
Tv = v in V - and v* = limn_~= Tn0.
Whittle (1979, 1980a,b) provides some generalizations of this theorem.
7.2.2. Identification and existence of optimal policies

In contrast to the positive case, v*-improving (conserving) decision rules are
optimal. This has important consequences for the existence of optimal station-
ary policies in the negative case. Recall that d* is conserving if
d* = arg max (r e + Pav*} .

dED
Theorem 7.13. Suppose d' is conserving. Then the stationary policy d' is
optimal.
In Example 7.11, re(l) ..... 1 and oe(2)---=0 SO that v e is a solution of the

optimality equation while e is suboptimal. Thus, in contrast to the positive
case, unimprovable policies, that is policies 7r for which Tv "~= v '~, are not
necessarily optimal.
Optimal policies exist under considerably more general conditions in the
negative case than in the positive case. Strauch (1966) obtained the following
result when the set of actions is finite in each state. Note that it is a
considerably stronger result than in the positive case where a counterexample
was available for S countable.
Theorem 7.14. Suppose for each s C S, A s is finite. Then there exists" an optimal
stationary policy.
By Theorem 7.12, there exists a solution to the optimality equation, so as a

consequence of Theorem 7.14, any condition which ensures the existence of
392 M . L . Puterrnan
conserving decision rules is sufficient to guarantee the existence of a stationary

policy. Further results for this case are provided by Strauch (1966) and Schal
(1975).
7.2.3. Computation
Value iteration: As a consequence of Theorem 7.12, value iteration is con-
vergent provided v = 0. As in the positive case, the absence of bounds on
v " - v* makes solution by value iteration impractical. Van Dawen (1985)
shows that if S is finite and v* > - % the convergence rate is geometric.
Policy iteration: In the negative case the situation regarding policy iteration is
the reverse of the positive case. That is, the improvement step is valid while
the termination criterion is not. If d is the current stationary policy and d' is
chosen in the improvement step, then Vd'>~ V~ SO successive iterates are
monotone. But d' can satisfy the stopping criterion
r d, + Pd,V d' = max {r d + Pdv d' } (7.5)

tied
and not be optimal. This is because ('7.5) implies only that v d' is a solution of
the optimality equation but since the optimality equation does not have a
unique solution, it need not equal v*.
To illustrate this, suppose, in Example 7.11, that policy iteration begins with
stationary policy e. Then
--1 + ve(2) = max {v~(l), --1 + ve(2)} = Tv~(1)
and
re(2) = Tve(2),
so that (7.5) is satisfied. Thus, the algorithm will terminate in the improvement
step with the suboptimal policy e.
Linear programming: The optimal expected return and optimal policies cannot
be determined by a direct application of linear programming in the negative
case. The primal linear programming problem derived from Theorem 7.12 is
given by
Maximize ~, a i r ( j )
j~s
subject to
e(s)~r(s,a)+ ~, p(jts, a)v(j), a!A~,s2S,

jes
and
o(s)~<O, s ~ S .
As illustrated by Example 7.11, the region bounded by the above constraints

is not convex, so the above LP cannot be solved by the simplex algorithm.
Kallenberg (1983) provides an algorithm based on average reward methods
that is applicable.
8. Undiscounted Markov decision problems---ll
This section presents the theory of Markov decision problems with average
and sensitive optimality criteria. Whittle (1983, p. 118) summarizes the difficul-
ty in analyzing such problems as follows:
"The field of average cost optimization is a strange one. Counterexamples
exist to almost all the natural conjectures and yet these conjectures are the
basis of a proper intuition and are valid if reformulated right or if natural
conditions are imposed."
Because results for these criteria are highly dependent on the structure of the
Markov chains induced by stationary policies; the reader is referred to Chapter
2 on stochastic processes or to a basic text such as Kemeny and Snell (1960) for
basic definitions. Emphasis throughout Section 8 will be problems with finite
state spaces, but extensions to countable state problems will also be con.
sidered.
8.1. Markov chains, Markov reward processes and evaluation equations

Let {X,, t = 1, 2 , . . . } be a Markov chain with state space S and transition
probability matrix P with entries p ( j l s ) . Elements of the n-step transition
probability matrix pn are denoted by p " ( j ]s). The expression, chain structure
of a Markov chain, refers to the nature of its decomposition into classes. A
Markov chain is said to be unichain if it contains one recurrent class plus
additional transient states and irreducible if it consists of a single (necessarily
recurrent) class. Otherwise a Markov chain is said to be multichain.
8.1.I. Some matrix theory

Two matrices that are derived from P play a particularly important role in
MDP theory. Define the limiting matrix P* by
P*-- lim 1 ~ p~ 1 (8.1)

N--*~ ~ / t=l
where the limit is pointwise in S. When P is aperiodic,
P * = lira p N
The deviation matrix H e is defined by
H e = (1 .... ( P - - e * ) ) ~(1 ..... P * ) . (8.2)

394 M.L. Puterman
It is the Drazin generalized inverse of I - P (Lamond, 1986, Lamond and

Puterman, 1989) and plays an important role in the modern theory of Markov
chains. Related to the deviation matrix is the fundamental matrix Z e =
(I- ( P - p . ) ) - i (Kemeny and Snell, 1960).
The chain structure of the Markov chain determines the form of P* and H e.
If the chain is irreducible, P* has identical rows with strictly positive entries.
The rows are the stationary distribution of the Markov chain and each entry is
the reciprocal of the long run frequency of the system being in the correspond-
ing state. When the system is unichain with recurrent class R, for all s E S,
jER,
with p * ( j ) strictly positive. When X t is multichain, P* has strictly positive

entries corresponding to states within the same recurrent class, zero entries if
states are in different classes or are both transient and positive values for at
least one entry whenever the origin state is transient and the destination state is
recurrent.
8.1.2. The average reward, bias and Laurent series expansion

For a Markov reward process (MRP), i.e., a Markov chain together with a
(expected) reward function r defined on S, the average reward or gain is
g(s) = l i m -~ E~ ~, r(X,) (8.3)
where lim inf or lim sup replaces the limit in (8.3) when necessary (see Section
5). Evaluating the expectation in (8.3) and expressing the result in matrix
terms yields
g = lim 1 ~ pt_lr (8.4)

N--,~ N =P*r~
t=l
Combining (8.4) and the above results on the structure of P* yields the
following important result about the form of g.
Proposition 8.1. I f s and j are in the same recurrent class, g(s) = g ( j ) . Further,
if the chain is irreducible or unichain, g(s) is constant.
Consequently, in the irreducible and unichain cases, the average reward can
be expressed as gl where g is a scalar and 1 is a vector of ones. In the
multichain case, the gain is constant on each recurrent class. This distinction
has a major influence on the resulting theory.
In the discounted case, Proposition 6.1 shows that the expected total
discounted reward is the unique solution of the linear system v = r + h P v and
consequently v = ( I - h P ) - ' r . Analogous results are now developed for the
average reward case. They are based on the relationship between the dis-
counted and average reward of Blackwell (1962) and Veinott (1969).
For this analysis, it is convenient to parameterize the problem in terms of the
interest rate p instead of in terms of the discount rate A. They are related by
A = (1 + p ) - i or p = (1 ~)/~-1. The quantity 1 + p is the value of 1 unit one
-
period in the future. Transforming to this scale yields
v~ = ( I - AP)-Ir = (1 + p)((1 + p)I - e ) - l r

= (1 + p ) ( p I + (I - P))-lr. (8.5)
The quantity ( p I + (I - P ) ) - ' is called the resolvent of I - P and has a Laurent

series expansion (Hille and Phillips, 1957) around p = 0 as follows.
Theorem 8.2. Let v be the eigenvalue of P less than one with largest modulus. If
0 ~< p ~< 1 - I 1, then
( p I + I - p)-a = P-'P* + ~ (_p).H~+l (8.6)

n=0
and
v~ = (1 + p P-~Y-I + ~ (-P)'Yn (8.7)

n=0
where
~,rn+l
y_l=P*r and y,=rle r, n=0,1,....
The expansion in (8.7) is the Laurent series expansion of the expected

discounted reward and is obtained by multiplying both sides of (8.6) by r.
Derivations of (8.6) in the finite state case appear in Miller and Veinott (1969)
and Lamond and Puterman (1989) and for the countable state case in Hordijk
and Sladky (1977). Note that Y-1 = g.
The quantity Y0 = Hpr is important for the analysis of MDP's with average
reward criteria and will be denoted by h. It is called the bias or transient reward
and can be interpreted by expanding ( I - ( P - p , ) ) - i in a Neumann series to
obtain
he = Her = ( I - ( P - p,))-l(i_ P*)r
=Z Z (e - e*)r
r*=O n=O
= ~, pn(r - g) = E [r(X~) ---g o (8.8)

n=O ._
The representation in (8.8) enables h e to be interpreted as the expected total

reward for an MRP with reward r - g. If P is aperiodic, the distribution of X n
converges to a limiting distribution, so eventually r(Xn) and g(Xn) will differ
by very little. Thus h can be thought of as the expected total reward 'until
convergence' or the expected total reward during the 'transient' phase of the
chains evolution.
An alternative representation for h e due to Howard (1960) and discussed by
Denardo (1973) provides further insight into the interpretation of the bias.
Define v N as the total expected N period reward for an MRP with terminal
reward zero. That is,
N-I
ON = Z Pnr
n=0
Now using the next to last equality in (8.8) yields

N-1
he= Z P"r-Ng+ ~ (P~-P*)r
n -0 n=N
so that
DN = N g q- h p "Jr"O ( 1 ) , (8.9)
where o(1) is a vector with components which approach zero pointwise as

N---~ o0. Thus as N becomes large, for each s E S, VN(S) approaches a straight
line with slope g(s) and intercept he(s). When g is constant, VN(S) --VN(j)
approaches he(s ) - h p ( j ) , so that h e is the asymptotic relative difference of
starting the process in two states s and j. For this reason, h e is often referred to
as the relative value. Another immediate consequence of this representation is
that v N - VN_ 1 grows at rate g for N large. This justifies calling g the gain of the
process.
For optimization, policies will be initially compared on the basis of their
respective gain rates. If several policies have identical gain rates then (8.9)
implies that a decision maker would prefer the policy with the largest bias.
Consequently the bias could be used to break ties. Veinott (1974) provides
additional insight into the interpretation of the bias as well as the higher order
terms in the Laurent series expansion of v,.
In the context of average reward Markov decision processes it will suffice to
use Blackwell's (1962) truncated Laurent expansion of v, which is a con-
sequence of (8.7).
Corollary 8.3.
v, = (1 .... a ) - l g + h + / ( a ) (8.10)
where f(A) is a vector which converges to zero componentwise as A~'I.

The following corollary is useful for extending existence results from the
discounted case to the average reward case (Derman, 1970, pp. 25-28) and
establishing structural results for average reward problems from those for
discounted problems. It follows immediately by multiplying both sides of (8.10)
by 1 - A and passing to the limit.
Corollary 8.4.
g=fimll(1-X)v,. (8.11)
8.1.3. The evaluation equations

Representation (8.7) can be used to establish the average reward and
sensitive policy evaluation equations. Substitution of (8.7) into the discounted
reward evaluation equation expressed as
r + [ ( P - I ) - pI]{(1 + p)-lva} = 0
and equating terms with like powers of p yields the following result (Veinott,
1969). The converse follows by multiplying the (n + 1)th equation by P* and
adding it to the nth equation.
Theorem 8.5. The coefficients of" the Laurent series expansion o f v a satisfy the
system o f equations
(P - I ) y _ 1 = 0, (8.12)
r - Y-1 + (P - l)Yo = 0, (8.13)
--yn+(P--I)yn+l=O, n>-O. (8.14)
Conversely, if w l, Wo. . . . , Wm, Win+1 satisfy equation (8.12), (8.13) and

(8.14) f o r n = O , 1 , . . . , m , m > 1 O t h e n w _ l = y 1, W o = Y o , ' ' ' , W m - I = Y m 1,
Wm = Ym and Wm~ 1 is unique up to a vector in the null space o f I - P.
The following two corollaries give the reduced form of Theorem 8.5 that wilt
be directly applicable to average reward problems. The first gives the equations
to be solved to determine the average reward for transition probability
matrices with general chain structure.
Corollary 8.6. The vectors g and h satisfy
(P--I)g=O (8.15)
and
r - g + (P - Oh = 0. (8.16)
398 M.L. Puterman
Conversely if g' and h' satisfy (8.15) and (8.16) then g ' = g and h ' = h + w
where ( P - 1 ) w = O .
In the case that P is unichain, any solution of (8.15) is a constant vector so

that the evaluation equation simplifies further. In this case g will be written as
gl where g is a scalar and 1 is a vector of l's of appropriate dimension.
Corollary 8.7 is the basis for policy evaluation in the unichain and recurrent
cases and serves as the basis for the optimality equation in these settings.
Corollary 8.7. Suppose P ~ unichain. Then the average reward and the bias
satisfy
r - gl + ( P - I)h = 0 . (8.17)
Conversely if the scalar g' and the vector h satisfy (8.17), then g' = g and
h' = h + w where ( P - I)w=O.
To implement and analyze policy improvement algorithms in the average

reward case requires a unique specification of h. Conditions which ensure this
are discussed in Section 8.3 and 8.8.
8.2. The average reward optimality equation--The unichain case

