.

R
els
ng
E
Artificial
ls. RIntelligence
R
ge s.
UnitE 4-1 Planning
n
E ngel

. R
els . R
ng
2020-2021 Odd BE CSE VII ls
semester
e
E ng
E
. R Engels. R
els . R
ng els
E ng
E
. R
els . R
ng els
E ng
E 1
 Unit 4 – Planning

• Planning and Learning: s. R

el
– Planning with State Space Search:Eng
• Partial Order Planning R
ls.
• Planning Graphs ge . R
E n els
• Examples ng
R E
• Blended/Self learning
ls.
ge . R
E n
– Forms of Learning els
ng
R Learning E
• .
Inductive
ls
ge . R
•
En
Explanation Based Learning
els
• Statistical Learning ng
R E
ls. • Learning With Complete Data
e
g . R
E n els
g
En 2
 Background

• Focus . R
els
– The focus here is deterministic planning n g
E
• Environment is fully observable
. R
• Results of actions is deterministic ls
ge . R
– Relaxing the above requiresEndealing with uncertainty els
ng
R contingency, exploration
• Problem types: sensor-less, E
ls.
• Planning ‘communities’ ge in AI . R
E n els
– Logic-based: Reasoning About Actions & Change ng
R E
s. representations: Classical AI Planning
– Less formal
l
ge . R
En
– Uncertainty (UAI): Graphical Modelsesuch ls as
ng
R E
• Markov Decision Processes (MDP), Partially Observable MDPs, etc.
ls.
g e . R
En els
ng
E 3
 Actions, events, and change
. R
l
• Planning requires a representation of time
e s
ng
– E
to express & reason about sequences of actions
. Rthe world
– to express the effects of actions son
l
ge . R
• Propositional Logic E n els
ng
– does not offer a representation
R for time E
.
ls needs to be repeated for eachR step
ge
– Each action description .
E n els
• Situation Calculus ng
R E
. FOL
– Is basedlson
g e . R
– EachEntime step is a ‘situation’ els
ng
E about actions & change
–R Allows to represent plans and reason
ls.
g e . R
En els
ng
E 4
 ‘Famous’ Problems
• Frame problem . R
els
– Representing all things that stay the same n g from one situation to the next
E
– Inferential and representational R
ls.
– What happens to a gold bar placed ge on a slide? . R
E n els
• Qualification problem ng
R E
s.
– Defining the circumstances
l under which an action is guaranteed to work
ge . R
– Example: what Eifnthe gold is slippery or nailed down, els etc.
ng
• Ramification R problem E
ls.
ge
– Proliferation of implicit consequences of. Ractions as actions may have secondary
E n els
consequences ng
R E
ls.– Examples: How about the dust on the gold?
g e . R
En els
ng
E 5
 Planning Languages
. R
els
• Languages must represent.. Eng
– States . R
els R
– Goals ng ls.
E ge
E n
– Actions . R
els R
ng ls.
• Languages must
E be ge
E n
– Expressive
. R for ease of representation
els R
g s.
En
– Flexible for manipulation by l
ng algorithms
e
R E
ls.
e R
E ng els.
ng
E 6
 State Representation
. R
• A state is represented with a conjunction e s
l of positive literals
n g
– fluents that are ground, functionless Eatoms
.R
– a fluent is a condition that canlschange over time
ge . R
• Using En els
ng
– Logical Propositions:. RPoor  Unknown E
els R
g
– FOL literals: At(Plane1,SFO)  At(Plane2,JFK) ls.
En ge
• FOL literalsRmust be ground & function-free E n
ls.
ge
– Not allowed: R
At(x,y) [non-ground]or sAt(Father(Fred),Sydney)[Function],
.
E n el
poor (negation, false fluent) ng
R E
.
•elsClosed World Assumption R
ng ls.
E – What are not stated aregeassumed false
En
 Action Representation
. R
els
• A set of ground (variable-free) actions
ng can be represented by a
E
single action schema R
ls.
– The schema is a lifted representation
ge . R
E n els
– it lifts the level of reasoning from propositional logic ng to a restricted
R E
subset of first-order
ls. logic
ge . R
En els
ng
R E
ls.
ge . R
En els
ng
R E
ls.
ge . R
En els
ng
E
 Action Representation
. R
els
• Action Schema ng At(AI,CJB),Plane(AI),
E Airport(CJB), Airport(MAA)
– Action name
. R
– Preconditions
els Fly(AI,CJB,MAA) R
– Effects ng ls.
E g e
• Example E n
At(AI,MAA),  At(AI,CJB)
. R
Action(Fly(p,from,to),els R
ng ls.
E e
PRECOND: At(p,from)  Plane(p)  Airport(from) gAirport(to)
E n
R
EFFECT:s.At(p,from)  At(p,to))
l
e Effects are split into ADD list and R DELETE list
• Sometimes, g ls.
En ge
RESULT(s, a) = [s − DEL(a)] ∪ n
ADD(a)
E
. R
els s- state, a-action, s’-new state R
g s.
En DEL(a) – actions available el
g at s, ADD(a) – actions available at s’
E n
 Applying an Action

