Intro to AI STRIPS Planning & Applications in Video-games Lecture6-Part1

Stavros Vassos, University of Athens, Greece stavrosv@di.uoa.gr May 2012

INTRODUCTION TO AI
STRIPS PLANNING
.. and Applications to Video-games!

Course overview
2

 Lecture 1: STRIPS planning, state-space search
 Lecture 2: Planning graphs, domain independent
heuristics
 Lecture 3: Game-inspired competitions for AI research,
AI decision making for non-player characters in games
 Lecture 4: Planning Domain Definition Language (PDDL),
examples with planners and Prolog code
 Lecture 5: Employing STRIPS planning in games:
SimpleFPS, iThinkUnity3D, SmartWorkersRTS
 Lecture 6: Planning beyond STRIPS

STRIPS planning
3

 What we have seen so far

 The STRIPS formalism for specifying planning problems
 Solving planning problems using state-based search

 Progression planning

 Effective heuristics for progression planning (based on
relaxed problems, planning graphs)
 PDDL tools for expressing and solving STRIPS problems

STRIPS planning
4

 What we have seen so far Classical planning

 There is complete knowledge about the initial state
 Actions are deterministic with exactly one outcome

 The solution is a linear plan (a sequence of actions)

STRIPS planning
5

 What we have seen so far Classical planning

 There is complete knowledge about the initial state
 Actions are deterministic with exactly one outcome

 The solution is a linear plan (a sequence of actions)

 Search “off-line”, then execute with “eyes closed”

STRIPS planning
6

Α
Β
Α Β C C

On(Α,Table)
On(Α,Β)
On(Β,Table)
On(Β,C)
On(C,Table)
Clear(Α)
Clear(Β)
Clear(C)

s0 g

STRIPS planning
7

Α Α
Β Β Β
Α Β C Α C C C

On(Α,Table) On(Α,Table) On(Α,Β)
On(Α,Β)
On(Β,Table) On(Β,C) On(Β,C)
On(Β,C)
On(C,Table) On(C,Table) On(C,Table)
Clear(Α) Clear(Α) Clear(Α)
Clear(Β) Clear(Β) Clear(Table)
Clear(C) Clear(Table)

Move(Β,Table,C) Move(Α,Table,Β)
s0 s1 s2 g

STRIPS planning: Search
8

On(Α,Β)
On(Β,C)

s0 s1 s2 g

STRIPS planning: Execute
9

On(Α,Β)
On(Β,C)

s0 s1 s2

10

 blackbox –o sokoban-domain.txt –f sokoban-problem.txt
----------------------------------------------------
Begin plan
1 (push c4-4 c4-3 c4-2 down box1)
2 (push c4-3 c3-3 c2-3 left box2)
3 (move c3-3 c3-2 down)
4 (move c3-2 c2-2 left)
…
28 (move c1-2 c1-3 up)
29 (push c1-3 c2-3 c3-3 right box1)
End plan
----------------------------------------------------

11

 blackbox –o sokoban-domain.txt –f sokoban-problem.txt
----------------------------------------------------
Begin plan
1 (push c4-4 c4-3 c4-2 down box1)
2 (push c4-3 c3-3 c2-3 left box2)
3 (move c3-3 c3-2 down)
…
28 (move c1-2 c1-3 up)
End plan
----------------------------------------------------

Planning beyond STRIPS
12

 What we have not seen so far

13


 Initial state with incomplete information

14


 Open world assumption, e.g., I don’t know anything about
block D, could be sitting anywhere
 Disjunctive information, e.g., On(A,B)  On(B,A)
 Existential information, e.g., I know there is a block on top of
A but I don’t know which one: x On(x,A)

15


 Open world assumption, e.g., I don’t know anything about
block D, could be sitting anywhere
 Disjunctive information, e.g., On(A,B)  On(B,A)
 Existential information, e.g., I know there is a block on top of
A but I don’t know which one: x On(x,A)

 Game-world: I know there is treasure hidden in some chest
but I don’t know which one