The section presents the optimality equation for Markov decision problems
with average reward criteria under the assumption that transition matrices
corresponding to all decision rules are unichain. This will be referred to as the
unichain case. A single optimality equation is sufficient to characterize optimal
policies and their average rewards making results simpler than in the multi-
chain case where a pair of nested functional equations is necessary.
The development herein parallels that in Section 6.1 where possible. It is
assumed throughout this section that the rewards and transition probabilities
are stationary, that the action space, rewards and transition probabilities are
such that appropriate maxima are attained and the state space is either finite or
countable. Results which hold under weaker conditions than unichainicity will
be noted.
Recall that a policy ~r* is average or gain optimal if
g~" = lirn -1v n~, > l i m - -1v n = g ~ ' for a l l ~ r E I l .
When the limits above do not exist, either weak or strong average optimal
policies (Section 5) are sought.
The optimality equation in a unichain average reward MDP is given by
O=maxtr(s,a)-g+ ~_~ p ( j I s , a ) h ( j ) - h ( s ) } , sES. (8.18)

a@A s I. jES
Ch. 8. M a r k o v D e c i s i o n Processes 399
It can be expressed in matrix-vector and operator notation as
0 = max {r d - gl + ( P a - I)h}-~ B ( g , h ) . (8.19)

dED
As in previous sections, the maximum in (8.19) is componentwise. When D

consists of a single decision rule, this equation reduces to (8.17). Recall that V
denotes the set of bounded real valued functions on S.
The following theorem is the average reward counterpart of Theorem 6.2 (cf.
Hordijk, 1974).
Theorem 8.8. Suppose there exist a scalar g and an h E V that satisfy
O~
> m a x {r a - gl + (Pd - I)h} . (8.20)
dED
Then
g/>sup
[ l i m s u p -1v :] .
wE/1 t_ n---~ n
If instead there exist a scalar g and an h E V that satisfy
0 ~ max {r e - gl + (Pa - I)h} (8.21)

dEl)
then
rElI n
Combining (8.20) and (8.21) yields:
Theorem 8.9. If equation (8.19) has a solution (g, h) E R a V, then

(a) there exists a scalar g* satisfying
g*=sup
rEH
rl i m -1:]
Ln--*
v
n
and
(b) g is unique and equals g*.
The following result is that the optimality equation determines average

optimal stationary policies. A decision rule d h is said to be h-improving if
rah- gl + (Pah - l ) h = max {r e - gl + (ed -- I ) h } (8.22)

dED
or equivalently
rah + Pahh = max {r a + Pah} .

dED
Theorem 8.10. Suppose there exist a scalar g* and an h * ~ V which satisfy

(8.19) and d* is h*-improving, then
e* ,~ = g .
g = max g
"rr@n
This result can be restated as follows. If the optimality equation possesses a

solution (g*, h*), then g* is unique and any stationary policy which uses an
h*-improving decision rule every period is strongly average optimal. It remains
to show that a solution to the optimality equation exists.
Three approaches have been used to establish existence of solutions to the
average reward optimality equation in the unichain case. They are:
(1) Policy iteration (Howard, 1960).
(2) Extensions of results for the discounted case (Theorem 6.6) obtained by
letting I increase to 1 and using representations for the discounted reward such
as (8.10) and (8.11) (Taylor, 1965).
(3) Fixed point theorems (Federgruen and Schweitzer, 1984b).
A sufficient condition for the existence of solutions to (8.19) is Ross's
(1968a) following generalization of Taylor (1965).
Theorem 8.11. Suppose there exist a finite N and an s o E S such that
Io (s) - v (s0)l < N (8.23)
for all s ~ S and 0 < A < 1. Then there exist an h* E V and a scalar g* which
satisfy (8.19) and
g* = lim
x]'l
(1 - A)v~(s0). (8.24)
Sufficient conditions for (8.23) to hold include:

(a) S is finite, A, is finite for each s ~_ S and every stationary policy is
unichain;
(b) S is finite, A s is compact for each s E S , r(s,a) and P ( J l s , a) are
continuous in a and every stationary policy is unichain;
(c) S is finite and the set of transition matrices corresponding to stationary
policies is a communicating system (Bather, 1973), that is, for every pair of
states s, j there exists a decision rule ~ and an integer n~>l such that
e (jls)>O;
(d) S is countable, rewards are uniformly bounded and the expected number
of transitions to reach state s o from s is uniformly bounded for all policies and
states s.
The results of this section can be summarized in the following theorem.
Theorem 8.12. Suppose (8.23) holds'. Then:

(a) there exists a unique g* and h * E V which satisfy the optimality equation
(8.19),
(b) there exists' h*-improving decision rules', and
(c) a stationary policy which uses an h*-improving decision rule is strongly

average optimal.
Theorem 8.12 is valid in the countable state case under weaker conditions
than (8.23) (Federgruen, Hordijk and Tijms, 1978, 1979). Related work
includes Hordijk (1974), Wijngaard (1977), Federgruen, Schweitzer and Tijms
(1983) and Deppe (1984).
The assumption that solutions to the optimality equation are bounded is
crucial for the existence of average optimal stationary policies. Counterexam-
pies have been provided by Fisher and Ross (1968) and Ross (1983), Bather
(1973), Sheu and Farn (1980) and Schweitzer (1985).
8.3. Policy iteration in the unichain case

Policy iteration is an efficient procedure for solving the optimality equation
and finding optimal policies in MDP's with average reward criterion. It
generates a sequence of stationary policies with monotonically non-decreasing
gains. It also is an important theoretical device for establishing existence of
solutions to the average reward optimality equations.
8.3.1. The algorithm

This algorithm was developed by Howard (1960) for finite state and action
MDP's. He demonstrated finite convergence under the assumption that all
policies are recurrent. In the countable state case, Derman (1966) used policy
iteration to constructively show the existence of a solution to the optimality
equation (8.19) under the assumptions that all states are recurrent under each
stationary policy and that the reward, gain and bias are uniformly bounded on
the set of stationary policies. Federgruen and Tijms (1978), Hordijk and
Puterman (1987) and Dekker (1985) demonstrated the convergence of policy
iteration for problems with compact action spaces.
The Policy Iteration Algorithm.

1. Set n = 0 and select an arbitrary decision rule d n E D.
2. (Policy evaluation) Obtain gd,, and an ha. by solving
0 = rd. -- gl + (Pal. -- 1 ) h . (8.25)

3. (Policy improvement) Choose d.+ 1 to satisfy
rd.+l + Pd.+lhd, = max {r d + Pdhd,} (8.26)
setting d,+l = d, if possible.

4. If d,+ 1 = d,, stop and set d* = d,. Otherwise increment n by 1 and return
to 2.
The above algorithm yields a sequence of decision rules {d,} and corre-
sponding gains { ga.}. The relative value functions {ha. } determined by solving
(8.25) are unique up to an additive constant. The choice of the additive
constant has no effect on the maximizing decision rule in (8.26) since for any h
satisfying (8.25) and any constant c,
r d + Pd(h + cl) = rd + Pdh + cl
for all d @ D. Computationally, it is convenient to set
ha,(So) = 0 (8.27)
for an arbitrarily selected s o. Solutions determined by different choice of s o

RV
differ by a constant. When ha, is determined by (8.27) it is denoted by hdn , the
superscript RV denoting relative value.
To implement the evaluation step under condition (8.27), solve the linear
system
r = (Q~o)w (8.28)
where Q~0 is the matrix I - P with the column corresponding to state s o

replaced by a column of l's. The solution of (8.28) is unique, satisfies (8.27)
and has ga, as its s0th component. It can be obtained by Gaussian elimination
or any appropriate iterative method. From a theoretical perspective, the
Blackwell (1962) specification that
Pdhd, = 0 (8.29)
is more convenient since it ensures that
hdn = H p a rdn
as in Section 8.1.2. When h is obtained from (8.25) and (8.29), it will be

denoted by hda, the superscript B denoting bias. It is easy to see that
h rtv = h B
dn without computing P*dn

Veinott (1969) provides a method for finding h B
8.3.2. Convergence of policy iteration

Convergence in the finite state and action case is a consequence of the
lexicographic monotonicity of the iterates of the above algorithm and the
finiteness of the set of stationary policies. If improvement occurs in a recurrent
s t a t e u n d e r stationary policy d,,+l, the gain for the improved policy is greater
than the gain for the previous policy (Howard, 1960). This is formally stated
as;
P r o p o s i t i o n 8.13. Suppose d,,+l is determined in step 3 of the policy iteration

algorithm. Then the following hold:
(a) ga.+l = g a l + P S . + B ( g a , ha, ) . (8.30)

(b) If B ( g a , ha.)(s)>O for a state s which is recurrent under d.+~, then
ga.+ l > g~."
(c) If B(ga , ha.)(s ) = 0 for all states s which are recurrent under d.+l, then
ga, +~= ga..
Representation (8.30) (Hordijk and Puterman, 1987) can be thought of as a

'Newton method' representation for the gains at sucessive steps of the policy
iteration. An immediate consequence of parts (b) and (c) is the following
convergence result.
Theorem 8.14. If all states are recurrent under every stationary policy and the
sets of states and actions are finite, then policy iteration converges in a finite
number of iterations.
When there are transient states associated with some (or all) stationary
policies, additional analysis is based on:
P r o p o s i t i o n 8.15. Suppose d n is determined in the improvement step of the

policy iteration algorithm and ha, is any solution of (8.25). Then
(a) h d.+
B 1= hB , B
d. -- P d . + l h a . + Hd.+lB(ga ., ha.) (8.31)
(b) If B ( g a , h a . ) ( s ) = 0 for all s that are recurrent under dn.l and

B ( g a , ha.)(So) > 0 for some s o which is transient under d.+ I, then
hB >h~.
dn+ 1
(c) If B(gan, han)(s)=O for all s that are recurrent under dn+ l and
B(ga , ha,)(s ) = 0 for all s which are transient under d,+l, then
hB = hB
dn+l dn
The result in Proposition 8.15(b) means that if there is no improvement in

states which are recurrent under the new policy and an improvement in a state
which is transient under the new policy, then the bias of the new policy will be
strictly greater than that of the previous policy. Surprisingly, this result does
not imply that the relative values are monotone increasing. Thus at sucessive
iterates, the algorithm produces a stationary policy with a larger gain and if this
is not possible then a policy with a larger bias. If neither of these alternatives
are possible the algorithm terminates.
404 M.L. Puterman
Theorem 8.16. Suppose all stationary policies are unichain and the sets o f states
and actions are finite. Then policy iteration converges in a finite n u m b e r o f
iterations.
The above results provide additional insight to the behavior of the iterates of
policy iteration. If s is recurrent and j is transient under action a, P ( ] [s, a) = 0.
Consequently, once the optimality equation is satisfied on all states that are
recurrent under a decision rule 6, which attains the maximum in the improve-
ment step, there will be no future changes in the gain. Consequently any
stationary policy which agrees with P~ on its recurrent states is average
optimal. Since, in subsequent iterations the bias is increased in transient states
of the maximizing policy until no further improvement is possible, one might
suspect that policy iteration terminates with a policy that is bias-optimal, that is
it has the largest bias among all policies with the same gain as 6. This
supposition is false (Example 2 in Denardo 1973 in which two policies have the
same gain but different recurrent classes). What is attained in this case is the
policy with the largest bias among policies which have the same recurrent class
as 6. To find bias optimal policies requires solution of an additional optimality
equation (Section 8.9).
8.4. Value iteration

This section provides sufficient conditions for convergence of value iteration
in the finite state case. The development is in the spirit of the survey articles by
Federgruen and Schweitzer (1978) and Schweitzer and Federgruen (1980).
Books by van der Wal (1981), Whittle (1983) and Bertsekas (1987) provide
valuable insight. Other important contributions include White (1963), Brown
(1965), Lanery (1967) and Bather (1973).
8.4.1. Convergence
Value iteration in the undiscounted case is based on the operator
T v = m a x [r a + P a v ] (8.32)
dED
on V, the space of bounded real valued functions on S. This operator is not a

contraction and consequently the iterative scheme v n+~ = To n is not necessarily
convergent. The quantity v n is the maximum expected total n-period return
when the terminal reward is v . For a fixed decision rule, v n can be expressed
as
n
O = Z eTr. +
m=0
= Z (e7 - + ne, r. + e, o (8.33)

m=O
When Pd is aperiodic, lim,~= P~ = P~ so it follows from (8.8) that for n

sufficiently large,
v" = H a r d + n g d + P ] v (8.34)
Consequently one might conjecture that
L = lim {v n -
n--~ oo
ng*} (8.35)
always exists. The following simple example shows this conjecture is false.
Example 8.17. Let S = {1, 2} and suppose there is a single decision rule d, with
rd = and Pa = [011
1 0 "
Then clearly gd = 0 and if
v =
0[a] b '
then
n even,
[Iab
On .= p n v O = b] '
]
|, n odd.
a J
Thus unless a = b = 0, lim,_,~ {o" - rig*} does not exist, but for a n y choice of
a and b, both lim,__,= v z" and lim,__,= v z"+l exist.
In this example, states 1 and 2 are both recurrent but each is periodic with
period 2. This suggests that periodicity causes problems for the convergence of
value iteration in average reward problems.
When the limit in (8.35) exists, value iteration can be used to solve the MDP
because:
(a) for N sufficiently large,
N N 1 g, o
v -v ~-L+Ng*-(L+(N--1)g*)= ,
(b) for n sufficiently large,
U N -Ng *~'~" h*
where h* is a solution of the optimality equation (8.19);

(c) for N sufficiently large, stationary policies which are vU-improving ale
optimal; and
(d) upper and lower bounds for the optimal gain are available.
When S is finite, the following result summarizes conditions when the limit in
(8.35) exists.
406 M.L. Puterman
Theorem 8.18. Let S be finite and let v n+l = Tv n. Then the limit in (8.35) exists
for any v E V if any o f the following conditions hold:
(a) For all s ~ S, P( j I s, a) > 0 for all a E A s and j ~ S (Bellman, 1957).
(b) There exists a state s o and an integer u >>-1 such that
(PdlPd2 "' Pdu)(S, SO) ~ ~ > 0 (8.36)
for any decision rules d 1, d 2 . . . . , d r and all s E S (White, 1963).