• Find a substitution list  for the variables. R

els
– of all the precondition literals n g
E
– with (a subset of) the literals in the current state description
. R
• Apply the substitution to theepropositions ls in the effect list R
n g ls.
• Add the result to the current E state description to generate ge the new
E n
state . R
els R
g ls.
• Example: En ge
– Current state: At(P1,JFK)  At(P2,SFO)  Plane(P1)n
E  Plane(P2)  Airport(JFK) 
. R
ls
Airport(SFO)
e R
g ls .
En
– It satisfies e
the precondition with ={p/P1,from/JFK,
g to/SFO)
E n
.–R Thus the action Fly(P1,JFK,SFO) is applicable
els – R
ng The new current state is: .
At(P1,SFO)
ls  At(P2,SFO)  Plane(P1)  Plane(P2) 
E Airport(JFK)  Airport(SFO) g e
En
 Goal & Action Representation
. R
s
• Goal is a partially specified state ngel
E
• A proposition satisfies a goalR if it contains all the atoms of the
ls.
goal and possibly others..ge . R
n E ls e
– Example: Rich  Famous  Miserable satisfies the goal g
n Rich  Famous
R E
ls.
g e . R
En els
ng
R E
ls.
ge . R
En els
ng
R E
ls.
ge . R
En els
ng
E
 Languages for Planning Problems
. R
• STRIPS els
n g
E
– Stanford Research Institute Problem Solver
. R
– Historically important els R
n g ls.
• ADL E ge
E n
R
. Languages
– Action Description els R
ng ls.
• PDDL E ge
E n
R
. Domain Definition Language
– Planning
els R
ng ls .
E
– Describes the four things we need e
g to define a search problem:
E n
R
.– the initial state, the actions that are available in a state, the result of
els R
g s .
En applying an action, and
gel the goal test
E n
 Example: Air Cargo
. R
• Initial state, Goal State
els
• Actions: Load, Unload, Fly ng
E
. R
els R
ng sA. PDDL description of
l an air cargo
E ge
n
E transportation
. R
els R planning problem:
ng ls. Loading and unloading
E e
E ng cargo and
. R flying it from place to
els R place
E ng els.
ng
R E
ls.
ge . R
E n els
ng Solution? Start with [Load(C1, P1, SFO)….]
E
 Example: Air Cargo
. R
els
ng A PDDL description of
E an air cargo
transportation
. R planning problem:
els R and unloading
ng .
Loading
lscargo and
E e
g flying it from place to
E n
. R
els R
place

ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
 Example: Spare Tire Problem
A PDDL description of

. R changing a flat tire

els problem:

ng Goal is to have a good

E spare tire properly
mounted onto the car’s
. R axle,
els R where the initial state has
ng ls. (a) a flat tire on the axle
E ge and (b) a good spare tire
E n
. R in the trunk

els R
ng ls. Four actions:
E ge 1. Removing spare from
E n trunk,
. R 2. Removing flat tire
els R from axle,
ng ls. 3. Putting spare on axle,
E ge and
E n
. R 4. Leaving car

els R
unattended overnight

ng ls. (which will result in

E ge both spare and flat
Solution?
E n tires stolen)
 Example: Spare Tire
. R Problem
els
ng
E
A PDDL description of
. R
els R
changing a flat tire

