Classical Planning - Final - Notes

Learning Objectives
At the end of the class you should be able to:

explain the model of deterministic planning
represent a problem using the STRIPs representation of
actions.
AI System 1 / 11
State-space Search
deterministic or stochastic dynamics

fully observable or partially observable
explicit states or features or individuals and relations
static or finite stage or indefinite stage or infinite stage
goals or complex preferences
perfect rationality or bounded rationality
flat or modular or hierarchical
single agent or multiple agents
knowledge is given or knowledge is learned
reason offline or reason while interacting with environment
2 / 11
Classical Planning
deterministic or stochastic dynamics

fully observable or partially observable
explicit states or features or individuals and relations
static or finite stage or indefinite stage or infinite stage
goals or complex preferences
perfect rationality or bounded rationality
flat or modular or hierarchical
single agent or multiple agents
knowledge is given or knowledge is learned
reason offline or reason while interacting with environment
3 / 11
Planning
Planning is deciding what to do based on an agent’s ability,

its goals, and the state of the world.
4 / 11
Planning

Initial assumptions:
► The world is deterministic. There are no exogenous events
outside of the control of the robot that change the state of the
world.
© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 4 / 11

Planning

world.
► The agent knows what state it is in. (Fully observable)
4 / 11
Planning

world.
► Time progresses discretely from one state to the next.

Planning

world.
► Goals are predicates of states that need to be achieved.

Planning

world.
► Goals are predicates of states that need to be achieved.
Planning is finding a sequence of actions to solve a goal.
4 / 11
Actions
A deterministic action is a partial function from states to

states.
5 / 11
Actions

states.
The preconditions of an action specify when the action can be
carried out.

Actions

states.
The preconditions of an action specify when the action can be
carried out.
The effect of an action specifies the resulting state.

Delivery Robot Example
Coffee
Shop
(cs) Sam's
Office
(off )
Mail Lab
Room (lab)
(mr )
Actions:
Features:
mc – move clockwise
RLoc – Rob’s location
mcc – move counterclockwise
RHC – Rob has coffee
puc – pickup coffee
SWC – Sam wants coffee
dc – deliver coffee
MW – Mail is waiting
pum – pickup mail
RHM – Rob has mail
dm – deliver mail
6 / 11
Explicit State-space Representation
State Action Resulting State

⟨lab, ¬rhc, swc, ¬mw , rhm⟩ mc ⟨mr , ¬rhc, swc, ¬mw , rhm⟩
⟨lab, ¬rhc, swc, ¬mw , rhm⟩ mc ⟨off , ¬rhc, swc, ¬mw , rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ c ⟨off , ¬rhc, swc, ¬mw , ¬rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ dm ⟨cs, ¬rhc, swc, ¬mw , rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ mc ⟨lab, ¬rhc, swc, ¬mw , rhm⟩
... c ...
mc
...
7 / 11
Explicit State-space Representation
State Action Resulting State

⟨lab, ¬rhc, swc, ¬mw , rhm⟩ mc ⟨mr , ¬rhc, swc, ¬mw , rhm⟩
⟨lab, ¬rhc, swc, ¬mw , rhm⟩ mc ⟨off , ¬rhc, swc, ¬mw , rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ c ⟨off , ¬rhc, swc, ¬mw , ¬rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ dm ⟨cs, ¬rhc, swc, ¬mw , rhm⟩
⟨off , ¬rhc, swc, ¬mw , rhm⟩ mc ⟨lab, ¬rhc, swc, ¬mw , rhm⟩
... c ...
What happens when we also wantmc to model its battery charge?
Want “elaboration tolerance”. . . .
7 / 11
STRIPS Representation
The state is a function from features into values. So it can be

represented as a set of feature-value pairs.
For each action:
precondition a set of feature-value pairs that must be true for
the action to be carried out.
8 / 11

For each action:
effect a set of a feature-value pairs that are made true by this
action.
8 / 11

For each action:
effect a set of a feature-value pairs that are made true by this
action.
STRIPS assumption: every feature not mentioned in the effect is
unaffected by the action.
8 / 11
Example STRIPS representation
Pick-up coffee (puc):

precondition: {RLoc = cs, RHC = False}. “The robot needs
to be at the coffee shop and not holding coffee in order to
pick up coffee.”
9 / 11

pick up coffee.”
effect: {RHC = True}. “After the robot picks up coffee, it is
holding coffee. Nothing else has changed.”
9 / 11

pick up coffee.”
Deliver coffee (dc):
precondition: {Rloc = off , RHC = True} “The robot needs
to be at the office and holding coffee in order to deliver
coffee.”
9 / 11

pick up coffee.”
Deliver coffee (dc):
precondition: {Rloc = off , RHC = True} “The robot needs
to be at the office and holding coffee in order to deliver
coffee.”
effect: {RHC = False, SWC = False} “After the robot
delivers coffee, it is not holding coffee, and Sam no longer
wants coffee. Nothing else has changed.”