(c) For any decision rule d, Pd is aperiodic (Schweitzer, 1965).
(d) Every stationary optimal policy is unichained and at least one o f them is
aperiodic (Denardo, 1973, Schweitzer and Federgruen, 1977).
Condition (a) is the least general; it implies (b)-(d). Condition (b) is

stronger than condition (c); conditions (c) and (d) are distinct. Condition (a)
ensures that the underlying Markov chain is irreducible and aperiodic under
every policy. Condition (b) means that there exists a state s o which can be
reached in u transitions with positive probability from each starting state for
any policy. Condition (d) allows non-optimal policies to be periodic and have
arbitrary structure but places restriction on the class of optimal policies while
condition (c) requires that all stationary policies be aperiodic but possibly
multichained.
The transition matrix in Example 8.17 violates each of the conditions in
Theorem 8.18. However, if Pd is replaced by
,,:['o
then condition (c) is satisfied and v n is (trivially) convergent for any v . In
practice, conditions (a) and (c) are easiest to check.
Schweitzer and Federgruen (1977) provide necessary and sufficient condi-
tions for the limit (8.35) to exist for all v , generalizing Brown (1965) and
Lanery (1967).
8.4.2. Determination o f optimal policies

An important practical issue in finite state problems is under what conditions
value iteration can be used to determine an average optimal stationary policy.
The following result provides a theoretical solution to this problem.
Theorem 8.19. Suppose the limit in (8.35) exists. Then:

(a) there exists an integer n o such that for all n >~no~ arty vn-improving
decision rule is average optimal (Odoni, 1969; van der Wal, 1981), and
(b) if d is vn-improving for infinitely many n, then d is average optimal
(Brown, 1965).
This result can be restated as follows. If
r d + Pdv" = Tv"
for n I> n o or for infinitely many n, then gd = g*"

The results in this theorem cannot be used in computation because no
bounds are available for no; i.e., the second condition is not verifiable. Even
when the limit in (8.35) exists, asymptotic behavior of the sets of v"-improving
decision rules can be eratic; they can be strict subsets of the set of maximal
gain rules for every n and they can oscillate periodically (Brown, 1965) or even
aperiodically (Bather, 1973) within the set of maximal gain decision rules.
Thus, convergence of the set of v"-improving decision rules cannot be used as a
termination condition for value iteration. When the limit in (8.35) fails to exist,
Lanery (1967) provides an example where non-maximal gain decision rules
appear infinitely often in the sequence of vn-improving decision rules.
Result (a) in Theorem 8.19 is a turnpike theorem for the expected total
reward criterion (Section 6.2.4). Whenever the horizon is known to exceed n o
periods, then any vn-improving decision rule (n i> no) is optimal in the first
period for a finite horizon problem under expected total reward criterion.
8.4.3. B o u n d s on the gain

In the unichain case, bounds on the optimal gain rate were given by Odoni
(1969) and Hastings (1976). They are valid in general and are given in the
following proposition.
Proposition 8.20. Suppose h is hounded. Then
L ( h ) =--rain ( Th(s) ~ h(s)) <~gd <<"g* ~ m a x ( Th(s) - h(s)) ~ U(h )

sES
(8.37)
where gd is any h-improving decision rule.
These bounds have been applied to value iteration as follows (Platzman,

1977).
Theorem 8.21. Suppose g* is constant and v n = Tv"-1. Then for all n,
min {v"(s) - vn-~(s)} ~<gd. <~g* <" max (v"(s) - v " - l ( s ) )

sES s~S
(8.38)
n-1
where dn is any v -tmproving decision rule. Further, if the ~mit (8.35) e x i s t ,
then
L ( v ~ --1) <<L ( v " ) < gal. ~ g* << U(v~) <~ U( v -1) (8.39)
and lira,_.= {U(v") - L(v~)} = O.

That g* is constant is difficult to check a priori. Conditions which imply it

include:
(a) all stationary policies are unichain,
(b) the set of policies is a communicating class, or
(c) S is simply connected, that is it can be partitioned into a communicating
class and a set of states which are transient under all policies (Platzman, 1977).
Thus when all stationary policies are aperiodic and g* is constant, the upper
and lower bounds converge to zero and are monotonic. This means that an
e-optimal policy can be found by stopping value iteration when U(v ~) -
L(v n) < e. At termination, ga, is within e of g* and the statinary policy d n is
e-optimal. When the limit (8.35) does not exist or the problem is multichain,
then limn__,= { U ( v " ) - L(v")) will not equal zero and the bounds cannot be
used as a stopping criterion.
In contrast to the discounted case, these bounds cannot be used for
elimination of suboptimal actions. Instead bounds on h* are required. Hastings
(1976) provided a procedure for temporary action elimination in average
reward problems.
8.4.4. Variants o f value iteration

When applying value iteration in the average reward case, there are two
potential problems.
(1) The divergence of v n might lead to numerical instability.
(2) The limit in (8.35) might not exist because of the periodicity of
transition matrices corresponding to some policies.
White (1963) provides a convergent value iteration scheme. Instead of using
v "+1= Tv n, his procedure renormalizes v" after each iteration so that one
component is zero. That is, a state s o is selected and successive iterates are
defined for all s E S by
w(s) = v ( s ) - v(So) and wn+l(s) = Tw"(s) - Tw"(So) . (8.40)
This procedure is called relative value iteration. In investigating its conver-

gence, it is convenient to use the span of a vector v E V as defined in Section
6.7. The following combines results of White (1963) and Platzman (1977).
Proposition 8.22. let w n be generated by (8.40). Then if g* is constant:
(a) w = l i m w"(s) , s~ S ,
exists;
(b) g = W(So) and h ( s ) = w(s) satisfy B ( g , h ) = 0; and
(c) if (8.36) holds,
s p ( w ~+1 - w ) < (1 - ~ ) 1 ' ~ sp(w ~ - w"-~), (8.41)
where 6 is defined in (8.36).

This result states that w n converges to a solution of the average reward

optimality equation and consequently finds the optimal gain g and a relative
value vector h. Part (c) implies that when (8.36) holds, the convergence rate is
linear with constant at least equal to ( 1 - 3) 1/v. Error bounds can be derived
from (8.41) and appear in Bertsekas (1987, p. 321) which also includes a proof
of this proposition under (8.36). Bounds like those in Theorem 8.22 are
available for relative value iteration.
When some policies, especially those that are average optimal have periodic
transition matrices, the limit in (8.35) need not exist for every starting value.
Schweitzer (1971) proposed transforming the problem so that all decision rules
have aperiodic transition matrices and the transformed problem has the same
relative costs and optimal policies as the original problem. The importance of
this transformation is that it ensures the limit (8.36) exists and value iteration is
globally convergent. The data transformation can be combined with relative
value iteration to obtain a scheme which is convergent for any unichain MDP
(Bertsekas, 1987, pp. 323-324).
8.5. Linear programming--Unichain case

The linear programming approach for the average reward MDP's is due to
de Ghellinck (1960) and Manne (1960) for problems in which the transition
probability matrix for each stationary policy is ergodic. Other work in this area
is by Denardo and Fox (1968), Denardo (1970), Derman (1970), Hordijk and
Kallenberg (1979, 1980) and Kallenberg (1983). The last reference is particu-~
larly noteworthy.
The primal problem is based on the result in Theorem 8.8 that g~> g*
whenever there exist (g, h) E R 1 x V satisfying
g >>-ra + (Pd - I ) h (8.42)
for all d E Do As a consequence of Theorem 8.9, g* is the smallest g for which

there exists an h E V satisfying (8.42). It follows that g* is an optimal solution
of the following primal linear programming problem.
Primal Linear Program.
Minimize g
subject to
g~ r(s, a) + ~ p(jls, a)h(j)- h(s),

jEs
and g and h unconstrained.
Its dual problem is as follows.

410 M.L. Puterman
Dual Linear Program.
Maximize ~ ~ r(s, a)x(s, a)

s~S a~A s
subject to
2 x(j,a)-2 2
a~A s sES aEA s
p(jls, a)x(s,a)=O, jCS,
E E x(s, .)= 1, (8.43)

s~S a~:A s
and x(s, a) >i O, a C A , , s C S.
By Theorem 8.11, there exists an optimal solution to the primal problem so

that by the duality theorem of linear programming, the dual problem processes
an optimal solution. A stationary average optimal policy for the MDP can be
found by using the simplex algorithm to obtain an extreme point solution x* of
the dual problem and then choosing a decision rule d* satisfying
if x(s, a ) > O , s ~ S * ,
d*(s) = {aarbitrary ifs~S- S* ,
where
S* = {s ~ S: x*(s, a) > 0 for some a G A,} . (8.44)
For any decision rule d* obtained in the above manner, x*(s, d*) is an
optimal solution to the dual problem and satisfies the equalities
x*(s, d*(s)) - ~ p(jls, d*(s))x*(s, d*(s)) = O, j E S,

sES
and
x*(s, d*(s))= 1. (8.45)

sES
These imply that x* is the stationary probability distribution corresponding to

the stationary policy d*. Since x* is zero on S - S*, such states are transient
under d*; S* is a recurrent class.
In the unichain case, there is not a one-tooone relationship between feasible
solutions to the dual problem and stationary policies. Under the assumption
that all stationary policies are recurrent, then the following result can be
obtained.
Ch. 8. M a r k o v D e c i s i o n Processes 411
Theorem 8.23. Suppose that the transition probability matrix of every stationary
policy is irreducible. Let x be any feasible solution to the dual problem, and
define the randomized stationary policy d by
P{d(s) = a} - x(s, a) , aEA,,sES. (8.46.)

2 x(s,a')
a'EA s
Then
x(s,a)= P{d(s)=a}~rd(s), aEAs, s@S , (8.47)
is a feasible solution to the dual problem where 7ra is the solution of
~rPd = ~r ai~d ~ ~ ( s ) = 1. (8.48)

sES
Further, given any randomized stationary policy d, x(s, a) defined by (8.47)/s a

feasible solution to the dual problem.
This theorem states that the mapping (8.47) is a one-to-one mapping

between stationary policies and solutions to the dual problem in the recurrent
case. Note that x(s, a) is the stationary probability that the system is in state s
and action a is chosen when using either a prespecified stationary policy in
which case 7r is computed by (8.48) or using the stationary policy with decision
rule defined by (8.46). For irreducible problems, S * = S, so in this case the
arbitrariness of the policy selected by the algorithm is removed. Consequently
policy iteration corresponds to the simplex method with block pivotting and the
simplex algorithm corresponds to a version of policy iteration in which only the
maximum improvement is selected.
For unichain problems, this relationship is not valid because the policy
iteration algorithm terminates with policies with special structure (Section
8.3.2) while the policy selected by the LP method is arbitrary on a subset of the
state space.
8.6. Modified policy iteration--Unichain case

For the average reward criterion, modified policy iteration is best thought of
as a variant of value iteration. It was proposed by Morton (1971) and
subsequently analyzed by van der Wal (1981) and Ohno (1985).
The algorithm to obtain an e-optimal policy is as follows:
Modified Policy Iteration Procedure of order m.