E ng els. problem:

ng Goal is to have a good

R E spare tire properly
ls. mounted onto the car’s
ge . R
E n els axle,

ng where the initial state

R E has (a) a flat tire on the
ls. axle and (b) a good
ge . R
E n els spare tire in the trunk

ng
R
Solution E
ls.
ge . R
E n els
ng
E
 Example: Blocks World
. R
• Initial state, Goal
els
• Actions: n g
E
• Move(b,x,y) (b-block, x-from, y – to) R
ls.
• MoveToTable(b,x) ge . R
En pre-condition and effects els
• Describe init, goal and actions(with
ng
R E
• Then give the solutionsplan
l .
g e . R
En els
ng
R E
ls.
g e of the blocks-world problem: . R
ls
A PDDL description
Goal is to E n one or more stacks of blocks, e
build
ng
R E
ls.can fit directly on top of another, or kept on the table
where the blocks can be stacked, but only one block

g e . R
En el s initial state goal state
g
En
 Example: Blocks. RWorld
els
• Initial state, Goal n g
E
• Actions: Move(b,x,y), MoveToTable(b,x)
. R
els R A PDDL description of the
n g ls. blocks-world problem:
E e
E ng Goal is to build one or

. R more stacks of blocks,

els R
g ls. where the blocks can be
En ge stacked, but only one
E n block can fit directly on
. R top of another, or kept on
els R the table
g ls .
En ge
E n
. R
e ls R
g s .
En gel Solution?
E n
 Example: Blocks. RWorld
els
• Initial state, Goal n g
E A PDDL description of the
• Actions: Move(b,x,y), MoveToTable(b,x)
. R blocks-world problem:

els Goal is to build one or

R stacks of blocks,
g s .
more
En el
g where the blocks can be
n
E stacked, but only one
. R
els R block can fit directly on
g ls.
En ge top of another, or kept on

E n the table

. R
e ls R
g ls. Notice how the solution plan finding
En ge becomes extremely straight forward

E n with the appropriate representation

. R
ls
e One solution is the sequence: R
g s.
En [MoveToTable (C,A), Move(B, Table, C),geMove(A,
l Table,B)]
En
 State-Space Search
• We want to see if the planning process can . R be solved using Search
els
Algorithms n g
E
– Problem must be formulated suitableRfor search
ls.
– Define initial, goal, transition model,
ge state space tree and search strategy . R
E n els
• Search the space of states (first chapters) ng
R E
– Initial state, goal test,ls. step cost, etc.
ge . R
– Actions are theEntransitions between state els
ng
• Actions are R
invertible (why?) E
ls.
– Move ge
forward from the initial state: . R
Forward State-Space Search or
E n e ls
Progression Planning ng
R E
.
ls – Move backward from goal state: Backward State-Space Search or Regression
g e . R
E n Planning els
g
En
 Forward and Backward State Space Search
. R
els Two approaches to
ng searching for a plan.
E (a) Forward (progression)

. R search, through the space

els of states, starting in the

R
ng ls. initial state and using the
E ge problem’s actions to
E n search forward for a
. R member of the set of goal
els R states
ng ls.
E ge (b) Backward (regression)
E n
. R search through sets of

