markov decision processes: a survey1/68 markov decision processes: a survey adviser : yeong-sung...
TRANSCRIPT
Markov Decision Processes: A Survey 1/68
Markov Decision Processes: A Survey
Adviser: Yeong-Sung LinGraduate Student: Cheng-Ta Lee
Network Optimization Research Group
March 22, 2004
Markov Decision Processes: A Survey 2/68
Outline
Introduction Markov Theory Markov Decision Processes Conclusion Future Work
Markov Decision Processes: A Survey 3/68
IntroductionDecision Theory
Probability Theory
+ Utility Theory
=
Decision Theory
Describes what an agent should believe based on evidence.
Describes what an agent wants.
Describes what an agent should do.
Markov Decision Processes: A Survey 4/68
Introduction
Markov decision processes (MDPs) theory has developed substantially in the last three decades and become an established topic within many operational research.
Modeling of (infinite) sequence of recurring decision problems (general behavioral strategies)
MDPs defined Objective functions Policies
Markov Decision Processes: A Survey 5/68
Markov Theory
Markov process A mathematical model that us useful in the
study of complex systems. The basic concepts of the Markov process are
those “state” of a system and state “transition”. A graphic example of a Markov process is
presented by a frog in a lily pond. State transition system
Discrete-time processContinuous-time process
Markov Decision Processes: A Survey 6/68
Markov Theory
To study the discrete-time process Suppose that there are N states in the system nu
mbered from 1 to N. If the system is a simple Markov process, then the probability of a transition to state j during the next time interval, given that the system now occupies state i, is a function only of i and j and not of any history of the system before its arrival in i.
In other words, we may specify a set of conditional probability pij.
where
N
jijP
1
1 10 ijp
Markov Decision Processes: A Survey 7/68
The Toymaker Example
First state: the toy is great favor. Second state: the toy is out of favor. Matrix form Transition diagram
6.04.0
5.05.0][ ijpP
Markov Decision Processes: A Survey 8/68
The Toymaker Example
, the probability that the system will occupy state i after n transitions.
If its state at n=0 is known. It follow that
)(ni
N
ii n
1
1)(
,...2,1,01)()1(1
npnnN
iijij
,...2,1,0)()1( nPnn
3
2
)0()2()3(
)0()1()2(
)0()1(
PP
PP
P
,...2,1,0)0()( nPn n
Markov Decision Processes: A Survey 9/68
The Toymaker Example
If the toymaker starts with a successful toy, then and , so that 1)0(1 0)0(2 01)0(
200
111
200
89
6.04.0
5.05.0
20
11
20
9)2()3(
20
11
20
9
6.04.0
5.05.05.05.0)1()2(
5.05.06.04.0
5.05.001)0()1(
P
P
P
Markov Decision Processes: A Survey 10/68
The Toymaker Example
Table 1.1 Successive State Probabilities of Toymaker Starting with a Successful Toy
Table 1.2 Successive State Probabilities of Toymaker Starting without a Successful Toy
)(1 n
)(2 n
n= 0 1 2 3 4 5 …
1 0.5 0.45 0.445 0.4445 0.44445 …
0 0.5 0.55 0.555 0.5555 0.55555 …
)(1 n
)(2 n
n= 0 1 2 3 4 5 …
0 0.4 0.44 0.444 0.4444 0.44444 …
1 0.6 0.56 0.556 0.5556 0.55556 …
Markov Decision Processes: A Survey 11/68
The Toymaker Example
The row vector with components is thus the limit as n approaches infinity of
i)(n
P
N
ii
1
1 212
211
6.05.0
4.05.0
121
9
41
9
52
Markov Decision Processes: A Survey 12/68
z-Transformation
For the study of transient behavior and for theoretical convenience, it is useful to study the Markov process from the point of view of the generating function or, as we shall call it, the z-transform.
Consider a time function f(n) that takes on arbitrary values f(0), f(1), f(2), and so on, at nonnegative, discrete, integrally spaced points of time and that is zero for negative time.
