ijcai-tutorial
60 Pages
English
Downloading requires you to have access to the YouScribe library
Learn all about the services we offer

Description

R5HFHQW.DPEKD5HFHQWZ$L*HG*UDGYDSL5HFHQWWLDQQS6XEEDPRSS$,,GYDQULDO$,3,-&.DPEKD9LODOD.DPEK6XEEQQ3R$,UD.DPEK6XEEDQ6XEEPZDJWL$$8$9L5HFHQWHGULDOIL7XWRIL7XW8$,$,-J5HFHQW $GYDQFHV LQ $, 3ODQQLQJ$ 8QLILHG 9LHZ&$,6XEEDUDR.DPEKDPSDWL$UL]RQD 6WDWH 8QLYHUVLW\rao@asu.eduhttp://rakaposhi.eas.asu.edu/ijcai99 tutorialQF HV VL QL QJ $ QL HZPlanning is hot...26% of the papers in AAAI 99. 20% of papers in IJCAI 99. New People. Conferences. Workshops. Competitions.Inter planetary explorations. Why the increased interest?Significant scale-up in the Significant strides in ourlast 4 5 years understanding– Before we could – Rich connections betweensynthesize about 5 6 planning and CSP(SAT)action plans in minutes OR (ILP)– Now, we can synthesize » Vanishing separation100 action plans in between planning &minutes SchedulingFurther scale up with New ideas for heuristic» –domain specific control of plannerscontrol – Wide array of approachesfor customizing plannerswith domain-specificknowledgeQF HV VL QL QJ $ QL HZLH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GYLH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GYU5HFHQWHPH5HFHQWOD$X**'LVMGYD8VH*JL&WLDUVSUHPQ$Q,SQW$,P3VS,/ODHPH.DPEKDL]DWLQQH.DPEKFRIUDOD6XEEDV6XEEHUVJI$FWLYH8?ZLRQ9LXVWHGPILFILO8*$LHG9LJQQQSZRP6XEE+6XEEDXULVWLFVOD&UDWLYH3IRLVWHG ...

Subjects

Informations

Published by
Reads 9
Language English

ULDO
.DPEK
OD
5HFHQW
WL
$
,-
*
3
*
P
GYD
$
L
7XW
WL
9L
5HFHQW
6XEE
S
$,
P
Q
S
D
$
$
,
5HFHQW
Q
7XWR
$,
$,
3
HG
.DPEKD
QQ
OD
Z
.DPEK
6XEED
QQ
UD
R
R
UD
.DPEKD
6XEED
S
6XEE
S
Z
,
J
L
$
GYD
8
5HFHQW
9L
ULDO
HG
R
IL

IL

8
,-&
$
D
J
5HFHQW $GYDQFHV LQ $, 3ODQQLQJ
$ 8QLILHG 9LHZ
&$,
6XEEDUDR.DPEKDPSDWL
$UL]RQD 6WDWH 8QLYHUVLW\
rao@asu.edu
http://rakaposhi.eas.asu.edu/ijcai99 tutorial
QF HV VL QL QJ $ QL HZ
Planning is hot...
26% of the papers in AAAI 99. 20% of papers in IJCAI 99.
New People. Conferences. Workshops. Competitions.
Inter planetary explorations. Why the increased interest?
Significant scale-up in the Significant strides in our
last 4 5 years understanding
– Before we could – Rich connections between
synthesize about 5 6 planning and CSP(SAT)
action plans in minutes OR (ILP)
– Now, we can synthesize » Vanishing separation
100 action plans in between planning &
minutes Scheduling
Further scale up with New ideas for heuristic» –
domain specific control of planners
control – Wide array of approaches
for customizing planners
with domain-specific
knowledge
QF HV VL QL QJ $ QL HZ
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GYI
V
S
5HFHQW
M
$
W
*
*
*
5H
GYD
ODQ
*
WLYH
L
6$7
WL
XV
D
LVWHG
S
ODQ
P
HUV
$
OD
,
H[W
Q
X
$,
&
3
LO
S
²
OD
L]DWLR
.DPEKD
+H
QQ
QF
.DPEK
L
R
S
UD
U
6XEED
²
6XEE
Q
J
FWLY
$
V
8
PL]LQJ
Z
XVWR
9L
R
HG
XQ
IL
O
IL
'LUHF
8
S
$

HG
,
9L
Q
J
W
QQ
P
Z
²
6XEE
XULVWLFV
6XEED
'LVMX
OD
WLY
UD
H
3
LQ
R
Q
.DPEK
W
.DPEKD
DV
$,
²
S
8V
P
HPH
S
I
D
GLV
Q
XQ
WL
H
L
ODQ
,
6
$
OXW
*
PL]LQ