1. Select v E V, set n = 0 and specify e > 0.
2. Determine a vn-improving decision rule d,,+~.
3. Compute L(v") and U(v") using (8.37). If U(v") = L(o ~) < e, stop and
set d " = dn+l, otherwise continue.
4. Obtain 0 n+l by
n+l z~ xm+l n
O = ~ldn+l ) O (8.49)
where
Tar =- r a + Pat) . (8.50)
5. Increment n by 1 and return to 2.
Modified policy iteration of order 0 is value iteration; in this case step 4 is

superfluous. When m is large the algorithm is similar, although not identical to
policy iteration.
The following result (van der Wal, 1981) is proved by showing that the
sequence of upper and lower bounds derived from (8.37) converge to g*.
Theorem 8.24. Suppose that S and A s for each s E S are finite, that for some
a > O, Pd >~t~lfor all d E D and that all stationary policies are unichain. Then if
{o n} is generated by modified policy iteration:
(a) L(v n) converges monotonically and exponentially fast to g* and
(b) U(o n) converges exponentially fast to g*.
The condition on Pu is referred to as strong aperiodicity; without it the

algorithm can cycle. The lower bound converges monotonically to g* but the
convergence Of the upper bound is not necessarily monotone (cf. Example 9.2
in van der Wal, 1981). As a consequence of Theorem 8.24, the stopping
criterion in step 3 is satisfied in a finite number of iterations. When implement
ing the algorithm, the bounds in 3 are evaluated at the same time as a
v"-improving decision rule is determined in step 2.
The asymptotic properties of the values generated by modified policy
iteration have not been investigated and the convergence of the algorithm for
non-finite problems or those with more general chain structures has not been
demonstrated.
8. Z Average reward--Multichain case

When there are multiple chains, that is, the Markov chains corresponding to
stationary policies have more than one recurrent class, two optimality equa-
tions are required to characterize optimal policies. Consequently, theory and
algorithms are more complex than in the unichain case.
Howard (1960) provided a policy iteration algorithm for solving this system
of equations. Other important references in this case include Blackwell (1962),
Denardo and Fox (1968), Veinott (1969), Denardo (1970), Schweitzer and
Federgruen (1978), Hordijk and Kallenberg (1979), Kallenberg (1983) and
Federgruen and Schweitzer (1984a,b).
Ch. 8. Markov Dec&ion Processes 413
Results are not as complete as in the unichain case and few are available
when the set of stationary policies is infinite. The assumption that the sets of
states and actions are finite is required for results in this section. Discussion will
focus on the optimality equation and policy iteration. The reader is referred to
Denardo and Fox (1968), Denardo (1970), Derman (1970), Dirickx and Rao
(1979) and Katlenberg (1983) for linear programming for multichain MDP's.
Little is known about value iteration and modified policy iteration for multi-
chain problems.
8. 7.1. Multichain optimality equations

For multichain average reward MDP's, a pair of nested optimality equations
is required to compute the optimal gain and determine optimal policies. They
are given by
max {(ed - l)g}=0 (8.51)

dED
and
max {r d - g + (Pal-- I)h) = 0

dEE
(8.5z)
where E = {d C D: P e g = g}.
A solution to this pair of nested functional equations is a pair ( g, h) ~ V x V
such that P e g <~ g for all d E D with equality holding for at least one d E D and
r e - g + Pah <~ h (8.53)
for all d ~_ D for which P d g = g with equality holding in (8.53) for at least one
such d.
In the unichain case the first optimality equation above is redundant and
E = D. This is because when all decision rules are unichain, P u g = g implies
that g is a constant so that equation (8.51) is satisfied for all d E D. The above
reduces to (8.18) in the unichain case. If D replaces E in (8.52) then possibly a
different decision rule attains the maximum in each equation.
Establishing that solutions to this pair of equations characterize average
optimal policies is not as straightforward as in the unichain case. A proof is
based on the existence of a Blackwell optimal stationary policy as defined in
(5.11). When S is finite, a policy It* is Blackwell optimal if there exists a
A * , 0 < ~ A * < I , such that or* is discount optimal for all A E [ A * , I ) . The
following important theorem is due to Blackwell (1962). An elegant non~
constructive proof using function theory was provided by Blackwell, a c o n
structive proof is based on the policy iteration algorithm in Section 8.9.
Theorem 8.25. There ex&ts a Blackwell optimal policy which is stationary
A closely related result is the following~

Theorem 8.26. If d* is Blackwell optimat men

(a) d* is average optimal in the class of all policies (Derman, 1970), and
(b) (g~, h~) satisfy the average reward optimality equations (8.51) and
(8.52).
The converse of this, that an average optimal policy is Blackwell optimal and
a policy determined by a solution to the optimality equation is Blackwell
optimal, holds when there exists a unique average optimal policy. Blackwell
(1962) provides a simple example in which an average optimal policy is not
Blackwell optimal. The above results establish the existence of a solution to the
optimality equations. In the next subsection, policy iteration will be used to
construct such a solution. The following theorem gives the optimality prop-
erties of solutions to the optimality equations.
Theorem 8.27. Suppose (g*, h*) satisfies (8.51) and (8.52) and d* attains the
maximum in (8.52) at (g*, h*). Then for all ~ E II,
lira inf (1 - A)(v~* - v~) >t 0

A'D
(8.54)
and
ga. ) g~.
8.8. Multichain policy iteration

Policy iteration in the multichain case consists of an improvement and an
evaluation step. In the improvement step, a decision rule is sought which
provides a strict improvement in (8.51) and if none is available, in (8.52).
When no improvement is possible in either equation the algorithm is termi-
nated. Proof of convergence is based on the partial Laurent expansion (8.10)
and was provided by Blackwell (1962). The algorithm is due to Howard
(1960). It is implemented componentwise.
The Policy Iteration Algorithm--Multichain case.

1. Set n = 0 and select an arbitrary decision rule d n ~ D.
2. (Policy evaluation) Obtain gal. and an he. by solving
O=(Pd,,-I)g and O=re,--g+(Pd,-I)h. (8.55)

3. (Policy improvement)
(a) Choose dn+ 1 C D to satisfy
Pa,,+, = max {Paga,}

dGD
(8.56)
and set dn+ 1 = d, if possible. If dn+ ~ = de,, let E n = {d E D: Page, = Pd,gd,}
and go to (b), otherwise increment n by 1 and return to 2.
(b) Choose d,+l E E~ to satisfy
re.+, + Pdn + l ha n = max

dEEn
{r a + Pdha.} (8.57)
setting dn+l = d n if possible.

4. If dn+ 1 = dn, stop and set d * = dn. Otherwise, increment n by 1 and
return to 2.
The above algorithm yields a sequence of decision rules {dn) and corre-
sponding gains { gd,}" The pair of matrix equations (8.55) uniquely determines
the gain, but the relative v a l u e s {hdn } are only unique up to a u satisfying
(Pa, - I)u = 0. If Pd, has k recurrent classes, then ha, will be unique up to k
arbitrary constants which can be determined by setting ha,(S ) = 0 for an
arbitrary s in each recurrent class of Pdn (Howard 1960). Blackwell's specifica-
tion (8.29) can also be used, but is computationally prohibitive. Veinott (1969)
provides a method for finding an h which satisfies (8.58). In practice, any h will
do.
The improvement step of the algorithm consists of two phases. First im-
provement is attempted through the first optimality equation (8.56), that is a
ga,-improving decision rule is sought. (Call a decision rule d' g-improving if
d' = arg maxd~ D {Pug}.) If no strict improvement is possible, an ha-improving
decision rule is found among all gan-improving rules and if no strict improve-
ment is possible, the iterations are stopped. Otherwise, the improved policy is
evaluated at the subsequent iteration.
In the unichain case, the first equation in (8.55) and part (a) of the
improvement step are redundant so that the algorithm reduces to the unichain
policy iteration algorithm.
Proofs that policy iteration is convergent in the multichain case for finite
state and action problems have been provided by Blackwetl (1962) using the
partial Laurent expansion and Denardo and Fox (1968) using detailed analysis
of the chain structure. The next result shows the monotone nature of the
multichain policy iteration algorithm and is the basis for a proof of con-
vergence.
Proposition 8.28. Let d E D and suppose either

(a) d' @ D is strictly gd-improving, or
(b) d' G D is gd-improving, Pd'gd = Paga and d' is' slrictly ha-improving,
Then
lim inf (1 - A)(o~' -~ o~) > O. (8.58)
As a consequence of this result, sucessive iterates of policy iteration satisfy
lim inf (1 - A)(v d'+~ - v ~dn) > 0 (8.59)

416 M.L. Puterman
until the algorithm terminates with d,+ 1 = d n at which point the optimality
equations are satisfied. Finite convergence is ensured since there are only
finitely many policies and the algorithm is monotone in the sense of (8.59).
This yields the following result.
Theorem 8.29. Suppose A s for each s E S and S are finite. Then the policy
iteration algorithm terminates in a finite number of iterations with a gain optimal
stationary policy and a pair (g*, h*) which satisfy the optimality equations
(8.51) and (8.52).
Based on the above arguments, one might speculate that this implies that the
gains of successive policies are strictly increasing. This is not the case because
(8.59) does not exclude the possibility that the gains of two successive policies
are identical and improvement occurs in the bias term.
Howard (1960, p. 69-74) and Denardo and Fox (1968, p. 477-479) have
analyzed the behavior of the iterative process in detail and have shown that:
(a) the gains of successive policies are monotone non-decreasing,
(b) if improvement occurs in step 3a of the algorithm, then it can only be in
transient states of dn+l, in which case ga,+l(s)> gd(S) where s is transient
under dn+ 1,
(c) if no improvement occurs in step 3a of the algorithm and it occurs in a
recurrent state of dn+l, in step 3b of the algorithm then gd.+l(s) > gd.(s) where
s is recurrent under dn+l, and
(d) if no improvement occurs in step 3a of the algorithm and it occurs in a
transient state of d,+~ in part 3b of the algorithm then ha,+l(s ) > hd,(s ) where s
is transient under d,+ 1.
In the special case that all policies are communicating, Haviv and Puterman
(1990) provide a modification of the unichain policy iteration algorithm which
avoids the pair of optimality equations.
At present, non-trivial conditions which imply the convergence of policy
iteration in the non-finite case are not available. Dekker (1985, pp. 109-110)
provides an example with finite states and compact actions in which an infinite
number of improvements occur in step 3a and converge to a suboptimal policy.
In it, the limiting policy has different ergodic classes than the optimal policy
and since improvements through 3a cannot create new ergodic classes, the
algorithm will not converge to an optimal policy.
8.9. Sensitive discount optimality

If (g*, h*) is a solution of the multichain average reward optimality equa-
tions (8.51) and (8.52) then any stationary policy which uses an h*-improving
decision rule is average optimal. Such a policy has the greatest bias among all
stationary policies with the same chain structure as d* but simple examples (cf.
Blackwell (1962), Denardo (1973)) show that such a policy need not have the
largest bias among all average optimal policies.
A policy with the largest bias among all average optimal policies is said to be
bias-optimal. Veinott (1966), Denardo (1970) and Kallenberg (1983) provide
methods for obtaining such policies in the finite state and action setting. Sheu
and Farn (1980) and Mann (1983) have also studied this criterion.
Since bias-optimal policies need not be unique, a decision maker might wish
to have some way of selecting a 'best' bias-optimal policy. Veinott (1969)
introduced the concept of sensitive discount optimality and using the Laurent
series expansion (8.7), showed that it provided a link between average
optimality, bias-optimality and Blackwell optimality. Contributors to this
theory include Miller and Veinott (1969), Veinott (1974), Hordijk and Sladky
(1977), Wijngaard (1977), van der Wal (1981), Federgruen and Schweitzer
(1984a) and Dekker (1985).
This section presents the theory of sensitive discount optimality in the finite
state and action case.
8.9.1. Discount optimality criteria

Recall from (5.12) that a policy 7r* is said to be n-discount optimal for
n=-l,0,1,.., if
lim inf (1 - A)-"[v;" - v~]/>0
for all ~r E H. This criterion can be reexpressed in terms of the interest rate
p = 1/(1 - A) (Section 8.1.2) as follows:
n
liminf p [v~~ * - v~]~>0. (8.60)
p$0
Let A be an arbitrary matrix. Then A is said to be lexicographically positive~

written A > 0, if A is not identically 0 and the first non-zero entry in each row
of A is strictly positive. If A and B are two matrices of the same dimension~ A
is said to be lexicographically greater than B written A > B if A - B is
lexicographically positive.
Let d and e denote two stationary policies. Using (8.7)~ the difference in
their discounted rewards can be written as
va -- va ----7 + ha -- he + ,=1 (_._p)n[y~ .....ye] (8.61)
where y~ is the nth term in the Laurent series expansion corresponding to

stationary policy 6 = d or e. Multiplying (8.61) by p and taking the lim inf as p
decreases to zero implies that
lira inf p[v~a - oh] i > 0 (8.62)

if and only if ga >~ge" When ga = ge, the limit in (8.62) will be zero, in which
case
d e
hm mf [v, - oh] ~> 0 (8.63)
p$0
if and only if h~ >/heB . if gd(s) > ge(S) for some s in S, then (8.63) will hold with
strictly inequality in component s, regardless of the values of h~(s) and heB (s)
(these quantities are defined in Section 8.3.1).
Similarily, if gd = ge and h~ = heB, then the lim inf's in both (8.62) and (8.63)
will be zero and
e
lim inf p[ vaa -- v xl >~0 (8.64)
050
if and only if y~ ~ y~. Conversely the sth component of (8.64) will be strictly
positive if either
(a) gd(s) > ge(s), or
(b) gd(S) = ge(S) and h~(s) > h~(s), or
(c) gd(S) = ge(s) and h~(s) = hBe(S) and y~(s) > ye~(s)o
These arguments can be repeated indefinitely to demonstrate that:
(1) The larger the value of n, the more selective the discount optimal
criteria. That is, if D* denotes the set of n-discount optimal stationary policies,
thenD* DD~* f o r n = 0 , 1 ,
n--1 . . . .
(2) A policy is (-1)-discount optimal if it maximizes the average reward,