Such a time function is shown in Fig. 2.4
Fig. 2.4 An Arbitrary discrete-time function
Markov Decision Processes: A Survey 13/68
z-Transformation z-transform F(z) such that Table 1.3. z-Transform Pairs
0
)()(n
nznfzF
n z1
1
z1
1
nn
2)1( z
z
2)1( z
z
)(nfn
)( zF
Time Function for n>=0 z-Transform
f(n) F(z)
f1(n)+f2(n) F1(z)+F2(z)
kf(n) (k is a constant) kF(z)
f(n-1) zF(z)
f(n+1) z-1[F(z)-f(0)]
1 (unit step)
n (unit ramp)
Markov Decision Processes: A Survey 14/68
z-Transformation
Consider first the step function
the z-transform is or
For the geometric sequence f(n)=αn,n 0≧ , or
00
,...3,2,1,01)(
n
nnf
32
0
1)()( zzzznfzFn
n
zzF
1
1)(
00
)()()(n
n
n
n zznfzF zzF
1
1)(
Markov Decision Processes: A Survey 15/68
z-Transformation
We shall now use the z-transform to analyze Markov processes.
In this expression I is the identity matrix.
,...2,1,0)()1( nPnn Pzzz )()0()(1
)0())((
)0()()(
zPIz
Pzzz
1))(0()( zPIz
Markov Decision Processes: A Survey 16/68
z-Transformation Let us investigate the toymaker’s problem by z-transformation.
5
3
5
22
1
2
1
P
zz
zzzPI
5
31
5
22
1
2
11
)10
11)(1(
2
11
)10
11)(1(
5
2
)10
11)(1(
2
1
)10
11)(1(
5
31
1
zz
z
zz
z
zz
z
zz
z
zPI
zzzz
zzzz
zPI
10
11
9
4
19
5
10
11
9
4
19
410
11
9
5
19
5
10
11
9
5
19
4
1
9
4
9
49
5
9
5
10
11
1
9
5
9
49
5
9
4
1
11
zzzPI
9
4
9
49
5
9
5
10
1
9
5
9
49
5
9
4
)(n
nH
Let the matrix H(n) be the inverse transform of (I-zP)-1 on an element-by-element basis
1))(0()( zPIz
)()0()( nHn
Markov Decision Processes: A Survey 17/68
z-Transformation
If the toymaker starts in the successful state 1, then π(0)=[1 0] and or ,
If the toymaker starts in the unsuccessful state 2, then π(0)=[0 1] and or ,
We have now obtained analytic forms for the data in Table 1.1 and 1.2.
9
5
9
5
10
1
9
5
9
4)(
n
nn
n
10
1
9
5
9
4)(1
n
n
10
1
9
5
9
5)(2
9
4
9
4
10
1
9
5
9
4)(
n
nn
n
10
1
9
4
9
4)(1
n
n
10
1
9
4
9
5)(2
Markov Decision Processes: A Survey 18/68
Laplace Transformation
We shall extend our previous work to the case in which the process may make transitions at random time intervals.
The Laplace transform of a time function f(t) which is zero for t<0 is defined by
Table 2.4. Laplace Transform Pairs
dtetfsF st
0
)()(
)(tfdt
d
ate as
1
s
1
atte 2)1(
1
a
2
1
s
)(tfe at
Time Function for t>=0 z-Transform
f(t) F(s)
f1(t)+f2(t) F1(n)+F2(n)
kf(t) (k is a constant) kF(s)
sF(s)-f(0)
1 (unit step)
t (unit ramp)
F(s+a)
Markov Decision Processes: A Survey 19/68
Laplace Transformation
dtetfsF st
0
)()(
00
01)(
t
ttf
0
1)(
sdtesF st
0)( tforetf at
0
)(
0
1)(
asdtedteesF tasstat
ji
ijiji
jijj dtatdtatdtt )(]1)[()(
ji
jijj aa
ji
ijijjjj dtatdtatdtt )(]1)[()(
)0()()()(
dtdtattdttji
ijijj
Markov Decision Processes: A Survey 20/68
Laplace Transformation
We shall now use the Laplace transform to analyze Markov processes.