Deterministic Planning
Given:
A description of the effects and preconditions of the actions
A description of the initial state
A goal to achieve
find a sequence of actions that is possible and will result in a state
satisfying the goal.
10 / 11
Deterministic Planning
Given:
A description of the effects and preconditions of the actions
A description of the initial state
A goal to achieve
find a sequence of actions that is possible and will result in a state
satisfying the goal.
Artificial Intelligence, Lecture 6.2 1/ 6

Forward Planning
Idea: search in the state-space graph.

The nodes represent the states

Forward Planning

There is an arc ⟨s, s J⟩ labeled with action A if
► A is an action that can be carried out in state s and
► s ′ is the state resulting from doing A in state s

Forward Planning

There is an arc ⟨s, s J⟩ labeled with action A if
► A is an action that can be carried out in state s and
► s ′ is the state resulting from doing A in state s
A plan is a path from the state representing the initial state to
a state that satisfies the goal.

Example state-space graph
Actions
〈cs, ¬rhc, swc, mw, ¬rhm〉
mc: move clockwise puc mcc
mcc: move counterclockwise mc
nm: no move 〈cs, rhc, swc, mw, ¬rhm〉〈mr, ¬rhc, swc, mw, ¬rhm〉
puc: pick up coffee
dc: deliver coffee 〈off, ¬rhc, swc, mw, ¬rhm〉
mcc
pum: pick up mail mc mcc mc
dm: deliver mail 〈cs, ¬rhc, swc, mw, ¬rhm〉
〈lab, ¬rhc, swc, mw, ¬rhm〉 Locations:

〈off, rhc, swc, mw, ¬rhm〉
cs: coffee shop
dc 〈mr, rhc, swc, mw, ¬rhm〉 off: office
lab: laboratory
mc
mcc mr: mail room
〈off, ¬rhc, ¬swc, mw, ¬rhm〉
Feature values
〈lab, rhc, swc, mw, ¬rhm〉 rhc: robot has coffee
〈cs, rhc, swc, mw, ¬rhm〉 swc: Sam wants coffee
mw: mail waiting
rhm: robot has mail

Forward planning representation
The search graph can be constructed on demand: you only

construct reachable states.


If you want a cycle check or multiple path-pruning, you need
to be able to find repeated states.


There are a number of ways to represent states:


► As a map from features into their values


► As a map from features into their values
► As a path from the start state

Improving Search Efficiency
Forward search can use domain-specific knowledge specified as:

a heuristic function that estimates the cost from a complete
state description to a goal.


domain-specific pruning of neighbors:
► don’t pick-up coffee unless Sam wants coffee.


► unless the goal involves time constraints, don’t do a “no
move” action..


► unless the goal involves time constraints, don’t do a “no
move” action..
► don’t go to the coffee shop unless “Sam wants coffee” is part
of the goal and Rob doesn’t have coffee

© D . L. Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.2 6/ 6
Regression Planning
Idea: search backwards from the goal description: nodes

correspond to subgoals, and arcs to actions.

Regression Planning

A subgoal is an assignment of values to some features.

Regression Planning

Search problem:
Nodes are subgoals

Regression Planning

Search problem:
Nodes are subgoals
There is an arc ⟨g,g J⟩ labeled with action A if
► A achieves one of the assignments in g

Regression Planning
Idea: searchbackwards from the goal description: nodes

Search problem:
Nodes are subgoals
► gJ is a proposition that must be true immediately before
action A so that g is true immediately after.

Regression Planning

Search problem:
Nodes are subgoals
The start node is the goal to be achieved.

Regression Planning

Search problem:
Nodes are subgoals
The start node is the goal to be achieved.
goal(g ) is true if g is a proposition that is true of the
initial state.

Defining nodes and arcs
A node g can be represented asa set of assignments of

values to variables:
[X 1 = v1 ,...,X n = vn ]
This is a set of assignments you want to hold.


[X 1 = v1 ,...,X n = vn ]

The last action achieves one of the X i = vi ,


[X 1 = v1 ,...,X n = vn ]

The last action achieves one of the X i = vi , and does not
achieve X j = vjJ where vjJ is different to vj .


[X 1 = v1 ,...,X n = vn ]

The neighbor of g along arc A must contain:


[X 1 = v1 ,...,X n = vn ]

► The prerequisites of action A


[X 1 = v1 ,...,X n = vn ]

► All of the elements of g that were not achieved by A


[X 1 = v1 ,...,X n = vn ]

► All of the elements of g that were not achieved by A
it must be consistent = have at most one value for each
feature.

