reinforcement learning partially observable markov decision processes (pomdp)
DESCRIPTION
Reinforcement Learning Partially Observable Markov Decision Processes (POMDP). 主講人：虞台文 大同大學資工所 智慧型多媒體研究室. Content. Introduction Value iteration for MDP Belief States & Infinite-State MDP Value Function of POMDP The PWLC Property of Value Function. - PowerPoint PPT PresentationTRANSCRIPT
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
主講人：虞台文
大同大學資工所智慧型多媒體研究室
Content IntroductionValue iteration for MDPBelief States & Infinite-State MDPValue Function of POMDPThe PWLC Property of Value
Function
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
Introduction
大同大學資工所智慧型多媒體研究室
Definition MDP A Markov decision process is a tuple
S a finite set of states of the world A a finite set of actions T: SA (S) state-transition function
R: SA R the reward function
, , ,S A T R
1( , , ) ( | , )t t tT s a s P s s s s a a
Complete Observability
Solution procedures for MDPs give values or policies for each state.
Use of these solutions requires that the agent is able to detect the state it is currently in with complete reliability.
Therefore, it is called CO-MDP (completely observable)
Partial Observability Instead of directly measuring the current
state, the agent makes an observation to get a hint about what state it is in.
How to get hint (guess the state)?– To do an action and take an observation.– The observation can be probabilistic, i.e., it
provides hint only.– The ‘state’ will be defined in probability
sense.
Observation Model
: ( )O S A
a finite set of observations the agent can experience of its world.
1 1( , , ) ( | , )t t tO s a o P o o s s a a
The probability of getting observation o given that the agent took action a and landed in state s’.
Definition POMDP
, , , , ,S A T R O
, , ,S A T R describes an MDP.
: ( )O S A is the observation function.
A POMDP is a tuple
How to find optimal policy in such an environment?
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
Value Iteration for MDP
大同大學資工所智慧型多媒體研究室
Acting Optimality
Finite-Horizon Model
Infinite-Horizon Discounted Model
0
maximize k
tt
E r
Maximize the expected
total reward of the next k steps.
0
maximize tt
t
E r
Maximize the expected
discounted total reward.
0 1
Are there any difference on the nature of their optimal policies?
Stationary vs. Non-Stationary Policies
Finite-Horizon Model
Infinite-Horizon Discounted Model
The optimal policy is dependent on the number of time steps remained.
The optimal policy is independent on the number of time steps remained.
Use non-stationary policy
Use stationary policy : S A
:t S A
Stationary vs. Non-Stationary Policies
Finite-Horizon Model
Infinite-Horizon Discounted Model
The optimal policy is dependent on the number of time steps remained.
The optimal policy is independent on the number of time steps remained.
Use non-stationary policy
Use stationary policy : S A
:t S A The remained time steps.
Value Functions
Finite-Horizon Model
Infinite-Horizon Discounted Model
Non-stationary policy
Stationary policy
, , 1( ,( ( )) ( , ( ), )) ( )tt ts S
tR s s T s s sV s V s
( , ( )) ( , ( ) ),) ()(s S
V s V sR s s T s s s
Optimal PoliciesFinite-Horizon Model
Infinite-Horizon Discounted Model
Non-stationary policy
Stationary policy
, , 1( ,( ( )) ( , ( ), )) ( )tt ts S
tR s s T s s sV s V s
( , ( )) ( , ( ) ),) ()(s S
V s V sR s s T s s s
* *1arg max ( , ) ( , ,( ( )))t t
as S
R s a T s as V ss
* *arg max (( ) (, ( , ) )) ,a
s S
R s a T s a ss V s
*1 arg m () )( ax ,
aRs s a *
1 arg m () )( ax ,a
Rs s a
Optimal PoliciesFinite-Horizon Model
Infinite-Horizon Discounted Model
Non-stationary policy
Stationary policy
* *1max ( , ) ( , , )( ) ( )t t
as S
V s V sR s a T s a s
* *max (( , ) ( , , )) ( )s S
V s V sR s a T s a s
* *1arg max ( , ) ( , ,( ( )))t t
as S
R s a T s as V ss
* *arg max (( ) (, ( , ) )) ,a
s S
R s a T s a ss V s
*1 arg m () )( ax ,
aRs s a *
1 arg m () )( ax ,a
Rs s a
Optimal PoliciesFinite-Horizon Model Non-stationary policy
* *1max ( , ) ( , , )( ) ( )t t
as S
V s V sR s a T s a s
* *
1arg max ( , ) ( , ,( ( )))t ta
s S
R s a T s as V ss
*1 arg m () )( ax ,
aRs s a *
1 arg m () )( ax ,a
Rs s a
How about t ?