N
iiji att
dt
d
1
)()( Attdt
d)()(
Atet )0()(
For discrete processes,
,...2,1,0)0()( nPn n
Pe A or PA ln
33
22
!3!2A
tA
ttAIe At
Attdt
d)()(
Asss )()0()(
)0())(( AsIs
1))(0()( AsIs
Markov Decision Processes: A Survey 21/68
Laplace Transformation
Recall the toymaker’s initial policy, for which the transition-probability matrix was
5
3
5
22
1
2
1
P
44
55A
44
55
s
sAsI
)9(
5
)9(
4)9(
5
)9(
4
1
ss
s
ss
ssss
s
AsI
99
4
9
5
99
4
9
49
9
5
9
5
99
5
9
4
1
ssss
ssssAsI
9
4
9
49
5
9
5
9
1
9
5
9
49
5
9
411
ssAsI
Markov Decision Processes: A Survey 22/68
Laplace Transformation
Let the matrix H(t) be the inverse transform (sI-A)-1
Then becomes by means of inverse transformation
1))(0()( AsIs
)()0()( tHt
9
4
9
49
5
9
5
9
5
9
49
5
9
4
)( 9tenH
Markov Decision Processes: A Survey 23/68
Laplace Transformation
If the toymaker starts in the successful state 1, then π(0)=[1 0] and or ,
If the toymaker starts in the unsuccessful state 2, then π(0)=[0 1] and or ,
We have now obtained analytic forms for the data in Table 1.1 and 1.2.
9
5
9
5
9
5
9
4)( 9tet
tet 91 9
5
9
4)( tet 9
2 9
5
9
5)(
9
4
9
4
9
5
9
4)( 9tet
tet 91 9
4
9
4)(
tet 92 9
4
9
5)(
Markov Decision Processes: A Survey 24/68
Markov Decision Processes
MDPs applies dynamic programming to the solution of a stochastic decision with a finite number of states.
The transition probabilities between the states are described by a Markov chain.
The reward structure of the process is described by a matrix that represents the revenue (or cost) associated with movement from one state to another.
Both the transition and revenue matrices depend on the decision alternatives available to the decision maker.
The objective of the problem is to determine the optimal policy that maximizes the expected revenue over a finite or infinite number of stages.
Markov Decision Processes: A Survey 25/68
Markov Process with Rewards
Suppose that an N-state Markov process earns rij dollars when it makes a transition from state i to j.
We call rij the “reward” associated with the transition from i to j.
The rewards need not be in dollars, they could be voltage levels, unit of production, or any other physical quantity relevant to the problem.
Let us define vi(n) as the expected total earnings in the next n transitions if the system is now in state i.
Markov Decision Processes: A Survey 26/68
Markov Process with Rewards
Recurrence relation
N
jjijiji nNinrpn
1
,...3,2,1,...,2,1)]1([)(
N
j
N
jjijijiji nNinprpn
1 1
,...3,2,1,...,2,1)]1()(
NirpqN
jijiji ,...,2,1
1
,...3,2,1,...,2,1)1()(1
nNinpqnN
jjijii
v(n)=q+Pv(n-1)
Markov Decision Processes: A Survey 27/68
The Toymaker Example
Table 3.1. Total Expected Reward for Toymaker as a Function of State and Number of Weeks Remaining
73
39R
6.04.0
5.05.0P
3
6q
)(1 n
)(2 n
n= 0 1 2 3 4 5 …
0 6 7.5 8.55 9.555 10.5555 …
0 -3 -2.4 -1.44 -0.444 0.5556 …
,...3,2,1,...,2,1)1()(1
nNinpqnN
jjijii
Markov Decision Processes: A Survey 28/68
Toymaker’s problem: total expected reward in each state as a function of week remaining
Markov Decision Processes: A Survey 29/68
z-Transform Analysis of the Markov Process with Rewards
The z-Transform of the total-value vector v(n) will be called where )(z
0
)()(n
nznvz