16


 Nondeterministic actions with more than one outcome

17


 An action succeeds with a degree of probability, e.g.,
move(x,b,y) action succeeds with a 90% probability
 An action may have more than one outcomes, e.g., moving a
block may lead to moving the intended block or a
neighbouring one

18


 An action succeeds with a degree of probability, e.g.,
move(x,b,y) action succeeds with a 90% probability
 An action may have more than one outcomes, e.g., moving a
block may lead to moving the intended block or a
neighbouring one

 Game-world: Picking a lock may result in the door opening or
the tool breaking

19


 Representation of the duration of actions

20


 Representation of the duration of actions
 How can we say that an action takes more time than another
one?
 How can we say that the goal should be reached within a
time limit?

21


 Exogenous events

22


 What if in the blocks world we decided to push one of the
blocks from time to time and change its location?
 What if in the blocks world there was another gripper that
could move blocks in order to achieve their goal?

23


 What if in the blocks world we decided to push one of the
blocks from time to time and change its location?
 What if in the blocks world there was another gripper that
could move blocks in order to achieve their goal?

 Game-world: the state of the game is altered not only by the
moves of our agent/NPC but also by the human player and
other agents

24


 Sensing actions

25


 Sensing actions
 These actions do not affect the world but instead the
knowledge of the agent about the world is updated
 E.g., sense which is the block that is on top of block A

26


 Sensing actions
 These actions do not affect the world but instead the
knowledge of the agent about the world is updated
 E.g., sense which is the block that is on top of block A

 Game-world: look-inside(chest1) could update the information
that the agent has about what is lying inside the chest

27


A more expressive solution
 Looking for a linear plan is the simplest case (and works well
only in classical planning problems)

28


A more expressive solution
 Looking for a linear plan is the simplest case (and works well
only in classical planning problems)

A solution can be
a tree of nested if-then-else statements, e.g.,
[if open(chest) then … else …]
 a more expressive program that specifies how the agent
should behave, e.g.,
[while open(chest) do … end while]

29

 Let’s see some scenarios that combine such features

30

 Three versions of the Vacuum Cleaner domain

31

 Version 1 of the Vacuum Cleaner domain

 Incomplete information about the initial state
 The cleaning bot does not know its position
 Deterministic actions
 Actions moveLeft, moveRight, clean always succeed with the
intuitive effects
 The bot does not get any other information about the state

32


 Conformant planning
 Find a sequence of actions that achieves the goal in
all possible cases

33


 Conformant planning
 Find a sequence of actions that achieves the goal in
all possible cases
 Plan: [moveLeft, clean, moveRight, clean]

34


 Incomplete information about the initial state
 The cleaning bot does not know its position
 Deterministic actions
 Actions moveLeft, moveRight, clean always succeed with the
intuitive effects
 At run-time the cleaning bot can see which state it is in

35


 Conditional planning
 Find a plan that also uses if-then-else statements, such
that it achieves the goal assuming that conditions can be
evaluated at run-time
 Plan: [ if isRight then clean else moveRight, clean ]

36


 Complete information about the initial state
 The cleaning bot is on the left, there is dirt on the right
 Nondeterministic actions
 Actions moveLeft, moveRight my fail without affecting the state
 At run-time the cleaning bot can see which state it is in

37


 Planning for more expressive plans
 Finda a plan that also uses while statements, such that it
eventually achieves the goal assuming that conditions can
be evaluated at run-time
 Plan: [ while isLeft do moveRight end while, clean ]

38

 We see that the resulting plan need not be a linear
sequence of actions

 How do we search for such plans?
 AIMA Section 12.3: Planning and acting in
nondeterministic domains
 AIMA Section 12.4: Conditional planning

39

sequence of actions


 Let’s
see an interesting extension of STRIPS that aims to
account for some of the problems we identified

40

 Planning with Knowledge and Sensing (PKS)
 [Petrick,
Bacchus 2002]
 http://homepages.inf.ed.ac.uk/rpetrick/software/pks/

 Extension of STRIPS that takes into account that
some information will be available at run-time
 Kf is like the normal STRIPS database but with open world
 Kw holds literals whose truth value will be known at run-time
 Kv holds literals with terms that will be known at run-time
 Kx holds exclusive or information about literals

 Works with conditional plans that take cases

41

sequence of actions


 Are these enough for building a real NPC?