How about Vt(s) Vt1(s) s?
How about t if Vt(s) Vt1(s) s?
To find an optimal policy, do we need to pay infinite time?
Value IterationThe MDP has finite number of states.
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
Belief States & Infinite-State MDP
大同大學資工所智慧型多媒體研究室
Agent Agent
POMDP Framework
World (MDP)
SESE
observationaction
bbelief state
SE: state estimator
Belief States
1 2( ) , , ( ), ( ), ( ) 0Ti ib s b s S b ss b
( ) 1s S
b s
There are uncountably infinite number of belief states.
State Space
1 2( ) , , ( ), ( ), ( ) 0Ti ib s b s S b ss b
( ) 1s S
b s
There are uncountably infinite number of belief states.
0 1
2-state POMDP
1( )b s 0 1
13-state POMDP
State Estimation
1 2( ) , , ( ), ( ), ( ) 0Ti ib s b s S b ss b
( ) 1s S
b s
State estimation:
Given bt, at and ot+1, bt+1=?
There are uncountably infinite number of belief states.
State Estimation
1( ) ( | , , )t tb s P s o a b
1 2( ), ( )( ),tt tTb s b sb
1 1 21 1( ),( )( ), Ttt tb s b s b
(
( | , , (
|
| ,
)
)
,
)t
t
tP s
P o
P o s
a
a a
b
b
b
( | , , )
( | , )
( | ( | ,, ) )t ts S
t
P s s a P
P
P o s a s
o a
a
b
b
b
( | , )( | , )
( | ,
( )
)ts S
t
P s s
P
a
o a
a bo sP s
b
( , , )( , , )
,
(
( |
)
)ts S
t
O T s a s bs a o
P a
s
o
b Normalization Factor
State Estimation
1( ) ( | , , )t tb s P s o a b
1 2( ), ( )( ),tt tTb s b sb
1 1 21 1( ),( )( ), Ttt tb s b s b
(
( | , , (
|
| ,
)
)
,
)t
t
tP s
P o
P o s
a
a a
b
b
b
( | , , )
( | , )
( | ( | ,, ) )t ts S
t
P s s a P
P
P o s a s
o a
a
b
b
b
( | , )( | , )
( | ,
( )
)ts S
t
P s s
P
a
o a
a bo sP s
b
( , , )( , , )
,
(
( |
)
)ts S
t
O T s a s bs a o
P a
s
o
b Normalization Factor
1 ( , , )t tSE a o b b1 ( , , )t tSE a o b b
,1 ( | , )
a o tt
tP o a T b
bb
,1 ( | , )
a o tt
tP o a T b
bb
Remember these.Remember these.