,...2,1,0)()1( nnPvqnv
)(1
1)]0()([1 zPq
zvzz
)(1
)0()( zzPqz
zvz
)0(1
)()( vqz
zzzPI
)0()()(1
)( 11 vzPIqzPIz
zz
qzPIz
zz 1)(
1)(
v(0)=0
Markov Decision Processes: A Survey 30/68
z-Transform Analysis of the Markov Process with Rewards
9
4
9
49
5
9
5
10
11
1
9
5
9
49
5
9
4
1
11
zzzPI
9
4
9
49
5
9
5
10
11
9
10
19
10
9
5
9
49
5
9
4
)1(
9
4
9
49
5
9
5
)10
11)(1(
9
5
9
49
5
9
4
)1(1
2
2
1
zzz
z
zz
z
z
zzPI
z
z
1
1
zPI
z
z
9
4
9
49
5
9
5
10
11
9
10
9
5
9
49
5
9
4
)(n
nnF
3
6q
,
4
5
10
11
9
10
1
1)(
n
nnv
n
nnv10
11
9
50)(1
n
nnv10
11
9
40)(2
9
50)(1 nnv
9
40)(2 nnv
The total-value vector v(n) is then F(n)q by inverse transformation of , and, sinceqzPI
z
zz 1)(
1)(
Let the matrix F(n) be the inverse transform of
We see that, as n becomes very large.
Both v1(n) and v2(n) have slope 1 and v1(n)-v2(n)=10.
Markov Decision Processes: A Survey 31/68
Optimization Techniques in General Markov Decision Processes
Value Iteration Exhaustive Enumeration Policy Iteration Linear Programming Lagrangian Relaxation
Markov Decision Processes: A Survey 32/68
Value Iteration
5.05.0][ 1 jp 39][ 1 jr
6.04.0][ 2 jp 73][ 1 jr
5.05.0][ 11 jp 39][ 1
1 jr
2.08.0][ 21 jp 44][ 2
1 jr
6.04.0][ 12 jp 73][ 1
1 jr
3.07.0][ 22 jp 191][ 2
2 jr
Original
Advertising? No Yes
Research? No Yes
Markov Decision Processes: A Survey 33/68
Diagram of States and Alternatives
Markov Decision Processes: A Survey 34/68
The Toymaker’s Problem Solved by Value Iteration
The quantity is the expected reward from a single transition from state i under alternative k. Thus,
The alternatives for the toymaker are presented in Table 3.1.
kip 1
kip 2
kir1
kir 2
kiq
State AlternativeTransition
ProbabilitiesReward
Expected Immediate
Reward
i k
1 (Successful toy)1 (No advertising) 0.5 0.5 9 3 6
2 (Advertising) 0.8 0.2 4 4 4
2 (Unsuccessful toy)1 (No research) 0.4 0.6 3 -7 -3
2 (research) 0.7 0.3 1 -19 -5
kiq
NirpqN
j
kij
kij
ki ,...,2,1
1
Markov Decision Processes: A Survey 35/68
The Toymaker’s Problem Solved by Value Iteration
We call di(n) the “decision” in state i at the nth stage. When di(n) has been specified for all i and all n, a “policy” has been determined. The optimal policy is the one that maximizes total expected return for each i and n.
To analyze this problem. Let us redefine as the total expected return in n stages starting from state i if an optimal policy is followed. It follows that for any n
“Principle of optimality” of dynamic programming: in an optimal sequence of decisions or choices, each subsequence must also be optimal.
,...2,1,0)]([max)1(1
nnrpnN
jj
kij
kij
ki
)(ni
Markov Decision Processes: A Survey 36/68
The Toymaker’s Problem Solved by Value Iteration
NirpqN
j
kij
kij
ki ,...,2,1
1
,...2,1,0)(max)1(1
nnpqnN
jj
kij
ki
ki
)(1 n
)(2 n
)(1 nd
)(2 nd
n= 0 1 2 3 4 …
0 6 8.2 10.22 12.222 …
0 -3 -1.7 0.23 2.223 …
- 1 2 2 2 …
- 1 2 2 2 …
Table 3.6 Toymaker’s Problem Solved by Value Iteration
Markov Decision Processes: A Survey 37/68
The Toymaker’s Problem Solved by Value Iteration
Note that for n=2, 3, and 4, the second alternative in each state is to be preferred. This means that the toymaker is better advised to advertise and to carry on research in spite of the costs of these activities.