0-discount optimal if maximizes the bias among all average optimal policies,
1-discount optimal if it maximizes the third term in the Laurent expansion
among all policies that are bias optimal, etc.
(3) A stationary policy d is n-discount optimal among the class of stationary
for all stationary policies, e, where v~d , n + 2 ls
d n + 2 e n + 2 - -
pohcles If v~' > v~' the
ISI x (n + 2) matrix with columns given by y,~, m = - 1 , 0 . . . . , n.
Another important consequence of (8.61) is that a stationary policy is
- d ~ coo
Blackwell optimal if v A' > v h" for all stationary policies e. That is, a Blackwell
optimal policy is n-discount optimal for all n and consequently is more selective
than any of the discount optimality criteria. If for some finite N, D~ contains a
single decision rule, then the stationary policy corresponding to that decisio~
rule is n-discount optimal for all n >i N and is Blackwell optimal.
The following result is an immediate consequence of Theorem 8.25.
Theorem 8.30. f l A , f o r each s ~ S and S are finite, then there exists" a stationary
n-discount optimal policy for each n.
8.9.2. Optimality equations

Corresponding to the n-discount optimality criteria are a series of nested
optimality equations extending those in Section 8.7. They are obtained by
substituting the Laurent expansion of an n-discount optimal policy into the

discount optimality equation and equating terms in like powers of p to obtain
max {(Pd - I ) Y - 1 } = 0, (8.65)

dED_ 1
max {r e -- y 1 Jv (Pa - I)y0} = 0 , (8.66)

d@D o
and for n/> 1,
max { - y , _ l + (Pa - I)Yn} =- 0 (8.67)

dED n
with D 1 D and for n->-0, Dn = { d E D n 1: d attains the max in equation

=
n - 1 at Y n - ~ when (Y-l, Y0, Yl,- - , Y,-~) satisfy equations m =

- 1 , 0 , . . . , n } . Note that equations (8.65) and (8.66) are the optimality
equations for multichain average reward problems.
The following result links solutions of the optimality equations to the terms
in the Laurent series expansion of Blackwell optimal policies (Dekker, 1985).
Theorem 8.31. Let d be a Blackwell optimal policy. Then for all n~

(yd_l ' ydo, . . . , yd) is a solution o f the first n + 2 optimality equations and is the
unique solution to the first n + 3 optimality equations where Yn +1 is arbitrary.
Further if ( Y - l , Yo, , Y n ) is any solution to the first n + 2 optimality equao
tions, then any d E Dn ~ is n-discount optimal.
In the case of bias-optimality the results simplify as follows.
Corollary 8.32. Suppose (Y*-,I, Y~, Y~) is a solution o f equations (8.65)-(8.67)

with n = 1. Then
(a) any decision rule d in D 1 which maximizes PdY~ is bias optimal~ and
(b) Y*-I and y~ are unique and equal the gain and bias o f any bias optimal
policy.
To compute n-discount optimal policies requires solution of the optimality

equations. Policy iteration methods are the most direct. Linear programming
methods are available for n = - 1 and n = 0; Federgruen and Schweitzer (1984)
have proposed a value iteration scheme.

The Policy Iteration Algorithm can be used to constructively establish the
existence of and compute n-discount optimal and Blackwell optimal policies,
The approach is to find the set of all -1-discount optimal policies and then
within this set, to find the set of all 0-discount optimal policies and continue
this process until the set of n-discount optimal policies is determined. The basic
420 M . L . Puterman .
algorithm is very similar to the multichain average reward policy iteration

algorithm.
m
For notational convenience, set r d = r d if m = 0, and 0 otherwise.
Policy Iteration--N-discount optimality.

1. Set m = - 1 , D_ 1 = D and Y*-z = 0 .
2. Set n = 0 and select a d, ~ D,,.
3. (Policy evaluation)
" Obtain
. yin d, by solving
d,, and Ym+l
r din. - - Y m, - ~ + ( p d n -- I ) y m = 0 , (8.68)
rin+l
d,, -- Yin + (Pd. -- l ) y m +1 = O, (8.69)
subject to Ym+l(S)= 0 for one s in each recurrent class of jog..
4. (Policy improvement)
(a) Choose a d~+l @ Din to satisfy
rd~+l + Pd,,+lydm~ = max

d~D m
{r'ff + Pdy d"} (8.70)
and set d,,+~ = d, if possible. If d n + ~ = d,,, set E'~ = { d E D m : d attains the

maximum in (8.70)} and go to (b), otherwise increment n by 1 and return to
step 3.
(b) Choose a d,+ 1 E E~ to satisfy
dn m+l dn
rd~+~m-~lAv p d . + l y m + 1 = deD.~max l r d + PdYm+l} (8.71)
and set d n+1 = d~ if possible. If d,,+l = d, go to 5, otherwise increment n by 1

and return to step 3.
5. Set Din+ 1 -= { d ' ~ Dr,: d ' attains the maximum in (8.71)}. If Din+ a is a
singleton or m = N, stop. Otherwise, set y * = yam' where d' E Din+~, increment
m by 1 and return to 2.
The algorithm terminates with, D . ~ , the set of N-discount stationary

optimal policies. When N :- - 1 , that is an average optimal policy is sought, the
above algorithm is identical to that in Section 8.8. The specification to
dn
determine Ym+l uniquely in Step 3 can be implemented by solving the system of
three equations given by (8.68), (8.69) and the identical equation to (8.69) at
m + 2 (Veinott, 1969).
Proof of convergence of the algorithm is similar to that in Section 8.8. For
each fixed m, the algorithm finds a sequence of policies that are monotonically
increasing in sense of terms of
m
limm fp
[v~d n + 1 - v ~ ~]~-0.
p$o
Since there are only finitely many stationary policies, this step terminates with
Dm+~, the set of m-discount optimal stationary policies. When m = N, the
algorithm terminates with the set of N-discount optimal policies.
Since Blackwell optimality corresponds to o0-discount optimality, the above
suggests that an infinite number of passes throught the above policy iteration
algorithm is necessary to obtain a Blackwell optimal stationary policy. Miller
and Veinott (1968) showed that this is not the case.
Theorem 8.33. Suppose d* is N-discount optimal where N is the number of

states in S. Then d* is Blackwell optimal.
Veinott (1974) and Lamond (1986) showed that the N in Theorem 8.33 can
be replaced by N - k, where k is number of recurrent classes in an ( N - k)-
discount optimal policy. The following immediate corollary to the above
theorem ties together many of the results in Section 8.
Corollary 8.34. Suppose A s for each s ~ S and S are finite. Then there exists a
stationary Blackwell optimal policy which can be determined by the policy
iteration algorithm in a finite number of iterations.
Veinott (1966), Denardo and Miller (1968), Lippman (1968), Sladky (1974),
Hordijk and Sladky (1977), Denardo and Rothblum (1979) and van der Wal
(1981) have investigated the relationship between discount optimality, overtak-
ing optimality and average optimality as defined in Section 5.
8.10. Computational results

This section illustrates policy iteration and value iteration in the undis-~
counted case by using these algorithms to find average optimal policies for an
infinite horizon version of the stochastic inventory model of Section 3.2. The
chain structure of stationary policies determines whether the single equation
policy iteration scheme of Section 8.3 or the multiple equation method of
Section 8.8 is appropriate and whether value iteration as defined in Section 8.4.
is convergent. This structure is investigated first.
Since self transitions are possible under all actions, all policies aie aperiodic.
Consider the stationary policy d = (0, 2, 1, 0) corresponding to the following
ordering rule. If the inventory at the start of the period is 0 units, then no
order is placed and if the inventory is 1 or more units, an order is placed which
instantaneously raises the stock level to 3 units. Although such a policy is
feasible, it is clearly impractical and non-optimal. This stationary policy is not
unichain because it partitions the state space into two recurrent classes: {0}
and {1, 2, 3}. All remaining stationary policies are unichain.
Even though the set of stationary policies does not satisfy the unichain
assumption, this problem is communicating since each state can be reached
from each other state with positive probability in a finite number of iterations
422 M.L. Puterman
under some policy. Consequently value iteration and modified policy iteration
are convergent. Since not all policies are unichain the multichain version of
policy iteration is required.

The algorithm in Section 8.8 is applied with d o = (0, 2, 1, 0). Solving the first
evaluation equation yields
gdo = and h Bdo~-

:il
Coincidentally, gdo is constant so that in the improvement stage of the
algorithm, equation (8.56) is superfluous. If the data is perturbed slightly, then
gdo will not be constant. This means that at the first pass through the algorithm
an improved policy will be obtained through (8.56) so that the first improve-
ment equation is necessary. Results (including the gain and bias) are reported
in Table 8.1.
Table 8.1
Policy iteration results
n gd,, (h~.(s)) d.(s)
s=0 s=l s=2 s=3 s=0 s=l s=2 s=3
0 0 0 0
0 0 2 l 0
(0) (-3.0) (-1.0) (5.0)
0 0 0 0
1 0 0 0 0
(0) (6.6667) (12.4444) (17.1852)
1.6 1.6 1.6 1.6
2 3 2 0 0
(-5.08) (-3.08) (2.12) (4.92)
2.2045 2.2045 2.2045 2.2045
3 3 0 0 0
(-4.2665) (-0.5393) (3.2789 (5.7335)
4 3 0 0 0
Since the average optimal policy is unique, it is n-discount optimal for all n.
Thus it is Blackwell optimal and discount optimal for all discount factors
sufficiently close to 1. It agrees with that found in the discounted case with
h=0.9.
8.10.2. Value iteration

Since all policies are aperiodic and communicating, value iteration and its
bounds are convergent. Table 8.2 reports the results of applying this algorithm
to the inventory example. The upper and lower bounds in (8.39) are the basis
for a stopping rule; An denotes the difference between these two bounds. A
0.01-optimal solution is sought and v is chosen to be (0, 0, 0, 0).
Table 8.2
Value iteration results
n v"(s) d'(s) A"
s=0 s=l s=2 s=3 s=0 s=l s=2 s=3
0 0 0 0 0 0 0 0 0
1 0 5.0 6.0 5.0 2 0 0 0 6.0000
2 2,0 6.25 10.0 10.50 3 0 0 0 4.2500
3 4.1875 8.0625 12.125 14.1875 3 0 0 0 1.8750
4 6.625 10.1563 14.1094 16.625 3 0 0 0 0.4531
5 8.75 12.5078 16.2617 18.75 3 0 0 0 0.1465
6 10.9453 14.6895 18.5068 20.9453 3 0 0 0 0.0635
7 13.1621 16.8813 20.7078 23.1621 3 0 0 0 0.0164
8 15.3647 19.0919 22.9081 25.3647 3 0 0 0 0.0103
9 17.5682 21.2965 25.1142 27.5685 3 0 0 0 0.0025
Observe that after 9 iterations, the difference between the bounds is 0.0025
so that the decision rule ( 3 , 0 , 0 , 0) (which is the unique optimal policy
identified by policy iteration) is guaranteed to have a gain that is within 0.0025
of optimum. That is, O<~g*--gdg<~O.O025. Estimates of g* and h* are
obtained using the method described in Section 8.4.1. That is
F2.2035] -2.2633-
g* ~ v 9 - v 8 = / 2"2046] h* ~ O9 1.4551
|2.2060l and - 9g* = 5.2602
/2.20353 7.7370
These calculations imply that 2 . 2 0 3 5 ~ g * ~<2.2060 which agrees with the

precise value of g* = 2.2045 obtained by solving the policy evaluation equa-
tions in the policy iteration algorithm. The values of h* are considerably
different those of policy iteration, but the differences of h* between any two
states are accurately estimated and agree with the exact values. For example,
comparing h* to its value at state 0 yields h * ( 1 ) - h * ( 0 ) = 4.7184, h*(2) ....
h*(0) =7.5235 and h * ( 3 ) - h * ( 0 ) = 10.0000 which is identical to the exact
values. Thus value iteration provides an accurate way of finding relative values.
If the precise bias is desired, the policy evaluation equations may be solved.
9. Semi-Markov decision processes and Markov renewal p r o g r a m m i n g
In the M D P ' s considered in the previous sections, the decision maker