State Estimation
1( ) ( | , , )t tb s P s o a b
1 2( ), ( )( ),tt tTb s b sb
1 1 21 1( ),( )( ), Ttt tb s b s b
(
( | , , (
|
| ,
)
)
,
)t
t
tP s
P o
P o s
a
a a
b
b
b
( | , , )
( | , )
( | ( | ,, ) )t ts S
t
P s s a P
P
P o s a s
o a
a
b
b
b
( | , )( | , )
( | ,
( )
)ts S
t
P s s
P
a
o a
a bo sP s
b
( , , )( , , )
,
(
( |
)
)ts S
t
O T s a s bs a o
P a
s
o
b Normalization Factor
1 ( , , )t tSE a o b b1 ( , , )t tSE a o b b
( | , ) ( , , ) ( , , ) ( )t ts S s SP o a O s a o T s a s b s
b( | , ) ( , , ) ( , , ) ( )t ts S s S
P o a O s a o T s a s b s
b
It is linearw.r.t bt
State Transition Function1 ( , , )t tSE a o b b1 ( , , )t tSE a o b b
ba
b’
( , , ) ( | , )a P a b b b b
( , , )SE a ob b( , , )SE a ob b
( | , , ) ( | , )o
P a o P o a
b b b
( , , )
( | , )o
SE a o
P o a
b b
b
It is linearw.r.t bt
State Transition Function
( , , ) ( | , )a P a b b b b
( , , )SE a ob b( , , )SE a ob b
( | , , ) ( | , )o
P a o P o a
b b b
( , , )
( | , )o
SE a o
P o a
b b
b
It is linearw.r.t bt
Suppose that ( , , ) ( , , ) i jSE a o SE a o i j b b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
POMDP = Infinite-State MDP
A POMDP is an MDP with tuple B a set of Belief states A the finite set of actions (the same as
the original MDP) : BA (B) state-transition function
: BA R the reward function1( , , ) ( | , )t t ta P a a b b b b b b
What is the reward function?
, , , B A
Reward Function
( , ) ( ) ( , )s S
R sa b s a
b
The reward function ofthe original MDP
Good news: It is Linear.
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
Value Function of POMDP
大同大學資工所智慧型多媒體研究室
Consider a 2-state POMDP:
Value Function over Belief Space
b0 1
V(b) How to obtain the value function in belief space?
Can we use the table-based method?
Finding Optimal Policy
POMDP = Infinite-State MDPThe general method of MDP:
– To determine the value function and, then followed by policy improvement.
Value functions– State value function– Action value function
Review Value Iteration
Based on finite-horizon value function.
It finds on each iteration.*t
What is*1 ?
The and *1 *
1V
( , ) ( ) ( , )s S
R sa b s a
bImmediateReward
*1( ) (( ) , )
s S
a R sbQ as
b
**
1 1( ) arg m (ax )
a
aQV bb
The and *1 *
1V*1
( ) (( ) , )s S
a R sbQ as
b
**
1 1( ) arg m (ax )
a
aQV bb
Consider a 2-state POMDP with two actions (a1, a2) and three observations (o1, o2 , o3).
b0 1
*1
aQ
a1
a2*
1V
b0 1
a1
a2
Horizon-1 Policy Trees
Consider a 2-state POMDP with two actions (a1, a2) and three observations (o1, o2 , o3).
*1V
b0 1
a1
a2
a2 a1
P1
*1
Horizon-1 Policy Trees
Consider a 2-state POMDP with two actions (a1, a2) and three observations (o1, o2 , o3).
*1V
b0 1
It is piecewise linear and convex.(PWLC)
a2 a1
P1
*1
s1
s2
(0,0)(1,0)
(1,0)
The and *1 *
1V*1
( ) (( ) , )s S
a R sbQ as
b
**
1 1( ) arg m (ax )
a
aQV bb
How about 3-state POMDP and more?
It is PWLC.
What is the policy?
The and *1 *
1V*1
( ) (( ) , )s S
a R sbQ as
b
**
1 1( ) arg m (ax )
a
aQV bb
How about 3-state POMDP and more?
What is the policy?
The PWLC A Piecewise Linear function consists of linear,
or hyperplane segments
– Linear function:
– kth linear segment:
– the -vector:
– each segment could be represented as
0 0 11i Ni
i Nx x x x
0
N
ii
ki x
0 1[ , ,..., ]k k k kN α
( )k tα
The PWLC A Piecewise Linear function consists of linear,
or hyperplane segments
– Linear function:
– kth linear segment:
– the -vector:
– each segment could be represented as
0 0 11i Ni
i Nx x x x
0
N
ii
ki x
0 1[ , ,..., ]k k k kN α
( )k tα
( ) max is PWLC.Tk
kf x α x( ) max is PWLC.T
kk
f x α x
The and *t *
tV
**
1( ) ( , ) ( , , ) ( )att
Q a a V
b
b b b b b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
Immediatereward
*1 ( ,( , ) ( | , ) ( )),t
o
a P o a oa SEV
bb b
Value of observation o for doing action a
on the current stat b. Prob. of observation o
for doing action aon the current stat b.