For this problem the convergence seems to have taken place at n=2, and the second alternative in each state has been chosen. However, in many problems it is difficult to tell when convergence has been obtained.
Markov Decision Processes: A Survey 38/68
Evaluation of the Value-Iteration Approach
Even though the value-iteration method is not particularly suited to long-duration processes.
Markov Decision Processes: A Survey 39/68
Exhaustive Enumeration
The methods for solving the infinite-stage problem.
The method calls for evaluating all possible stationary policies of the decision problem.
This is equivalent to an exhaustive enumeration process and can be used only if the number of stationary policies is reasonably small.
Markov Decision Processes: A Survey 40/68
Exhaustive Enumeration
Suppose that the decision problem has S stationary policies, and assume that Ps and Rs are the (one-step) transition and revenue matrices associated with the policy, s=1, 2, …, S.
Markov Decision Processes: A Survey 41/68
Exhaustive Enumeration
The steps of the exhaustive enumeration method are as follows. Step 1. Compute vs
i, the expected one-step (one-period) revenue of policy s given state i, i=1, 2, …, m.
Step 2. Compute πsi, the long-run stationary probabilities of the tran
sition matrix Ps associated with policy s. These probabilities, when they exist, are computed from the equations
πs Ps =πs
πs1 +πs
2 +…+πsm =1
where πs =(πs1 , πs
2 , …, πsm ).
Step 3. Determine Es, the expected revenue of policy s per transition step (period), by using the formula
Step 4. The optimal policy s* id determined such that
m
i
si
si
sE1
}{* ss EMaxE
Markov Decision Processes: A Survey 42/68
Exhaustive Enumeration
We illustrate the method by solving the gardener problem for an infinite-period planning horizon.
The gardener problem has a total of eight stationary policies, as the following table shows:
Stationary policy, s Action
1 Do not fertilize at all.
2 Fertilize regardless of the state.
3 Fertilize if in state 1.
4 Fertilize if in state 2.
5 Fertilize if in state 3.
6 Fertilize if in state 1 or 2.
7 Fertilize if in state 1 or 3.
8 Fertilize if in state 2 or 3.
Markov Decision Processes: A Survey 43/68
Exhaustive Enumeration
The matrices Ps and Rs for policies 3 through 8 are derived from those of policies 1 and 2 and are given as
100
5.05.00
3.05.02.01P
100
150
3671R
55.04.005.0
3.06.01.0
1.06.03.02P
236
047
1562R
100
5.05.00
1.06.03.03P
100
150
1563R
100
3.06.01.0
3.05.02.04P
100
047
3674R
55.04.005.0
5.05.00
3.05.02.05P
236
150
3675R
100
3.06.01.0
1.06.03.06P
100
047
1566R
55.04.005.0
5.05.00
1.06.03.07P
236
150
1567R
55.04.005.0
3.06.01.0
3.05.02.08P
236
047
3678R
Markov Decision Processes: A Survey 44/68
Exhaustive Enumeration
Step1: The values of vs
i can thus be computed as given in the following table.
si
si=1 i=2 i=3
1 5.3 3 -1
2 4.7 3.1 0.4
3 4.7 3 -1
4 5.3 3.1 -1
5 5.3 3 0.4
6 4.7 3.1 -1
7 4.7 3 0.4
8 5.3 3.1 0.4
Markov Decision Processes: A Survey 45/68
Exhaustive Enumeration Step 2:
The computations of the stationary probabilities are achieved by using the equations
πs Ps =πs
πs1 +πs
2 +…+πsm =1
As an illustration, consider s=2. The associated equations are
The solution yields
In this case, the expected yearly revenue is
21
23
22
21 05.01.03.0
22
23
22
21 4.06.06.0
23
23
22
21 55.03.01.0
123
22
21
256.2)4.0221.3317.46(59
13
1
222 i
iiE
59
22,
59
31,
59
6 23
22
21
Markov Decision Processes: A Survey 46/68
Exhaustive Enumeration Step 3&4:
The following table summarizes πs and Es for all the stationary policies.