UDF
GYD
LRQ
&O
I
O
GLV
DVVLFD
GLV
DVVLFDO
Q
$
WLYH
S
D
O
ª

S
&R
OD
P
5HFHQW

QQ
6$7
5HFHQW
&
LQ
S
GV
HPH
Q
Q
ZRUO
OD
UREO
S
UREOH
L]DWL
ZRU
RP
P
UV
²
+
0RGHOLQJ
H
0RGHOLQJ
XULVWLFV
O
'LVM
DVVLFDO
&
DVVLFD
F
URY
WLYH
O
UH
FO
I
J
I
FR
LVWHG
Q
HPH
J
DVV
FRU
OD
U
S
UHFWQHVV
Q
UHFWQHVV
HU
*
V
5
5
5H
8VH
Q
ILQ
R
HUV
ILQ
Q
Q
Q
RU
GLV
HPH
I
Q
M
I
X
W
Q
RU
FWLYH
I
S
SRU
S
DQ
Q
6XS
²
6X
6R
QLQ
J
J
3
Q
LRQ
J
R
*
W
Q
UDFW
FXVWRPL]DWL
XVW
R
Q
UPD
U
UPDO
UR
O
P
)UDPHZ
M
)UDPHZR
P
FXVWRPL]DW
M
UN
F
N
F
*
&
&
S
G
O
RQ
DQ
DWHG
V
M
'LUH
X
*
DWH