*, ,11 ( )a o
tV b
The and *t *
tV
**
1( ) ( , ) ( , , ) ( )att
Q a a V
b
b b b b b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
Immediatereward
*1 ( ,( , ) ( | , ) ( )),t
o
a P o a oa SEV
bb b
Value of observation o for doing action a
on the current stat b. Prob. of observation o
for doing action aon the current stat b.
*, ,11 ( )a o
tV b
PWLC
PWLC?
Yes, it is.But, I will defer the proof.Yes, it is.But, I will defer the proof.
The and *2 *
2V
**
12( ) ( , ) ( , , ) ( )aQ a a V
b
b bb b b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
( | , ) ( , , )( , , )
0
P o a SE a oa
otherwise
b b bb b
*, ,1( , ) ( )a o
o
a V
b b
*1( , ) ( ( ,| , , ))) (
o
a P o a V SE a o
b b b
** *, ,
2 12( ) arg max ( ) arg max ( , ) ( )a a o
a ao
V Q a V
b b b b
The and *2 *
2V
**
12( ) ( , ) ( | , ) ( )a
o
Q a P o a V
b bb b
0 1
a1
*1V
0 1
a1
a2
Compute 1*2
aQ
1( , )a bb
b’o1
o3 o2
( , , )SE a o bb ( , , )SE a o bb
( , , )SE a o bb ( , , )SE a o bb
**
12( ) ( , ) ( | , ) ( )a
o
Q a P o a V
b bb b
The and *2 *
2V
0 1
a1
*1V
0 1
a1
a2
Compute 1*2
aQ
1( , )a bb
b’o1
o3 o2
What action will you take if the observation is oi after a1 is taken? What action will you take if the observation is oi after a1 is taken?
The and *2 *
2V
**
12( ) ( , ) ( | , ) ( )a
o
Q a P o a V
b bb b
Consider individual observation (o) after action (a) is taken.
*, , *1 1( ) ( | , ) ( )a oV P o a V b b b
*1( | , ,( )() , )SE a oP o a V bb
Define
**, ,
12( ) ( , ) ( )a a o
o
Q a V
bb b
The and *2 *
2V
0 1
a1( , )a b
0 1
a1a2*, ,
1 ( )a oV b
Transformed value function
0 1
a1a2 *
1 ( )V b
**, ,
12( ) ( , ) ( )a a o
o
Q a V
bb b
The and *2 *
2V*
*, ,12
( ) ( , ) ( )a a o
o
Q a V
bb b
0 1
a1
( , )a b
1 1,*,1 ( )a oV b
0 1
1 3,*,1 ( )a oV b
0 1
1 2,*,1 ( )a oV b
0 1
The and *2 *
2V*
*, ,12
( ) ( , ) ( )a a o
o
Q a V
bb b
0 1
a1
( , )a b
1 1,*,1 ( )a oV b
0 1
1 3,*,1 ( )a oV b
0 1
1 2,*,1 ( )a oV b
0 1
1*2
aQ
o1o2o3
Horizon-2 Tree for Action 1
12 2( , , )aa a 1 1 2( , , )a a a 2 21( , ),aa a
1*2
aQ
o1o2o3
a1
o1 o2o3
P1 P1 P1
*1
Horizon-2 Tree for Action 1
12 2( , , )aa a 1 1 2( , , )a a a 2 21( , ),aa a
1*2
aQ
o1o2o3
12 2( , , )aa a 1 1 2( , ),a a a
2*2
aQ
o1o2o3
The and *2 *
2V
1*2
aQ2*2
aQ
The and *2 *
2V
1*2
aQ2*2
aQ
a1 a2 a1 a2
*2V
Horizon-2 Policy Tree
a1 a2 a1 a2
*2V
o1 o2 o3
P1 P1 P1
*1
P2
*2
Can you figure out How to determine the value function for horizon 3 from the above discussion?