Policy 2 yields the largest expected yearly revenue. The optimum long-range policy calls for applying fertilizer regardless of the system.
s1 s
2 s3 sE
2.256=*sE
S
1 0 0 1 -1
2 6/59 31/59 22/59
3 0 0 1 0.4
4 0 0 1 -1
5 5/154 69/154 80/154 1.724
6 0 0 1 -1
7 5/137 62/167 70/137 1.734
8 12/135 69/135 54/135 2.216
Markov Decision Processes: A Survey 47/68
Policy Iteration
The system is completely ergodic, the limiting state probabilities πi are independent of the starting state, and the gain g of the system is
where qi is the expected immediate return in state i defined by
N
iii qg
1
NirpqN
jijiji ,...,2,1
1
Markov Decision Processes: A Survey 48/68
Policy Iteration
A possible five-state problem.
The alternative thus selected is called the “decision” for that state; it is no longer a function of n. The set of X’s or the set of decisions for all states is called a “policy”.
Markov Decision Processes: A Survey 49/68
Policy Iteration
It is possible to describe the policy by a decision vector d whose elements represent the number of the alternative selected in each state. In this case
An optimal policy is defined as a policy that maximizes the gain, or average return per transition.
3
1
2
2
3
d
Markov Decision Processes: A Survey 50/68
Policy Iteration
In five-state problem diagrammed, there are different policies.
However feasible this may be for 120 policies, it becomes unfeasible for very large problem.
For example, a problem with 50 states and 50 alternatives in each state contains 5050( 10≒ 85) policies.
The policy-iteration method that will be described will find the optimal policy in a small number of iterations.
It is composed of two parts, the value-determination operation and the policy-improvement routine.
12051234
Markov Decision Processes: A Survey 51/68
Policy Iteration The Iteration Cycle
NirpqN
jijiji ,...,2,1
1
Markov Decision Processes: A Survey 52/68
The Toymaker’s Problem
Let us suppose that we have no a priori knowledge about which policy is best.
Then if we set v1=v2=0 and enter the policy-improvement routine.
It will select as an initial policy the one that maximizes expected immediate reward in each state.
For the toymaker, this policy consists of selection of alternative 1 in both state 1 and 2. For this policy
1
1d
6.04.0
5.05.0P
3
6q
73
39R
Markov Decision Processes: A Survey 53/68
The Toymaker’s Problem
We are now ready to begin the value-determination operation that will evaluate our initial policy.
Setting v2=0 and solving these equation, we obtain
We are now ready to enter the policy-improvement routing as shown in Table 3.8
211 5.05.06 vvvg 212 6.04.03 vvvg
1g
101 v
02 v
N
jj
kij
ki vpq
1
State Alternative Test Quantity
i k
112
6+0.5(10)+0.5(0)=11 4+0.8(10)+0.2(0)=12
212
-3+0.4(10)+0.6(0)=1 -5+0.7(10)+0.3(0)=2
Markov Decision Processes: A Survey 54/68
The Toymaker’s Problem
The policy-improvement routine reveals that the second alternative in each state produces a higher value of the test quantity than does the first alternative. For this policy,
We are now ready to the value-determination operation that will evaluate our policy.
With v2=0, the results of the value-determination operation are
The gain of the policy is thus twice that of the original policy, we have found the optimal policy.
For the optimal policy, v1=10, v2=0, so that v1-v2=10. This means that, even when the toymaker is following the optimal policy by using advertising and research.
2
2d
3.07.0
2.08.0P
5
4q
211 2.08.04 vvvg 212 3.07.05 vvvg
2g
101 v
02 v
2
2d
Markov Decision Processes: A Survey 55/68
Linear Programming
The infinite-stage Markov decision problems, can be formulated and solved as linear programs.