chooses actions at a fixed discrete set of time points. Semi-Markov decision
processes (SMDP's) generalize M D P ' s by allowing the decision maker to
choose actions at a n y epoch necessitating observing the system in continuous'
time. The effect of choosing an action is to determine the probability distribw
tion of both the subsequent state and the remaining time in the current state
For a specified policy the system evolves by remaining in a state for a random
amount of time and then jumping to a different state. These models are called
semi-Markov because for fixed Markov policies the system states evolve
according to a semi-Markov process.
Analysis of SMDP's depends on the set of admissible policies or controls.
When actions are allowed at any tirfie, continuous time control theory methods
are appropriate. If actions can be chosen only immediately following transi-
tions and the time horizon is infinite, or a fixed number of transitions, discrete
time M D P methods can be adapted. When actions are allowed only after
transitions, the models are often referred to as Markov renewal programs
(MRP's), however not all authors distinguish MRP's and SMDP's in this way.
When the time horizon is finite and fixed, MRP's also require control theory
methods for analysis. This section presents results for infinite horizon Markov
renewal programs.
MRP's are mostly widely used to model equipment replacement, queueing
control and inventory control problems. Notable special cases are continuous
time MDP's (CTMDP's) in which the transition times are exponentially
distributed and MDP's in which all transition times are constant and equal.
Semi-Markov decision processes were introduced by Jewell (1963), other
contributors to the theory include H o w a r d (1963), de Cani (1964), Schweitzer
(1965, 1971), Fox (1966) and D e n a r d o (1971). They are also treated in the
books of Kallenberg (1983) and H e y m a n and Sobel (1984). Numerous papers,
most notably in the infinite horizon average reward case, have been based on
applying a transformation (Schweitzer, 1971) to convert SMDP's to MDP's. A
good reference on semi-Markov processes is ~inlar (1975).
9.1. Problem formulation

Let S be either a finite or countable set of states. Whenever a transition
occurs into state s, the decision maker chooses an action from the set A s. As a
consequence of choosing action a in state s, the probability that the system is in
state j at the next transition is p ( ] I s, a), and the probability that a transition
occurs within t units of time is F(t]s, a). It is assumed that there exists a
constant y, 0 < y < 1 such that
F(0[s,a)<~,y for a l l a 6 2 A ~ a n d s ~ S . (9.1)
The transition structure can be expressed in terms of the joint probability

q(t, j] s, a ) = p ( ] Is, a)F(t] s, a). Alternatively the problem can be defined in
terms of q in which case p and F can be derived.
The economic consequence of choosing action a in state s is that the decision
maker receives a lump sum reward k(s, a), and a continuous reward at rate
c(s, a) per unit time as long as the system is in state s.
A Markov decision rule d is a function such that d(s) 62 A s for each s 62 S.
For a specified d let Pd(][S) = p ( j l s , d(s)), Fd(tls ) = F(tls , d(s)), qd(t]s, ]) =
q(tls, d(s), j), ka(s ) = k(s, d(s)) and ca(s ) = c(s, d(s)). T o avoid technicalities,
it will be assumed that the decision making period begins at the time of the first
transition and no rewards are received until that time. Let 7r = (da, d 2 , . . . ) be
an arbitrary policy which corresponds to using d n at the time of the nth
transition. Corresponding to this policy is a semi-Markov process { Y~; t/> 0}
which gives the state of the system at time t and a process {U~; t/> 0} which
gives the action chosen at time t. This process can be further described in terms
of a sequence of jointly distributed random variables {(X~, r~); n = 1, 2 , . . . }
where X2 is the state of the system immediately following the nth transition
and r~ is length of time the system is in state X~. It is convenient to define o-~
as the total time until the nth transition starting at the time of the first
transition, that is
n -- 1
qr E
O- n ~--- 3" n
j=l
and o-~ = 0.
9.2. The discounted case

Let a > 0 denote the continuous time interest rate. That is, the present value
of one dollar received at time t is e -at. For policy ~r, the expected infinite
horizon discounted reward given that the first transition is into state s is
denoted by v~(s) and given by
v2(s ) = e c ( Y t , U~) dt + e k(Y~, U~,j) . (9.2)

n=l
The first term in (9.2) corresponds to the continuous portion of the reward and
the second term to the fixed reward received only at decision epochs.
The objective in this problem is to characterize
v*(s) = sup v'~(s)

~rEfl
for all s ~ S and to find a policy ,~r* with the property that
(s) = v2(s)
for all s ~ S.
This problem is transformed to a discrete time problem by allowing the
discount factor to be state and action dependent and analyzing the problem in
terms of its embedded chain. Define ra(s ) to be the expected total discounted
reward until the next transition if the system just entered state s and decision
rule d is selected. It is given by
re(s ) = k a ( s ) + cd(s)Ea~LJo[f e T M dt ]
= ke(s ) + ca(s){a-2(1 - E,d[e-~'d])} (9.3)
where ~.e is the time until the first transition given that the system just entered
state s and decision rule d is used. Define the expected discounted holding time
in state s if action a is selected by
Ms, a) = f o e ~' d F ( t l s , a) . (9.4)
For d C D, define Ad(S) = A(s, d(s)). Note that 2te(S) is the Laplace transform of
r e. Thus
re(s ) = kd(s ) + ce(s){a-111 _ Aa(s)] } .
Using this quantity, v~(s) can be re-expressed as
v~ s = E s
d1
ETrr --~'r 7r'
= tel(s) + le v~ (Xz) ]
= re,(s) +
j~s
where 7r' = (d2, d3, . . .). For a stationary policy d, the infinite horizon expec-
ted total discounted reward can be obtained by solving the equation
v~(s) = re(s ) + ~ A a ( s ) p a ( j l s ) v ~ ( j ) o
jeS
This can be expressed in matrix terms as
v = r e + Mev (9.5)
where M e is the matrix with entries Z d ( s ) p a ( j l s ) . This differs from the discrete
time evaluation equation (6.8) by the state and action dependent discount rate
(9.4). From a computational point of view this causes the matrix M a to have
unequal row sums bounded by A* = SUpa,,A(S, a). The efficiency of numerical
methods for solution of (9.5) has been investigated by Porteus (1980a, 1983)~
The optimality equation for Markov renewal programs is given by
v = max {r d + May } ~ T v . (9.6)

dED
If (9.1) holds, ~* < 1 and if in addition IIr~ll ~ M < m for all d E D, T defined
in (9.6) is a contraction operator on the space of bounded real valued functions
on S, so consequently all results of Section 6 apply. This means that for MRP's:
(1) The optimality equation has a unique solution v*.
(2) There exists a stationary policy which is optimal.
(3) The problem can be solved by value iteration, policy iteration, modified
policy iteration, linear programming and their variants.
(4) Bounds and action elimination methods are valid.
(5) Extensions to unbounded rewards are possible.
9.3. The average reward case

Let T be the time the system has been under observation since the first
decision epoch. For a fixed zr E H, define the expected total reward up to time
T by
I-t"T ~'% 1
:LJ0 z
rt=l n
(97
where v r is the random variable representing the number of decisions made up
to time T using policy 7r. For each zr E H, define the average expected reward
or gain by
~-.~inf ~v;(s),
g~(s) = lim 1 s~s.
The objective in the average reward case is to characterize the optimal average
expected reward
g*(s) = sup g'~(s) , s E S,

"n'CH
and determine a policy ~r* with the property that
g=~ (s) = g*(s) , sE S .
Let rd(s ) be the expected total reward until the next transitior~ when the
system is in state s and decision rule d is used. It is given by
r~(s) = k~(s) + c~(S)es~[~] = k~(s) + c~(s)H~(s) (9.8)
where Ha(s ) is defined as follows. For each a E A , and s ~ S, tile expected

time in state s until the next transition, H(s, a) is given by
H(s, a) ~
F t d F ( t i s , a)~ (9.9)
For d E D, define He(S ) ~- H(s, d(s)). Under (9.1), "0 ~: inf~,,H(s, a) > 0o
428 M.L. Puterman
The gain of a fixed, stationary policy d is uniquely determined by the

equations
(Pd-l)g=O and ru-gH a+(Pa-l)h=O. (9.10)
The second equation in (9.10) uniquely determines h up to an element in the

null space of Pa - I. The derivation of (9.10) is based on the partial Laurent
series expansion of v~ (Denardo, 1971).
The corresponding optimality equations are
max{(Pa-l)g)=0 and max{r a - g H a + ( P a - I ) h ) = O

dED dEE
(9.11)
where E = {d E D: Pdg = g}.

Both the evaluation equations and the optimality equations differ from the
MDP case by the inclusion of the term H a in the second equation. These
equations can be solved using the policy iteration methods of Section 8. The
only modification is that H(s, a) must be evaluated for each state-action pair.
To obtain other theoretical results and establish the convergence of value
iteration, the problem can be converted into an 'equivalent' MDP by applying
the following transformation (Schweitzer, 1971).
Define a transformed MDP indicated by " - " as follows:
F(s, a) = r(s, a)/H(s, a)

and
~(jis, a) = ~l'[p(jls, a) - 6(j[s)]/H(s, a) + 6(jls)
where 6(j[s)= 1 if j = s and 0 otherwise and
0 < 77' < H(s, a)/(1 - P(SiS, a))
for all a ~ A s and s E S for which p(s Is, a) < 1.