els R
relevant states, starting at

ng ls. the set of states

E ge representing the goal and
E n using the inverse of the
. R actions to search backward
els R for the initial state
ng ls.
E ge
E n
 State-Space Search (3)
.
• Remember that the language has no functions R symbols
els
ng
• Thus number of states is finite E
• And we can use any complete . Rsearch algorithm (e.g., A*)
els g R .
– [Completeness – does the els if one exists]
Enalgorithm always find a solution
ng
– We need an admissible R heuristic E
ls.
. R
– The solution is nagepath, a sequence of actions: total-order planning
E e ls
g
• Problem: SpaceR and time complexity En
ls.
e
– onlyngpositive preconditions andls. R
E ge
–R only one literal effect En
ls.
g•e With the above identified . R we can use any appropriate search
En els
ng
algorithm for planning.
E
 STRIPS & State-Space
. R Search
els
g and
• STRIPS : Easy to focus on ‘relevant’ propositions
n
– Work backward from goal (using Effects)E
. R
– Work forward from initial state (using PRECONDITIONS)
els R
– Facilitating bidirectional search g s.
En gel
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
 Relevant and Consistent Actions
. R
els
• An action is relevant n g
E
– In Progression planning whenR its preconditions match a subset of
ls.
the current state ge . R
En els
– In Regression planning, when its effects match a subset ng of the
R E
ls
current goal state .
ge . R
n
• The purpose of Eapplying an action is to ‘achieves els a desired literal’
ng
R E
. careful that the action does not undo a desired literal as a
– Shouldlsbe
ge . R
side
Eneffect els
ng
• A. Rconsistent action is an action that E does not undo a desired literal
els R
g s.
En ge l
E n
 Backward State-Space Search
. R
• Given els
ng
E
– A goal G description
. R
– An action A that is relevantelsand consistent R
n g ls.
• Generate a predecessor state where E ge
E n
. R
ls
– Positive effects (literals)
e of A in G are deleted R
ng ls.
– PreconditionEliterals of A are added unless ngethey already appear
E
R any variables in A’s effects to match literals in G
ls.
– Substituting
g e . R
En
– Substituting els
any variables in A’sgpreconditions to match substitutions
E n
. R in A’s effects
els R
g .
En • Repeat until predecessor gel description matches initial state
s
E n
 Heuristic to Speed up Search
. R
• We can use A*, but we need an admissible e s
l heuristic
n g
– Must not overestimate cost from current s toEgoal g
– Such as the straight line distance (hSLD) we
. R used earlier )
els R
• Such an admissible heuristicngcan be derived with ls.
E e
g heuristic for the
– by defining a relaxed problem whose solution can be used as admissible
E n
. R
given problem
els R
g ls.
• Think of a search Enproblem as a graph ge
– where the nodes are states and the edges are actionsEn
. R
• The problem e s
l is to find a path connecting theRinitial state to a goal state
n g ls.
• ThereEare two ways we can relax this problem ge to make it easier:
n
E it strictly easier to find a path, or
–. R by adding more edges to the graph, making
els– by grouping multiple nodes together, R forming an abstraction of the state space that has
g s .
En ge
fewer states, and thus is easierl to search
E n
26
 Heuristic to Speed up Search
. R
Approach els
ng
E
1. Divide-and-conquer: sub-goalR independence assumption
ls.
e
– Problem relaxation isngdone .R
through one of thelsfollowing
E ge
– By removing R En
ls.
e
2. … all preconditions
g . R
En els
ng
3. … all preconditions
. R and negative effects
E
els R
g .
4. …Ennegative effects only: Empty-Delete-List
els
g
E n
. R
els R
ng ls.
E ge
E n
27
 1. Subgoal Independence Assumption
. R
ls
• A key idea in defining heuristics is decomposition:
e
ng
E
The cost of solving a conjunction of subgoals is the sum of the
. R
costs of solving each subgoal
els independently R
g s.
• Optimistic En el
ng E
– Assume that there R
.are no subplans that interact negatively when put-
els R
g ls.
together En ge
n
E achieved by an action in
– Example:Rone action in a subplan delete goal
ls.
e subplan
another
g . R
En els
• Pessimistic (not admissible) Eng
. R
els – Redundant actions in subplansR can be replaced by a single action in
g s.
En merged plan gel
E n
28
 2. Problem Relaxation: Removing Preconditions
. R
els
g
• Remove preconditions from action descriptions
n
E
– All actions are applicable
R
ls.in every state, and
• Every action becomes applicable
e R
ng
– Every literal in the goal is Eachievable ls.
in one step ge
• Any single goal fluentRcan be achieved in one step E n
ls.
ge
• if there is an applicable R
action—if not, the problem is .impossible
E n els
• Number of steps to achieve the conjunction of literals ng in the goal is equal to the
R E
.
number oflsunsatisfied literals
ge . R
• Alert En e ls
ng (1)
E
– Some actions may achieve several literals
R
ls.– Some action may remove the effect of another action (2)
ge . R
En – For many problems an accurate e s
l heuristic is obtained by considering (1) and ignoring (2)
ng
E 29
 3. Remove Preconditions & Negative Effects
. R
• Considers only positive interactions among e s
l actions to achieve multiple
n g
subgoals E
. R
• The minimum number of actions els required is R
g s.
En positive effects that satisfy the
– Sum of the union of the actions’ gel goal
E n
. R
e ls R
g ls.
En ge
E n
. R
e ls R
g ls .
En ge
E n
. R
els R
g s.
En gel
E n
30
 4. Removing Negative Effects (Only)
. R
els
n g
E
• Remove all negative effects of actions (no action may destroy the effects
of another) . R
els R
g s.
• Known as the Empty-Delete-List En heuristic gel
E n
• Requires running a simple
. R planning algorithm
els R
g ls.
• Quick & effectiveEn ge
• Usable in progression E n
. R or regression planning
els R
g ls .
En ge
E n
. R
els R
g s.
En ge l
E n
31
 . R
els
ng
E
. R
els
ng
E
. R
els
ng
E
. R
PARTIAL
ngel ORDERING
s GRAPHS – PLANNING
E
GRAPHS
R
ls.
ge
En
32
 Regression
. R
els
• Regressing a goal, G, thru anng action, A
E
• Yields the weakest precondition
R G’
. ls
– Such that: if G’ is true ge before A is executed . R
En els
– G is guaranteed to be true afterwards ng
R E
ls.
g e . R
En els
g