We have defined the policy of MDP and can be defined
by . Each state has k decisions, so D can be
characterized by assigning values in the
matrix, , where each row must contain a single
1with the rest of the elements zero. When an element =1, it can be interpreted as calling for decision k when the system is in state i .
Nd
d
d
D
2
1
10 orDik
NKNN
K
K
ddd
ddd
ddd
D
21
12221
11211
ikD
Markov Decision Processes: A Survey 56/68
Linear Programming
When we use linear programming to solve the MDP problem, we will define the formulation as .
The linear programming formulation is best expressed in terms of a variable , which is related to as follows.
Let be the unconditional probability that the system is in state i and decision k is made; that is, .
From the rules of conditional probability, . Furthermore, . So that
N
i
K
k
kiiki qdE
1 1
ikw
ikD
}{ kdecisionandistatePwik
ikw
ikiik dw
K
kiki w
1
K
k ik
ik
i
ikik
w
wwd
1
Markov Decision Processes: A Survey 57/68
Linear Programming
There exist several constraints on1.
2.
3.
ikw
N
ii
1
1 , so that
N
i
K
kikw
1 1
1
.from the results on steady-state probabilities,
N
iijij p
1
, so that NjforpwwN
i
kij
K
kik
K
kik ,,2,1,
1 11
KkandNiwik ,,2,1,,2,1,0
Markov Decision Processes: A Survey 58/68
Linear Programming
The long run expected average revenue per unit time is given by , hence the problem to choose the that , subject to the constrains.
1.
2.
3. This is clearly a linear programming problem that can be
solved by the simplex method. Once the is obtained, the
N
i
K
k
kiik
N
i
K
k
kiiki qwqdE
1 11 1
ikw
N
i
K
kikw
1 1
1
NjforpwwN
i
kij
K
kik
K
kjk ,,2,1,0
1 11
KkandNiwik ,,2,1,,2,1,0
N
i
K
k
kiik qwMaximize
1 1
K
k ik
ikik
w
wd
1
ikw
Markov Decision Processes: A Survey 59/68
Linear Programming
The following is an LP formulation of the gardener problem without discounting:Maximize E=5.3w11+4.7w12+3w21+3.1w22-w31+0.4w32
subject tow11 + w12 - (0.2w11 + 0.3w12 + 0.1w22 + 0.05w32) = 0w21 + w22 - (0.5w11 + 0.6w12 + 0.5w21 + 0.6w22 + 0.4w32) = 0w31 + w32 - (0.3w11 + 0.1w12 + 0.5w21 + 0.3w22 + w31 + 0.55w32) = 0
w11 + w12 + w21 + w22 + w31 + w32 = 1
wik>=0, for all I and k The optimal solution is w11 = w21 = w31 = 0 and w12 = 0.1017, w22 = 0.5254,
and w32 = 0.3729. This result mean that d12=d22=d32=1. Thus, the optimal policy selects alternative k=2 for i=1, 2, and 3. The optimal values of E is 4.7(0.1017)+3.1(0.5254)+0.4(0.3729)=2.256.
Markov Decision Processes: A Survey 60/68
Largrangian Relexation
If the linear programming method can not find the optimal solution with the additional constraints .
we can use Lagrangian relaxation to bind the constraints to the object function, and then solve this new sub problem without the additional constraints added .
By adjusting the multiplier of Lagrangian relaxation, we can get the upper bound and the lower bound of this problem.
We will use the multiplier of Lagrangian relaxation to rearrange the revenue of Markovian decision process, and then do the original Markovian.
Decision Process model to find the optimal policy .
Markov Decision Processes: A Survey 61/68
Comparison
Characteristic
Methods
Calculates simply
large problemOptimal
policy Additional constraints
Value Iteration
Exhaustive Enumeration
Policy Iteration
Linear Programming
Lagrangian Relaxation
Markov Decision Processes: A Survey 62/68
Semi-Markov Decision Processes
So far we have assumed that decisions are taken at each of a sequence of unit time intervals.