The and *3
b
1 1*, ,2 ( )a oV b
a11 2*, ,
2 ( )a oV b1 3*, ,
2 ( )a oV b
a2
2 1*, ,2 ( )a oV b
2 2*, ,2 ( )a oV b
2 3*, ,2 ( )a oV b
*3V
1*3( )aQ b
2*3
( )aQ b
*3 ( )V b
a1 a2 a1 a2
*2V
The and *3 *
3V
1o
2o
3o
1o
2o
3o
1*3( )aQ b 2
*3( )aQ b
The and *3 *
3V
1o
2o
3o
1o
2o
3o
1*3( )aQ b 2
*3( )aQ b
*3V *
3V
How for *t *
tVand ?
*3V
Horizon-3 Policy Tree
o1 o2 o3
P1 P1 P1
*1
P2
*2
P3
o1 o2 o3
P2P2
*3
Reinforcement Learning
Partially ObservableMarkov Decision Processes
(POMDP)
The PWLC Property of Value Function
大同大學資工所智慧型多媒體研究室
Value Function for POMDP
**
1( ) ( , ) ( , , ) ( )att
Q a a V
b
b b b b b
* *1( ) max ( , ) ( | , ) ( ,( , ))t t
ao
SE a oV a P o a V
b bb b
*1 ( ,( , ) ( | , ) ( )),t
o
a P o a oa SEV
bb b
*1 ( ) max ( , ) max ( ) ( , )i iia a
V a b s R s a b b
Value Function for POMDP
*1 ( ) max ( , ) max ( ) ( , )i iia a
V a b s R s a b b
Let 1 2( , ), ( , ),T
a R s a R s ar
*1 ( ) max T
aa
V b r b
Value Function for POMDP
Let 1 2( , ), ( , ),T
a R s a R s ar
*1 ( ) max T
aa
V b r b
Let ,1 kk aα r *1 ,1( ) max T
kk
V b α b
Theorem
*( )tV b is PWLC.
Proof*( )tV b is PWLC.
By induction
We already know*
1 ( )V b is true.
Assume*
1( )tV b is also true.
We then show *( )tV b must be true.
Proof
* *1( ) max ( , ) ( | , ) ( ,( , ))t t
ao
SE a oV a P o a V
b bb b
From the assumption, we have
*1 , 1( ) m( , , x) a T
t k tk
SE aV o b α b
b
, 1 ( , , )max Tk t
kSE a o bα
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
Proof
, 1arg ma ( , ,( , , )x) Tk t
ko aa SE o b bαLet
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
*, 1( , , ) ( , , )( ) max ( , ) ( | , ) a o
Tt t
ao
SV o a oa EP a
b bb b b α
,( ,| , )
,(
) a oSE a oP o a
T b
bb
,( ,| , )
,(
) a oSE a oP o a
T b
bb
, 1( , ,, )max ( , ) Tt a o
aa
ooa
b bα Tb
Proof
, 1arg ma ( , ,( , , )x) Tk t
ko aa SE o b bαLet
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
*, 1 (( ) max ( , ) ( | , ) m )a , ,x T
t k ta k
o
V a SEo aP oa
b b α bb
*, 1( , , ) ( , , )( ) max ( , ) ( | , ) a o
Tt t
ao
SV o a oa EP a
b bb b b α
, 1( , ,, )max ( , ) Tt a o
aa
ooa
bb bα T
, 1, ) ,( ,max T Ta t aa
ao
oo
b T br α
Proof
( , , ,)*
, 1( ) max T Tt a t
ao
aa ooV
b bb r Tα
Let ( , , ), , 1 ,ii ia o a oT T
k t a to
b Tα r α
*,( ) max T
t k tk
V b α b
*( )tV b is PWLC.