The choice of "q' ensures that/7(s Is, a) > 0 so that all stationary policies have
aperiodic chains. The sets of optimal policies for the original and transformed
problems are identical and g* = g* and h = ~/'h.
Because of this, the following results hold for the Markov renewal program.
ruing problem with average reward criteria:
(1) Whenever (8.23) holds,
(a) the first optimality equation is redundant and there exists a solution to
the second optimality equation, and
(b) there exists an optimal stationary policy.
(2) When S is finite, value iteration converges in the sense of Section 8.4, it
can be used to determine optimal policies, and bounds are available.
(3) Linear programming can be used to determine optimal policies in the
finite state and action case.
Bibliography
Bather, J. (1975). Optimal decision procedures for finite Markov chains. Adv. Appl. Probab. 5,
328-339, 521-540, 541-553.
Bellman, R.E. (1957). Dynamic Programming. Princeton University Press, Princeton, NJ.
Bertsekas, D.P. (1987). Dynamic Programming, Deterministic and Stochastic Models. Prentice-
Hall, Englewood Cliffs, NJ.
Blackwell, D. (1961). On the functional equation of dynamic programming. J. Math. Anal. Appl.
2, 273-276.
Blackwell, D. (1962). Discrete dynamic programming. Ann. Math. Statist. 35, 719-726.
Blackwell, D. (1965). Discounted dynamic programming. Ann. Math. Statist. 36, 226-235.
Blackwell, D. (1967). Positive dynamic programming. Proc. 5th Berkeley Symp. Mathematical
Statistics and Probability 1, 415-418.
Brown, B.W. (1965). On the iterative method of dynamic programming on a finite space discrete
Markov process. Ann. Math. Statist. 36, 1279-1286.
~inclar, E. (1975). Introduction to Stochastic Processes. Prentice-Hall, Englewood Cliffs, NJ.
De Cani, J.S. (1964). A dynamic programming algorithm for embedded Markov chains when the
planning horizon is at infinity. Management Sci. 10, 716-733.
Dekker, R. (1985). Denumerable Markov decision chains: Optimal policies for small interest
rates. Unpublished Ph.D. Dissertation, University of Leiden.
Dembo, R. and M. Haviv (1984). Truncated policy iteration methods. Oper. Res. Lett. 3,
243 -246.
Demko, S. and T.P. Hill (1981). Decision processes with total cost criteria. Ann. Probab. 9,
293-301.
Denardo, E.V. (1967). Contraction mappings in the theory underlying dynamic programming.
S I A M Rev. 9, 169-177.
Denardo, E.V. and B. Fox (1968). Multichain Markov renewal programming. S l A M J. Appl.
Math. 16, 468-487.
Denardo, E.V. and B.L. Miller (1968). An optimality condition for discrete dynamic programming
with no discounting. Ann. Math. Statist. 39, 1220-1227.
Denardo, E.V. (1970). Computing a bias-optimal policy in a discrete-time Markov decision
problem. Oper. Res. 18, 279-289.
Denardo, E.V. (1971). Markov renewal programs with small interest rates. Ann. Math. Statist. 42,
477 -496.
Denardo, E.V. (1973). A Markov decision problem. In: T.C. Hu and S.M. Robinson (Eds.),
Mathematical Programming. Academic Press, New York.
Denardo, E.V. and U.G. Rothblum (1979). Overtaking optimality for Markov decision chains
Math. Oper. Res. 4, 144-152.
Denardo, E.V. (1982). Dynamic Programming, Models and Applications. Prentice-Hall, Engle-
wood Cliffs, NJ.
D'Epenoux, F. (1963). Sur un probl~me de production et de stockage dans l'aleatoire. Rev.
Francaise Automat. lnformat. Rech. Oper. 14 (English Transl.: Management Sci. 10, 98-108).
Deppe, H. (1984). On the existence of average optimal policies in semi-regenerative decision
models. Math. Oper. Res. 9, 558-575.
Derman, C. (1966). Denumerable state Markovian decision processes--Average cost criterion.
Ann. Math. Statist. 37, 1545-1554.
Derman, C. and R. Straueh (1966). A note on memoryless rules for controlling sequential decision
processes. Ann. Math. Statist. 37, 276-278.
Derman, C. (1970). Finite state Markovian decision processes. Academic Press, New York.
Dirickx, Y.M.J. and M.R. Rao (1979). Linear programming methods for computing gain-optimal
policies in Markov decision models. Cah. Centre d'Etudes Rech. Oper. 21, 133-142.
Dubins, L.E. and L.J. Savage (1965). How to Gamble if You Must: Inequalities for Stochastic
Probesses. McGraw-Hill, New York.
Eagle, J.E. (1975). A Utility Criterion for the Markov Decision Process. Unpublished Ph.D
Dissertation, Dept. of Engineering-Economic Systems, Stanford University.
430 M.L. Puterman
Federgruen, A., A. Hordijk and H.C. Tijms (1978). Recurrence conditions in denumerable state
Markov decision processes. In: M.L. Puterman (Ed.), Dynamic Programming and Its Applica-
tions. Academic Press, New York, 3-22.
Federgruen, A. and P.J. Schweitzer (1978). Discounted and undiscounted value-iteration in
Markov decision problems: A survey. In: M.L. Puterman (Ed.), Dynamic Programming and Its
Applications. Academic Press, New York, 23-52.
Federgruen, A. and H.C. Tijms (1978). The optimality equation in average cost denumerable state
semi-Markov decision problems, recurrency conditions and algorithms. J. Appl. Probab. 15,
356-373.
Federgruen, A., A. Hordijk and H.C. Tijms (1979). Denumerable state semi-Markov decision
processes with unbounded costs, average cost criteria. Stochastic Process. Appl. 9, 223-235.
Federgruen, A. and P.J. Schweitzer (1980). A survey of asymptotic value-iteration for undis-
counted Markovian decision processes. In: R. Hartley, L.C. Thomas and D.J. White (Eds.),
Recent Developments in Markov Decision Processes. Academic Press, New York, 73-109.
Federgruen, A., P.J. Schweitzer and H.C. Tijms (1983). Denumerable undiscounted semi-Markov
decision processes with unbounded rewards. Math. Oper. Res. 8, 298-313.
Federgruen, A. and J.P. Schweitzer (1984a). Successive approximation methods for solving nested
functional equations in Markov decision problems. Math. Oper. Res. 9, 319-344.
Federgruen, A. and J.P. Schweitzer (1984b). A fixed point approach to undiscounted Markov
renewal programs. SIAM J. Algebraic Discrete Methods 5, 539-550.
Fisher, L. and S.M. Ross (1968). An example in denumerable decision processes. Ann. Math.
Statist. 39, 674-676.
Fleming, W.H. and R. Rishel (1975). Deterministic and Stochastic Optimal Control. Springer-
Verlag, New York.
Flynn, J. (1976). Conditions for the equivalence of optimality criteria in dynamic programming.
Ann. Statist. 4, 936-953.
Fox, B.L. (1966). Markov renewal programming by linear fractional programming. SIAM J. Appl.
Math. 14, 1418-1432.
Grinold, R. (1973). Elimination of suboptimal actions in Markov decision problems. Oper. Res.
21, 848-851.
Harrison, J.M. (1972). Discrete dynamic programming with unbounded rewards. Ann. Math.
Statist. 43, 636-644.
Hartley, R., L.C. Thomas and D.J. White (Eds.) (1980). Recent Developments in Markov Decision
Processes. Academic Press, New York.
Hartley, R. (1980). A simple proof of Whittle's bridging condition in dynamic programming. J.
Appl. Probab. 17, 1114-1116.
Hastings, N.A.J. (1968). Some note son dynamic programming and replacement, Oper. Res.
Quart. 19, 453-464.
Hastings, N.A.J. (1969). Optimization of discounted Markov decision problems. Oper. Res.
Quart. 20, 499-500.
Hastings, N.A.J. (1976)~ A test for suboptimal actions in undiscounted Markov decision chains.
Management Sci. 23, 87-91.
Hastings, N.A.J. and J.A.E.E. van Nunen (1977). The action elimination algorithm for Markov
decision processes. In: H.C. Tijms and J. Wessels (eds.), Markov Decision Theory, Mathemati-
cal Centre Tract No. 93. Mathematical Centre, Amsterdam, 161-170.
Haviv, M. and M.L. Puterman (1990). Improved policy iteration methods for communicating
Markov decision processes. Annals of Operations Research, Special Issue on Markov Decision
Processes, to appear.
Hernfindez-Lerma, O. (1989). Adaptive Markov Control Processes. Springer-Verlag, New York.
Heyman, D.P. and M.J. Sobel (1984). Stochastic Models in Operations Research, Vol. II.
McGraw-Hill, New York.
Hille, E. and R.S. Phillips (1957). Functional Analysis and Semi-Groups, American Mathematical
Society Colloquim Publications, Vol. 31. AMS, Providence, RI.
Hinderer, K. (1970). Foundations of Non-Stationary Dynamic Programming with Discrete Time
Parameter. Springer-Verlag, New York.
Hopp, W.J., J.C. Bean and R.L. Smith (1988). A new optimality criterion for non-homogeneous
Markov decision processes. Oper. Res. 35, 875,-883.
Hordijk, A. (1974). Dynamic Programming and Markov Potential Theory. Mathematical Centre
Tract No. 51. Mathematical Centre, Amsterdam.
Hordijk, A. and K. Sladky (1977). Sensitive optimality criteria in countable state dynamic
programming. Math. Oper. Res. 2, 1-14.
Hordijk, A. and L.C.M. Kallenberg (1979). Linear programming and Markov decision chains.
Management Sci. 25, 352-362.
Hordijk, A. and L.C.M. Kallenberg (1980). On solving Markov decision problems by linear
programming. In: R. Hartley, L.C. Thomas and D.J. White (Eds.), Recent Developments in
Markov Decision Processes. Academic Press, New York, 127-143.
Hordijk, A. and M.L. Puterman (1987). On the convergence of policy iteration in undiscounted
finite state Markov decision processes; the unichain case. Math. Oper. Res. 12, 163-176.
Howard, R. (1960). Dynamic Programming and Markov Processes. MIT Press, Cambridge, MA.
Howard, R.A. (1963). Semi-Markovian decision processes. Proc. Internat. Statist. Inst., Ottawa,
Canada.
Howard, R.A. and J.E. Matheson (1972). Risk sensitive Markov decision processes. Management
Sci. 8, 356-369.
Hubner, G. (1977). Improved procedures for eliminating suboptimal actions in Markov program
ruing by the use of contraction properties. Transactions of the Seventh Prague Conference on
Information Theory, Statistical Decision Functions, Random Processes, 257-263.
Hubner, G. (1988). A unified approach to adaptive control of average reward decision processes.
OR Spektrum 10, 161-166.
Jacquette, S.C. (1973). Markov decision processes with a new optimality condition: Discrete time.
Ann. Statist. 3, 496-505.
JeweU, W.S. (1963). Markov-renewal programming I: Formulation, finite return models; Ii:
Infinite return models, example. Oper. Res. 11, 938-971.
Kallenberg, L.C.M. (1983). Linear Programming and Finite Markov Control Problems, Mathe-
matical Centre Tract No. 148. Mathematical Centre, Amsterdam.
Kantorovich, L.V. (1952). Functional Analysis and Applied Mathematics, Translated by C.D.
Benster, NBS Report 1509, National Bureau of Standards, Los Angeles, CA.
Kemeny, J.G. and J.L. Snell (t960). Finite Markov Chains. Van Nostrand-Reinhold, New York.
Kreps, D.M. and E. Porteus (1977). On the optimality of structured policies in countable stage
decision processes, II: Positive and negative problems. SIAM J. Appl. Math. 32, 457-466.
Lamond, B.L. (1984). MDPLAB, an interactive computer program for Markov dynamic program-
ming. Working Paper 1068, Faculty of Commerce, University of British Columbia.
Lamond, B.L. (1986). Matrix methods in queueing and dynamic programming. Unpublished
Ph.D. Dissertation, Faculty of Commerce, University of British Columbia.
Lamond, B.L. and M.L. Puterman (1989). Generalized inverses in discrete time Markov decision
processes. SIAM J. Mat. Anal. Appl. 10, 118-134.
Lanery, E. (1967). Etude asymptotique des systrmes Markovien h commande. Rev. Fran~aise
Inform. Rech. Oper. 1, 3-56.
Lippman, S.A. (t968). Criterion equivalence in discrete dynamic programming. Oper. Res. 17,
920 -923.
Lippman, S.A. (1975). On Dynamic Programming with Unbounded Rewards. Management. Sci.
21, 1225-1233.
Liusternik, L. and V. Sobolev (1961). Elements of Functional Analysis. Ungar, New York.
MacQueen, J. (1966). A modified dynamic programming method for Markov decision problems. J.
Math. Anal. Appl. 14, 38-43.
Mandl, P. (1967). An iterative method for maximizing the characteristic root of positive matrices.
Rev. Roumaine Math. Pures Appl. 12, 1312-1317.
Mandl, P. (1974). Estimation and control in Markov chains. Adv. in AppL Probab. 6, 40-60~
Mann, E. (1983). Optimality equations and bias optimality in bounded Markov decision processes.
Preprint No. 574, University of Bonn.
Manne, A. (1960)~ Linear programming and sequential decisions. Management Sci. 6, 259-267
432 M.L. Puterman
Miller, B.L. and A.F. Veinott, Jr. (1969). Discrete dynamic programming with a small interest
rate. Ann. Math. Statist. 40, 366-370.
Mine, H. and S. Osaki (1968). Some remarks on a Markovian decision process with an absorbing
state. J. Math. Anal. Appl. 23, 327-333.
Monahan, G.E. (1982). A survey of partially observabel Markov decision processes: Theory,
models, and algorithms. Management Sci. 28, 1-16.
Morton, T.E. (1971). On the asymptotic convergence rate Of cost differences for Markovian
decision processes. Oper. Res. 19, 244-248.
Morton, T.E. and W.E. Wecker (1977). Discounting ergodicity and convergence for Markov
decision processes. Management Sci. 23, 890-900.
Morton, T. (1978). The non-stationary infinite horizon inventory problem. Management Sci. 24,
1474-1482.
Odoni, A.R. (1969). On finding the maximal gain for Markov decision processes. Oper. Res~ 17,
857-860.
Ohno, K. (1985). Modified policy iteration algorithm with nonoptimality tests for undiscounted
Markov decision processes. Working Paper, Dept. of Information System and Management
Science, Konan University, Japan.
Ohno, K. and K. Ichiki (1987). Computing optimnal policies for tandem queueing systems. Oper.
Res. 35, 121-126.
Ornstein, D. (1969). On the existence of stationary optimal strategies. Proc. Amer. Math. Soc. 20,
563 -569.
Ortega, J.M. and W.C. Rheinboldt (1970). Iterative Solutions of Nonlinear equations in Several
Variables. Academic Press, New York.
Platzman, L. (1977). Improved conditions for convergence in undiscounted Markov renewal
programming. Oper. Res. 25, 529-533.
Pliska, S.R. (1976). Optimization of multitype branching processes. Management Sci. 23,
117-125.
Pliska, S.R. (1978). On the transient case for Markov decision processes with general state spaces.
In: M.L. Puterrnan (Ed.), Dynamic Programming and Its Application. Academic Press, New
York, 335-350.
Porteus, E. (1971). Some bounds for discounted sequential decision processes. Management Sci.
18, 7-11.
Porteus, E. and J. Totten (1978). Accelerated computation of the expected discounted return in a
Markov chain. Oper. Res. 26, 350-358.
Porteus, E. (1980a). Improved iterative computation of the expected discounted return in Markov
and semi-Markov chains. Z. Oper. Res. 24, 155-170.
Porteus, E. (1980b). Overview of iterative methods for discounted finite Markov and semi-Markov
decision chains. In: R. Hartley, L.C. Thomas and D.J. White (Eds.), Recent Developments in
Markov Decision Processes. Academic Press, New York, 1-20.
Porteus, E. (1981). Computing the discounted return in Markov and semi-Markov chains. Naval
Res. Logist. Quart. 28, 567-578.
Porteus, E. (1983). Survey of numerical methods for discounted finite Markov and semi-Markov
chains. Presented at Twelfth Conference on Stochastic Processes and Their Applications, Ithaca~
NY.
Puterman, M.L. (Ed.) (1978). Dynamic Programming and Its Applications. Academic Press, New
York.
Puterman, M.L. and S.L. Brumelle (19'78). The analytic theory of policy iteration. In: M.L.
Puterman (ed.), Dynamic Programming and Its Application. Academic Press, New York.
Puterman, M.L. and M.C. Shin (1978). Modified policy iteration algorithms for discounted
Markov decision problems. Management Sci. 24, 1127-1137.
Puterman, M.L. and S.L. Brumelle (1979). On the convergence and policy iteration in stationary
dynamic programming. Math. Oper. Res. 4, 60-69.
Puterman M.L. and M.C. Shin (1982). Action elimination procedures for modified policy
iteration algorithms. Oper. Res. 30, 301-318.
Puterman, M.L. (1991). Markov Decision Processes. Wiley, New York.
Ross, S. (1968a). Non-Discounted denumerable Markovian decision models. Ann. Math. Statist.
39, 412-423.
Ross, S.M. (1968b). Arbitrary state Markovian decision processes. Ann. Math. Statist. 39,
2118-2122.
Ross, S.M. (1983). Introduction to Stochastic Dynamic Programming. Academic Press, New York.
Rothblum, U.G. and A.F. Veinott, Jr. (1975). Cumulative average optimality for normalized
Markov decision chains. Working Paper, Dept. of Operations Research, Stanford University.
Rothblum, U.G. (1979). Iterated successive approximation for sequential decision processes. In:
J.W.B. van Overhagen and H.C. Tijms, (Eds.), Stochastic Control and Optimization. Vrije
Universiteit, Amsterdam, 30-32.
Rothblum, U.G. (1984). Multiplicative Markov decision chains. Math. Oper. Res. 9, 6-24.
Schal, M. (1975). Conditions for optimality in dynamic programming and for the limit of n-stage
optimal policies to be optimal. Z. Wahrsch. Verw. Gebiete 32, 179-196.
Schweitzer, P.J. (1965). Perturbation theory and Markov decision chains. Unpublished Ph.D.
Dissertation, Massachusetts Institute of Technology.
Schweitzer, P.J. (1971). Iterative solution of the functional equations of undiscounted Markov
renewal programming. J. Math. Anal. Appl. 34, 495-501.
Schweitzer, P.J. and A. Federgruen (1977). The asymptotic behavior o1 undiscounted value
iteration in Markov decision problems. Math. Oper. Res. 2, 360-381.
Schweitzer, P.J. and A~ Federgruen (1978). The functional equations of undiscounted Markov
renewal programming. Math. Oper. Res. 3, 308-321.
Schweitzer, P.J. and A. Federgruen (1979). Geometric convergence of value iteration in multich-
ain Markov decision problems. Adv. in Appl. Probab. 11, 188-217.
Schweitzer, P.J. (1985). On undiscounted Markovian decision processes with compact action
spaces. Rev. R A I R O Rech. Oper. 19, 71-86.
Seneta, E. (1981). Non-negative Matrices and Markov Chains. Springer-Verlag, New York.
Shapiro, J. (1968). Turnpike planning horizons for a Markovian decision model. Management Sci.
14, 292-300.
Shapley, L.S. (1953). Stochastic games. Proc. Nat. Acad. Sci. U.S.A. 39, 1095-1100.
Sheu, S.S. and K.-J. Farn (1980). A sufficient condition for the existence of a stationary 1-optimal
plan in compact action Markovian decision processes. In: R. Hartley, L.C. Thomas and D.J.
White (Eds.), Recent Developments in Markov Decision Processes. Academic Press, New York,
111-126.
Sladky, K. (1974). On the set of optimal controls for Markov chains with rewards. Kybernetika 10,
350-367.
Smallwood, R. and E. Sondik (1973). The optimal control of partially observable Markov
processes over a finite horizon. Oper. Res. 21, 1071-1088.
Sobel, M.J. (1982). The variance of discounted Markov decision processes. J. Appl. Probab. 19,
794 -802.
Sondik, E.J. (1971). The optimal control of partially observable Markov processes. Ph.D.
Dissertation, Department of Engineering-Economic Systems, Stanford University.
Sondik, E. (1978). The optimal control of Partially observable Markov processes over the infinite
horizon: Discounted costs. Oper. Res. 26, 282-304.
Strauch, R. (1966). Negative dynamic programming. Ann. Math. Statist. 37, 871-890.
Taylor, H.M. (1965). Markovian sequential replacement processes. Ann. Math. Statist. 36,
1677-1694.
Tijms, H.C. and J. Wessels (eds.) (1977). Markov Decision Theory. Tract 93, Mathematical
Centre, Amsterdam.
van Dawen, R. (1986a). Finite state dynamic programming with the total reward criterion. Z~
Oper. Res. 30, A1-A14.
van Dawen, R. (1986b). Pointwise and uniformly good stationary strategies in dynamic program
ruing models. Math. Oper. Res. 11, 521-535.
van der Wal, J. and J.A.E.E. van Nunen (1977). A note on the convergence of the value oriented
successive approximations method. COSO Note R 77-05, Department of Mathematics, Eind~
hoven University of Technology.
434 M.L. Puterman
van der Wal, J. (1984). Stochastic Dynamic Programming. Tract 139, Mathematical Centre,
Amsterdam.
van der Wal, J. (1984). On stationary strategies in countable state total reward Markov decision
processes. Math. Oper. Res. 9, 290-300.
van Hee, K. (1978). Markov strategies in dynamic programming. Math. Oper. Res. 3, 37-41.
van Nunen, J.A.E.E. (1976a). A set of successive approximation methods for discounted Marko-
vian decision problems. Z. Oper. Res. 20, 203-208.
van Nunen, J.A.E.E. (1976b). Contracting Markov Decision Processes. Tract 71, Mathematical
Centre, Amsterdam.
van Nurlen, J.A.E.E. and J. Wessels (1978). A note on dynamic programming with unbounded
rewards. Management Sci. 24, 576-560.
Veinott, Jr., A.F. (1966). On finding optimal in discrete dynamic programming with no discount
ing. Ann. Math. Statist. 37, 1284-1294.
Veinott, Jr., A.F. (1968). Extreme points of Leontief substitution systems. Linear Algebra AppI. 1,
181-194.
Veinott, Jr., A.F. (1969). On discrete dynamic programming with sensitive discount optimality
criteria. Ann. Math. Statist. 40, 1635-1660.
Veinott, Jr., A.F. (1974). Markov decision chains. In: G.B. Dantzig and B.C. Eaves (Eds.),
Studies in Optimization. American Mathematical Association, Providence, RI.
White, D.J. (1963). Dynamic programming, Markov chains, and the method of successive
approximations. J. Math. Anal. Appl. 6, 373-376.
White, D.J. (1978). Elimination of non-optimal actions in Markov decision processes. In: M.L.
Puterman (Ed.), Dynamic Programming and Its Applications. Academic Press, New York,
131-160.
White, D.J. (1985a). Monotone value iteration for discounted finite Markov decision processes. J.
Math. Anal. Appl. 109, 311-324.
White, D.J. (1985b). Real applications of Markov decision processes. Interfaces 15, 73-83.
White, D.J. (1988). Mean, variance, and probabilistic criteria in finite Markov decisionprocesses:
A review. J. Optim. Theory Appl. 56, 1-29.
Whittle, P. (1979). A simple condition for regularity in negative programming. J. Appl. Probab.
16, 305-318.
Whittle, P. (1980a). Stability and characterisation condition in negative programming. J. Appl.
Probab. 17, 635-645.
Whittle, P. (1980b). Negative programming with unbounded costs: A simple condition for
regularity. In: R. Hartley, L.C. Thomas, D.J. White (Eds.), Recent Developments in Markov
Decision Processes. Academic Press, New York, 23-34.
Whittle, P. (1983). Optimization Over Time, Dynamic Programming and Stochastic Control, Vol.
II. J. Wiley and Sons, New York.
Wijngaard, J. (1977). Sensitive optimality in stationary Markov decision chains on a general state
space. In: H.C. Tijms and J. Wessels (Eds.), Markov Decision Theory, Mathematical Centre
Tracts No. 93. Mathematical Centre, Amsterdam, 85--94.
Yosida, K. (1968). Functional Analysis. Springer-Verlag, New York.