42

 What happens when an exogenous event changes
something in the state while a plan is executed?

43

 MiniGame domain

44

 MiniGame domain
 up
 up
 up
 pickup
 right
 right
 right
 stab

45

 MiniGame domain
 up
 up
 up
 pickup
 right
 right
 right
 stab

46

 MiniGame domain
 up
 up
 up
 pickup
 right
 right
 right
 stab

47

 MiniGame domain

48

 The human player picks up the weapon that was part of
the plan for the NPC
 The player pushes the NPC out of the position it thinks its
located
…

49

 Before executing the next action check that the
preconditions of the actions are satisfied
preconditions of all remaining actions will be satisfied
 Specify some special conditions that should hold at each
step of the plan in order to continue executing it

50

preconditions of the actions are satisfied
preconditions of all remaining actions will be satisfied
 Specify some special conditions that should hold at each
step of the plan in order to continue executing it

 AIMA Section 12.5: Execution monitoring and replanning

51

 The approaches we have seen so far look for a plan
that features simple programming constructs

52

 What if we could also provide the planner with a
“sketch” of how the plan should look like?
 Note that this makes sense only for a particular
application, i.e., it is domain dependant

53

 What if we could also provide the planner with a
“sketch” of how the plan should look like?
 Note that this makes sense only for a particular
application, i.e., it is domain dependant

 In this way we can also specify a behavior for an
agent that works in an “on-line” manner
 First,
find a way to get a weapon. Execute the plan.
 Then, find a way to get to the player. …

54

 MiniGame domain

55

 Golog: High-level agent programming language

 search (
(turn; π x. move(x) )*;
π x. pick-up(x);
?(π x. gun(x) and npc-holding(x));
);
search (
(turn; π x. move(x) )*;
?(npc-at(x) and player-at(y) and adjacent (x,y));
);
shoot-player

56


57


 Based on situation calculus, a first-order logic formalism

 Much
more expressive than STRIPS for specifying a
domain and an initial situation

 Many extensions in the literature, and a few working
systems available, e.g.,
 http://www.cs.toronto.edu/cogrobo/main/systems/index.html

Course overview
58

 Lecture 1: STRIPS planning, state-space search
 Lecture 2: Planning graphs, domain independent
heuristics
 Lecture 3: Game-inspired competitions for AI research,
AI decision making for non-player characters in games
 Lecture 4: Planning Domain Definition Language (PDDL),
examples with planners and Prolog code
 Lecture 5: Employing STRIPS planning in games:
SimpleFPS, iThinkUnity3D, SmartWorkersRTS
 Lecture 6: Planning beyond STRIPS

Bibliography
59

 Material
 Artificial
Intelligence: A Modern Approach 2nd Ed. Stuart Russell,
Peter Norvig. Prentice Hall, 2003 Sections 11.2, 12.3, 12.4, 12.5
 References
A knowledge-based approach to planning with incomplete
information and sensing. Ronald P. A. Petrick, Fahiem Bacchus. In
Proceedings of the International Conference on AI Planning and
Scheduling Systems (AIPS), 2002
 Golog: A Logic Programming Language for Dynamic Domains.
Hector J. Levesque, Raymond Reiter, Yves Lesperance, Fangzhen
Lin, Richard B. Scherl. Logic Programming, Vol. 31, No. 1-3. 1997

Intro to AI STRIPS Planning & Applications in Video-games Lecture6-Part1

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Intro to AI STRIPS Planning & Applications in Video-games Lecture6-Part1 (20)

Recently uploaded (20)

Intro to AI STRIPS Planning & Applications in Video-games Lecture6-Part1