precond
E n
G’ A G

effect
. R
e ls R
ng ls .
E
Represents a set of world g e
E n Represents a set of world
. R states states
els R
g s.
En gel
E n
© Daniel S. Weld 33
 Planning Graphs
. R
s
• All of the heuristics can suffer from einaccuracies
l
ng
E
• A special data structure called a Planning Graph can be
. R
– used to give better heuristicelsestimates R
g s.
– applied to any of the search En techniques or another algorithmgel
E n
GRAPHPLAN . R
els R
• A planning problem g .
asks if we can reach eals goal state from the
En g
initial state E n
. R
ls
• A treeEnof ge all possible actions from . R
ls initial state to successor states,
ge
and their successors, and so on n
E to answer
. R
ls
e – Can we reach state G from Rstate S0
g s.
En el
g size and hence impractical
n
– This tree is of exponential
E 34
 Planning Graphs
. R
• A planning graph = polynomial size e s
lapproximation to the
ng
exhaustively indexed tree E
– can be constructed quicklyels. R R
g s.
– May not answer definitively En whether G is reachable gfromel S0, but can
E n
estimate the number . R of steps
els R
g .
• The estimate is En always correct when it greports els the goal is not
E n
reachable ls. R
ge . R
• And it Ennever overestimates the e s
lnumber of steps
ng
E
. R it is an admissible heuristic
• lsSo
g e . R
En els
ng
E 35
 Planning Graphs - 2
. R
• Planning graph is a directed graphgeorganized ls into levels
E n
– First a level S0 for the initial state, consisting of nodes representing
each fluent (a condition that . R change over time) that holds in S
can
els R 0
n g ls.
– then a level A0, consisting E of nodes for each ground gaction e that might
E n
be applicable in Sls0. R
ge . R
– then alternatingEn levels Si followed by Ai gels
E n
– until we . R
reach a termination condition
els R
g .
• PlanningEn graphs work only forgpropositional els planning problems
E n
.–R The problems with no variables
els R
g s.
En gel
E n
36
 Planning Graphs - 3
. R
• Roughly
els
– ng I
Si contains all the literals that could hold atEtime
R
• depending on the actions executed at preceding
.
time steps
ls R in Si
– If it is possible that either P or ¬Pgecould hold, then both will be represented
n ls.
E
– Ai contains all the actions that could have their preconditions satisfied e
g at time I
E n
• Why roughly? . R
els R
g ls.
En only a restricted subset of thegepossible
– Planning graph records negative interactions
among actions; E n
. R
– therefore,
e s
l a literal might show up at level Sj when
R actually it could not be true until a
g if at all
laterEnlevel, ls.
ge
– It is guaranteed thought that A literalEwilln never show up too late
. R
•elsDespite the possible error, the Rlevel j at which a literal first appears is a good
ng ls.
E estimate of how difficult itgeis to achieve the literal from the initial state
En 37
 Planning Graphs – Example problem
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
38
 Planning Graphs – Example
. R
• The planning graph for the “have cake and eat cake too” problem
e ls
• Up to level S2 ng
• Rectangles indicate actions E
• small squares indicate persistence actions R
l
• Straight lines indicate preconditions and effectss.
g e . R
• Mutex links are shown as curved gray lines
E n els
• (not all mutex links are shown)
ng
R the persistence actions for those literals
• If two literals are mutex at Si, then E
ls .
will be mutex at Ai e
g link . R
• need not draw thatnmutex
E e ls
g
R En
ls.
ge . R
E n els
ng
R E
ls.
ge . R
E n els
ng
E 39
 Planning Graphs – Example
R. its effects at
ls
• Each action at level Ai is connected to its preconditions at Si and
e
Si+1 n g
• So a literal appears because an action caused it AND a E literal can persist if no
action negates it
• Known as persistence action or no-op ls.
R
ge . R
• For every literal C, add to the problem anpersistence
E
action with precondition C
els
and effect C
ng
R
• Level A0 in figure shows one “real” action, Eat (Cake) E
ls. actions drawn as small square boxes
e
• along with two persistence
g . R
E n els
g
R En
ls.
ge . R
E n els
ng
R E
ls.
ge . R
E n els
ng
E 40