Markov Decision Processes

Uploaded by

Copyright:

Available Formats

Markov Decision Processes

Uploaded by

Document Information

Original Description:

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

Markov Decision Processes

Uploaded by

Copyright:

Available Formats

D.P. Heyman, M.J. Sobel, Eds., Handbooks in OR & MS, Vol.

Markov Decision Processes

This chapter presents theory, applications and computational methods for

transition probability functions. The optimality criteria considered include

This chapter is broad in scope. In Section 2, Markov decision processes are

2.1. Problem definition and notation

r,(s, a) = ~ r,(s, a, j)pt(jls, a)

The collection of objects (T, St, As,t, pt(j[s,a), rt(s,a)) is a Markov

the set of allowable actions. A deterministic decision rule can be regarded as a

2.2. Stationary Markov decision problems

ra(s ) = ~ r(s, a)P{d(s) = a}

pe(jls) = ~ p(jls, a)P{d(s) : a}.

It is convenient to use vector space notation for analysis and presentation of

P(s, j) = PalPal2"'" Pan(s, j) := P ( X , ~, : j Ix7 = s). (2.2)

3.1. A two state M a r k o v decision process

decision maker receives an immediate reward of 5 units and at the next

al, 1 //~{s,.s; ..... ~ a2,1

Asl,t={a~,l,al,2}, As2,t={a2.1} ~ t627'.

r,(si,al.l)--5 , r,(s~,al,2)=lO , r,(s2, a z , l ) = - l . , t62T.

p,(jls~, a1.1) = 0.5, j~Sl,S2~

3.2. A stochastic inventory control problem

g+c(u) if u > 0 , (3.1)

units of product are demanded in a month and the inventory u exceeds

F(u) = ~ f ( j ) p j + f(u)P{O,>~ u ) . (3.2)

This problem is formulated as a Markov decision process as follows.

States (the amount of inventory on hand at the start of a month):

St={O, 1,2,...,M}, t=l,2,...,N+l.

Actions (the amount of additional stock to order in m o n t h t):

Expected rewards (expected revenue less ordering and holding costs):

G(s, a) = F(s + a) - O(a) - h(s + a) , t=1,2,... ,N;

(the value of terminal inventory);

Transition probabilities (see explanation below):

A brief explanation of the derivation of the transition probabilities follows.

The inventory is constrained to be 3 or fewer units and the manager wishes

below where an x corresponds to an infeasible action. Transition probabilities

Equation (4.1b) shows t h a t / 4 , can be defined inductively in terms of kt,_~.

the process up to decision epoch t. Clearly, requiring policies to depend on

v[v(s ) = sup v~(s), s ~ S 1 . (4.4)

By the definition of the supremum, such a policy is guaranteed to exist. The

4.2. Policy evaluation

The expectation in (4.5) is with respect to the process determined by policy

The Finite Horizon Policy Evaluation Algorithm.

UN+I(hN+I) = rN+I(SN#I) for all h N + 1 = ( h N , a N , SN+I) ~ HN+ 1 .

2. Substitute t - 1 for t and compute u[(h,) for each h, ~ H, by

uT(h,) = r,(st, dt(ht)) + ~ P,(jlst, d,(h,))u,\~(h,, dr(h,), j)

This inductive scheme reduces the problem of computing expected rewards

4.3. The optimality equation and the principle of optimality

u*(h,) = sup u~(ht). (4.7)

for t = 1 , . . . , N and h t U= t t t. When t = N + 1, the boundary condition uN+ 1 =

(4.6) by replacing supremum over all actions by the action corresponding to a

Theorem 4.1. S u p p o s e u t is a solution o f (4.8) f o r t = 1 . . . . , N a n d UN+1 =

Theorem 4.2. S u p p o s e u*, t = 1 , . . . , N , are solutions o f the optimality equa-

r,(s,, d * ( h t ) ) + ~ p,+~(jist, d*(ht))u,*~(h,, d*(h,), j)

= max IG(s~, a ) + ~ p t ( j [ s , , a)ut*+l(h ~, a, j ) } . (4.9)

(b) For each t = l , 2 , . . . , N + l ,

U,7r* (ht)=u*(h,), h, E H t, (4.11)

Equation (4.9) is often expressed as

d*(h,)=argmax{r,(st, a)+ ~, pt(jlst, a)ut*Fl(ht, a, j)}.

Theorem 4.3. Let e > 0 be arbitrary and suppose u*, t = 1 . . . . . N + 1, are

r,(s,, d * ( h t ) ) + ~ p,+~(jist, d(ht))u,~(h,, d*(h,), j)

d(h,)=argmax{r,(st, a)+ ~, pt(jlst, a)utFl(ht, a, j)}.

v~(s) = v(s) for all s E S .