We will allow decisions to be taken at varying integral multiples of the unit time interval.
The interval between decisions may be predetermined or random.
Markov Decision Processes: A Survey 63/68
Partially Observable MDPs
MDPs assume complete observable (can always t
ell what state you’re in).
We can’t always be certain of the current state.
POMDPs are more difficult to solve than MDPs
Most real-world problems are POMDPs
Markov Decision Processes: A Survey 64/68
Applications on MDPs
Capacity Expansion Decision Analysis Network Control Queueing System Control
Markov Decision Processes: A Survey 65/68
Conclusion
MDPs provide and elegant formal framework for sequential decision making.
We present a powerful tool for formulating models and finding the optimal policies.
Five algorithm were presented Value Iteration Exhaustive Enumeration (optimal policy) Policy Iteration (optimal policy) Linear Programming (optimal policy) Lagrangian Relaxation (optimal policy)
Markov Decision Processes: A Survey 66/68
Future Work
Sensor Networks Maximize system lifetime of sensor networks Maximize cover the area of sensor networks Minimize response time of sensor networks
Markov Decision Processes: A Survey 67/68
References1. Hamdy A. Taha, “Operations Research: an Introduction,” third edition, 1982.2. Hillier and Lieberman,”Introduction to Operations Research,” fourth edition, H
olden-Day, Inc, 1986.3. R. K. Ahuja, T. L. Magnanti, and J. B. Orlin, “Network Flows,” Prentice-Hall, 1
993.4. Leslie Pack Kaelbling, “Techniques in Artificial Intelligence: Markov Decision
Processes,” MIT OpenCourseWare, Fall 2002.5. Ronald A. Howard, “Dynamic Programming and Markov Processes,” Wiley, N
ew York, 1970.6. D. J. White, “Markov Decision Processes,” Wiley, 1993.7. Dean L. Isaacson and Richard W. Madsen, “Markov Chains Theory and Appli
cations,” Wiley, 19768. M. H. A. Davis “Markov Models and Optimization,” Chapman & Hall, 1993.9. Martin L. Puterman, “Markov Decision Processes: Discrete Stochastic Dynam
ic Programming,” Wiley, New York, 1994.10. Hsu-Kuan Hung, Adviser: Yeong-Sung Lin ,“Optimization of GPRS Time Sl
ot Allocation”, June, 2001.11. Hui-Ting Chuang, Adviser: Yeong-Sung Lin ,“Optimization of GPRS Time Sl
ot Allocation Considering Call Blocking Probability Constraints”, June, 2002.
Markov Decision Processes: A Survey 68/68
References
12. 高孔廉,「作業研究 --管理決策之數量方法」,三民總經銷,民國 74年四版。13. 李朝賢,「作業研究概論」,弘業文化實業股份有限公司出版,民國 66年 8
月。14. 楊超然,「作業研究」,三民書局出版,民國 66年 9月初版。15. 葉若春,「作業研究」,中興管理顧問公司出版,民國 86年 8月五版。16. 薄喬萍,「作業研究決策分析」,復文書局發行,民國 78年 6月初版。17. 葉若春,「線性規劃理論與應用」,民國 73年 9月增定十版。18. Leonard Kleinrock, “Queueing Systems Volume I: Threory,” Wiley, New York,
1975.19. Chiu, Hsien-Ming, “Lagrangian Relaxation,” Tamkang University, Fall 2003.20. L. Cheng, E. Subrahmanian, A. W. Westerberg, “Design and planning under
uncertainty: issues on problem formulation and solution”, Computers and Chemical Engineering, 27, 2003, pp.781-801.
21. Regis Sabbadin, “Possibilistic Markov Decision Processes”, Engineering Application of Artificial Intelligence, 14, 2001, pp.287-300.
22. K. Karen Yin, Hu Liu, Neil E. Johnson, “Markovian Inventory Policy with Application to the Paper Industry”, Computers and Chemical Engineering, 26, 2002, pp.1399-1413.