 Planning Graphs – Example
•
R
Level A0 contains all the actions that could occur in state S0 ls.
ge from occurring together
•
•
Also records conflicts between actions that would prevent
E n them
Grey links indicate mutual exclusion (or mutex) links
• For example, Eat (Cake) is mutually exclusive withRthe persistence of either Have(Cake) or ￢ Eaten(Cake)
• Also Have(Cake) and Eaten(Cake) are mutex ls .
g estates . R
• n
S1 represents a belief state: a set of possible
E els
•
n
Continue alternating between state level Si and action level Ai until two consecutiveglevels are identical
• Graph has leveled off R E
ls. choosing among actions
•
e
Planning graph does not require
g . R
E n
• which would entail combinatorial search
els
• ng
It just records the impossibility of certain choices using mutex links
E
. R
els R
E ng els.
ng
R E
ls.
ge . R
E n els
ng
E 41
 Mutex Relation
. R if any of the following three
e ls
• Mutex relation holds between two actions at a given level
conditions holds: ng
• Inconsistent effects:
E
– One action negates an effect of the other . R
e ls have inconsistent effects because theyRdisagree on the
ng
– Ex: Eat (Cake) and the persistence of Have(Cake)
ls.
effect Have(Cake) E ge
• Interference: R En
.
ls is the negation of a precondition of the other
e
– one of the effects of one action
g . R
En with the persistence of Have(Cake) by negating
– Ex: Eat (Cake) interferes e ls its precondition
n g
• Competing needs:
R E
s .
– One of the lpreconditions of one action is mutually exclusive with a precondition of the other
ge . R
En
– Ex: Bake(Cake) and Eat (Cake) are mutex because they
els compete on the value of the Have(Cake) precondition
• Inconsistent support ng
R E
ls.– A mutex relation holds between two literals at the same level if one is the negation of the other
g . R achieve the two literals is mutually exclusive
e – or if each possible pair of actions that could
En els
ng
E 42
 Planning Graph complexity
. R
e ls
ngthe planning problem
• A planning graph is polynomial in the size Eof
• For a planning problem with l literalsRand a actions
.
elsl2 mutex links, and
– each Si has no more than l nodes and
g . R
En (including the no-ops), (a + l)2 mutex
– each Ai has no more than a + l nodes e s
llinks, and
ng
R
2(al + l) precondition and effect links E
ls.
• Thus, an entire graph gewith n levels has a size of O(n(a +. R
l) 2
)
E n els
• The time to build the graph has the same complexity ng
R E
ls.
g e . R
En e ls
ng
R E
ls.
ge . R
En els
ng
E 43
 The GRAPHPLAN algorithm
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
e ls R
g .
algorithm repeatedly adds a level toelasplanning graph with EXPAND-GRAPH
En
• The GRAPHPLAN
g GRAPHPLAN calls EXTRACT-SOLUTION to search for a
• Once all the goals show up as nonmutex in the graph,E n
R
. that solves the problem
lsplan
R again, terminating with failure when there is no reason to go on
e• If that fails, it expands another level and tries
g s.
En g el
E n
44
 Planning Graphs – Spare tire example
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
45
 Planning Graphs – Spare tire example
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
46
 Planning Graphs – Spare tire example
. R
• GRAPHPLAN initializes the planning graph to a one-level e s
l (S0) graph representing the initial state
n g the problem description’s initial state are
E
– The positive fluents and the relevant negative fluents from
shown
. R (such as Tire(Spare)) and the irrelevant negative literals
ls
– Not shown are the unchanging positive literals
e R
ng in S0, so we need not call EXTRACT-SOLUTION
• The goal At(Spare, Axle) is not present ls.
E ge
– There is no solution yet
R En
.
• EXPAND-GRAPH adds into Als0 the three actions whose preconditions exist at S 0
ge . R
n
– all the actions exceptEPutOn(Spare, Axle) els
ng
– along with persistence actions for all the literals in S
R 0 E
• The effects of ls . actions are added at level S
the
ge 1
. R
En
• EXPAND-GRAPH then looks for mutex relations els adds them to the graph
and
n g
R
• At(Spare, Axle) is still not present in S E
ls . 1