QF

F
3
WLY
L
WLYH
S
U
,/
P
HG
$XWR

I

L

H
63
I

LQ
$XW
5HFHQW
Overview
& O DQ QL JSJS HP
3 3UR YL
HIL QHP HQW 3 W3O OD QQL )R RU
&R QM XQ HUH QHP HQW ODQQH QQHUV
XQ HU QHP HQW QQH UV
HIL QHP HQW R WR QV
ROX WLR QH[ WLR QV
FW & RP OHG &63 /3
3OD
&X WR RQ
RP RQ
SSR UW QRQ F OGV
QF HV VL QL QJ $ QL HZ
Planning : The big picture
Synthesizing goal directed behavior
Planning involves
– Action selection; Handling causal dependencies
– Action sequencing and handling resource
allocation (aka SCHEDULING)
Depending on the problem, plans can be
– action sequences
– or “policies” (action trees, state action mappings
etc.)
QF HV VL QL QJ $ QL HZ
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GY
LR
VW HU
LQD
.DPEK
QQ
5HFHQW
WL
$
$
*
*
*
P
GYD
GYD
*
IL
L
9L
WL
6XEE
D
OD
S
3
P
5HFHQW
$
5HFHQW
,
,
Q
IL
$,
8
3
HG
S
J
OD
Z
.DPEKD
6XEED
QQ
UD
.DPEK
R
R
.DPEKD
UD
S
6XEED
S
6XEE
$,
J
L
$
$
8
$
Z
Q
9L
HG
5HFHQW
Planning & (Classical Planning)
(Static)
Environment (Observable)
Goals
perception action
(perfect) (deterministic)
What action next?
I = initial state G = goal state
(prec) O (effects)
i
[ I ] O O O O [ G ]
i j k m
QF HV VL QL QJ $ QL HZ
Why care about classical Planning?
Many domains are approximately classical
– Stabilized environments
It is possible to handle near-classical domains
through replanning and execution monitoring
Classical planning techniques often shed light on
effective ways of handling non classical planning
worlds
– Currently, most of the efficient techniques for handling
non classical scenarios are still based on
ideas/advances in classical planning
Classical planning poses many interesting
computational challenges
QF HV VL QL QJ $ QL HZ
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GYIL
5HFHQW
OD
5HFHQW
WL
$
J
*
3
*
P
GYD
$
L
8
WL
9L
D
6XEE
S
$,
P
Q
S
D
$
$
,
.DPEK
Q
IL
$,
$
3
HG
.DPEKD
QQ
OD
Z
.DPEK
6XEED
QQ
UD
R
R
UD
.DPEKD
6XEED
S
6XEE
S
Z
,
J
L
$
GYD
8
5HFHQW
9L
HG
5HFHQW
The (too) many brands of classical planners
Planning as Theorem Proving Planning as Search
(Green’s planner )
Search in the space of States
(progression, regression, MEA)
(STRIPS, PRODIGY, TOPI)
Search in the space of Plans
(total order, partial order,
Search in the space ofprotections, MTC)
Task networks (reduction (Interplan,SNLP,TOCL,
of non primitive tasks)UCPOP,TWEAK)
(NOAH, NONLIN,
O Plan, SIPE )
Planning as (constraint) Satisfaction
(Graphplan, IPP, STAN, SATPLAN, BLackBOX )
QF HV VL QL QJ $ QL HZ
Advantages of the Unified View
To the extent possible, this tutorial shuns brand names
and reconstructs important ideas underlying those
brand names in a rational fashion
Better understanding of existing planners
– Normalized comparisons between planners
– Evaluation of trade offs provided by various
design choices
Design of novel planning algorithms
– Hybrid planners using multiple refinements
– Explication of the connections between planning,
CSP, SAT and ILP
QF HV VL QL QJ $ QL HZ
8QLI\LQJ
)UDPHZRUN
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GYUD
.DPEK
P
5HFHQW
WL
$
.DPEKD
Þ
GYD
"
P
GYD
5HFHQW
J
R
8
9L
IL
6XEE
HG
S
9L
D
Z
D
$
$
,
5HFHQW
Q
IL
$,
.DPEK
3
HG
QQ
S
OD
Z
6XEE
6XEED
QQ
UD
OD
R
6XEED
.DPEKD
3
S
$,
S
Q
L
J
L
$
5HFHQW
8
WL
,
$
$
0RGHOLQJ &ODVVLFDO 3ODQQLQJ
$FWLRQV 6WDWHV &RUUHFWQHVV
QF HV VL QL QJ $ QL HZ
Modeling Classical Planning
*States are modeled in terms of (binary)
state variables
At(A,M),At(B,M) -- Complete initial state, partial goal state
¬In(A), ¬In(B)
*Actions are modeled as state
transformation functions
-- Syntax: ADL language (Pednault)
EarthEarth -- Apply(A,S) = (S \ eff(A)) + eff(A)
(If Precond(A) hold in S)
At(A,E), At(B,E),At(R,E)
Effects At(R,M), ¬At(R,E)
¬In(o )In(o ) 11 In( x) At( x,M)x
& ¬At(x, E)Unload(o )Load(o ) 11 Fly()Prec.
At(o ,l ), At(R,l) In(o )1 1 1 1 At(R,E)
QF HV VL QL QJ $ QL HZ
Appolo 13
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GY
ZLOO
.DPEK
S
5HFHQW
WL
$
GYD
*
Q
*
P
GYD
:
L
PH
WL
RQ
D
QW
S
$,
P
Q
S
D
$
*
,
ZLO
Q
$
$,
WKH
3
DQ
.DPEKD
LFD
OD
*
.DPEK
DWL
QQ
5HFHQW
R
R
UD
.DPEKD
6XEED
S
6XEE
S
Z
,
J
L
$
:
8
H
9L
H
HG
D
IL
VX
IL
*
8
F
$
$
J
5HFHQW
HG
5HFHQW
9L
S
QQ
UHV
OD
UHVH
Z
W
6XEE
DWLR
6XEED
*
3
ª
UD
Some notes on action representation
Action ASTRIPS Assumption : Actions must specify all the
state variables whose values they change... Eff: If P then R
If Q then W
No disjunction allowed in effects
– Conditional effects are NOT disjunctive
» (antecedent refers to the previous state &
Action A1
consequent refers to the next state)
Prec: P, Q
Eff: R, WQuantification is over finite universes
Action A2– essentially syntactic sugaring
Prec: P, ~Q
Eff: R, ~WAll actions can be compiled down to a canonical
representation where preconditions and effects are
Action A3
Prec: ~P, Qpropositional
Eff: ~R, W
– Exponential blow up may occur (e.g removing
Action A4
conditional effects)
Prec: ~P,~Q
Eff: ODV XPH KW HFD QR QLF DO HU OUH HQ RQ
QF HV VL QL QJ $ QL HZ
Checking correctness of a plan:
The State based approaches
Progression Proof: Progress the initial state over the action
sequence, and see if the goals are present in the result
progress progress
At(A,E) At(B,E) In(A)
At(R,E) Load(A) Load(B)At(R,E) At(R,E)
At(B,E) In(A) In(B)
Regression Proof: Regress the goal state over the action
sequence, and see if the initial state subsumes the result
regress regress
At(A,E) At(B,E) In(A)
At(R,E) Load(A) Load(B)At(R,E) In(B)
At(B,E) In(A)
QF HV VL QL QJ $ QL HZ
LH 9 LHG LI 8Q LQ DQ 3O LQDW DP DR DU FH DQ $GY
LH 9 LHG LI 8Q LQ DQ 3ODW DP DR DU LQ FH DQ $GY
VV