ge – so again EXTRACT-SOLUTION is not called . R

En• EXPAND-GRAPH is called again, gadding s
el A1 and S2 resulting in the planning graph shown
En 47
 Planning Graphs – Spare tire example – Mutex examples
. R
• Inconsistent effects: els
ng
E
– Remove(Spare, Trunk ) is mutex with LeaveOvernight because one has the effect At(Spare,
Ground) and the other has its negation
. R
• Interference: els R
– Remove(Flat, Axle) is mutex withEn
g ls.
ge
LeaveOvernight because one has the precondition At(Flat,
Axle) and the other has its negation as an effect. E n
. R
• Competing needs: els R
g .
ls one has At(Flat, Axle) as a
– PutOn(Spare, Axle)En is mutex with Remove(Flat, Axle) because e
ng
precondition and the other has its negation
R E
• Inconsistent ls.
support:
ge . R
En Axle) is mutex with At(Flat, Axle) gineSls2
– At(Spare,
– Rbecause the only way of achieving At(Spare,En Axle) is by PutOn(Spare, Axle)
ls.
ge – and that is mutex with the persistence
. R action that is the only way of achieving At(Flat, Axle)
En• Thus, the mutex relations detect e ls
ng the immediate conflict
E two objects in the same place at the same time
– that arises from trying to put 48
 Planning Graphs – Spare tire example
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
ng ls.
E ge
E n
. R
e ls R
g .
En from the goal are present in S , and nonegeoflsthem is mutex with any other
• All the literals 2

E
• Hence a solution might exist, and EXTRACT-SOLUTIONn will try to find it
• Can. Rformulate EXTRACT-SOLUTION as a Boolean constraint satisfaction problem (CSP)
ls as a as a backward search problem, whereR
•geOr
n ls. to a level in the planning graph and a set of unsatisfied goals
E • each state in the search contains a
g e
pointer
En 49
 Planning Graphs – Spare tire example
. R
els
ng
E
. R
els R
ng ls.
E ge
E n
. R
els R
• Start at S ng
with the goalEAt(Spare, Axle) els.
• The only choice for achieving the goal set is PutOn(Spare, Axle) Eng
2

s
• Can only take to. aRsearch state at S with goals At(Spare, Ground) and ￢ At(Flat, Axle)
e l Ground) can be achieved only by Remove(Spare,
1
R Trunk ), and
g
• At(Spare, .
• ￢ At(Flat, els
En Axle) can be achieved only by eithergRemove(Flat , Axle) or LeaveOvernight
E n
• But LeaveOvernight is mutex with Remove(Spare, Trunk )
R
. • So the only solution is to choose Remove(Spare, Trunk ) and Remove(Flat, Axle)
ls R At(Spare, Trunk ) and At(Flat, Axle)
e• Brings us to a search state at S with the goals
g s.
En • Both of these are present in the state, e lso we have a solution
0

ng
• The actions Remove(Spare, TrunkE ) and Remove(Flat, Axle) in level A , followed by PutOn(Spare, Axle) in A
0 1 50
 Planning Graphs – Summary
. R
• Planning is a search problem els
n g
E
– Possible to apply search heuristics to improve the process
. R
• One approach is to construct els a “graph” of the searchR space
n g ls.
– and use this to guide searchE ge
E n
. R
ls
• This leads to thegeconcept of a planning graph R
ls.
En ge
E n
. R
els R
g ls .
En ge
E n
. R
els R
g s .
En gel
E n
51
 Unit 4 – Planning - Summary

• Planning and Learning: s. R

el
– Planning with State Space Search:Eng
• Partial Order Planning R
ls.
• Planning Graphs ge . R
E n els
• Examples ng
R E
• Blended/Self learning
ls.
ge . R
E n
– Forms of Learning els
ng
R Learning E
• .
Inductive
ls
ge . R
•
En
Explanation Based Learning
els
• Statistical Learning ng
R E
ls. • Learning With Complete Data
e
g . R
E n els
g
En 52
 References
. R
els
• AI n g
E
– Artificial Intelligence – ElaineR Rich, Kevin Knight, B. Nair 3rd Edition
ls.
• AIMA ge . R
En els
ng
– Artificial IntelligenceR - A Modern Approach 3rd Edition E - RUSSELL &
ls.
NORVIG ge . R
En els
ng
R E
ls.
ge . R
En els
ng
R E
ls.
ge . R
En e ls
ng
E 53