Formalizing arcs using STRIPS notation
⟨g,g J⟩ is an arc labeled with action A where g is

[X 1 = v1 ,...,X n = vn ] and A is an action, if


∃i X i = vi is on the effects list of action A


∀j X j = vjJ is not on the effects list for A, where vjJ /= vj


g J = preconditions(A) ∪ {X k = vk ∈ g : X k = vk ∈ /
effects(A)}


g J = preconditions(A) ∪ {X k = vk ∈ g : X k = vk ∈ /
effects(A)} if it is consistent

Regression example
Locations:
Actions [¬ swc] cs: coffee shop Feature values
mc: move clockwise off: office rhc: robot has coffee
mcc: move counterclockwise dc lab: laboratory swc: Sam wants coffee
puc: pick up coffee mr: mail room mw: mail waiting
dc: deliver coffee [off,rhc] rhm: robot has mail
pum: pick up mail
dm: deliver mail mc mcc
[cs,rhc] [lab,rhc]
mc mcc mcc
puc mc
[mr,rhc] [off,rhc] [mr,rhc]
[cs] [off,rhc]

Loop detection and multiple-path pruning
Goal G1 is simpler than goal G2 if G1 is a subset of G2.

► It is easier to solve [cs] than [cs, rhc].
If you have a path to node N have already found a path
to a simpler goal, you can prune the path N.

Improving Efficiency
You can define a heuristic function that estimates how

difficult it is to solve a goal from a state.

Improving Efficiency
You can define a heuristic function that estimates how

difficult it is to solve a goal from a state.
You can usedomain-specific knowledge to remove
impossible goals, e.g.
► It is often not obvious from an action description to
conclude whether an agent can hold multiple items at
any time.

Comparing forward and regression planners
Which is more efficient depends on:

► The branching factor


► How good the heuristics are


► How good the heuristics are
Forward planning is unconstrained by the goal (except as
a source of heuristics).
Regression planning is unconstrained by the initial state
(except asa source of heuristics)

Planning as a CSP
In forward planning, the nodes considered are constrained to

be reachable, even if they don’t lead to goal.

Planning as a CSP

In regression planning, the nodes considered are constrained
to be ones from which we can achieve the goal, even if they
are not reachable.

Planning as a CSP

are not reachable.
When representing planning as a CSP, we can constrain the
states by the starting state and the goal.

Planning as a CSP

are not reachable.
. . . but we can only do this if we know the number of steps.

Planning as a CSP

are not reachable.
Search over planning horizons (number of time steps).

Planning as a CSP

are not reachable.
Search over planning horizons (number of time steps).
For each planning horizon, create a CSP constraining possible
actions and features

CSP Variables
Choose a planning horizon k.

Create a variable for each state feature and each time from 0
to k.
Create a variable for the action for each time in the range 0 to
k − 1.

CSP for Delivery Robot for a planning horizon of 2
RLoc 0 RLoc1 RLoc2
RHC0 RHC1 RHC2
Action0 Action1
SWC0 SWC1 SWC2
MW 0 MW 1 MW 2
RHM0 RHM1 RHM2
State 0 Action 0 State 1 Action 1 State 2
RLoci — Rob’s location Movei — Rob’s move action

RHCi — Rob has coffee PUCi — Rob picks up coffee
SWCi — Sam wants coffee DelC — Rob delivers coffee
MWi — Mail is waiting PUMi — Rob picks up mail
RHMi — Rob has mail DelMi — Rob delivers mail

Constraints
precondition constraints between state variables at time t and

action variable at time t, specify constraints on what actions
are available from a state.

Constraints

effect constraints between state variables at time t, action
variable at time t and state variables at time t + 1 constrain
the resulting state to be one that satisfies the effects.

Constraints

frame constraints among state variables at time t, action
variables at time t, and state variables at time t + 1 for values
of variables that do not change.

Constraints

frame constraints among state variables at time t, action
variables at time t, and state variables at time t + 1 for values
of variables that do not change.
initial state constraints that are usually domain constraints on
the initial state (at time 0).
goal constraints that constrains the final state to be a state
that satisfies the goals that are to be achieved.

Example Constraints
precondition constraint RLoci = cs if Actioni = puc

is violated when

Example Constraints

is violated when RLoci /= cs ∧Actioni = puc

Example Constraints

effect constraint rhci +1 if Acti = puc
is violated when

Example Constraints

is violated when RHCi+1 = false ∧Acti = puc.

Example Constraints

/ cs ∧Actioni = puc
is violated when RLoci =
frame constraint Rob has mail at any time if it had mail before
and the action wasn’t

Example Constraints

and the action wasn’t to pickup mail or deliver mail:
/ { pum, dm}
RHMi +1 = RHMi if Acti ∈
violated when

Example Constraints

and the action wasn’t to pickup mail or deliver mail:
/ { pum, dm}
RHMi +1 = RHMi if Acti ∈
violated when
RHMi +1 /= RHMi ∧Acti /= pum ∧Acti /= dm

Classical Planning - Final - Notes

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Classical Planning - Final - Notes

Uploaded by

Copyright:

Available Formats

Learning Objectives

At the end of the class you should be able to:

deterministic or stochastic dynamics

deterministic or stochastic dynamics

Planning is deciding what to do based on an agent’s ability,

Planning is deciding what to do based on an agent’s ability,

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 4 / 11

Planning is deciding what to do based on an agent’s ability,

Planning is deciding what to do based on an agent’s ability,

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 4 / 11

Planning is deciding what to do based on an agent’s ability,

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 4 / 11

Planning is deciding what to do based on an agent’s ability,

A deterministic action is a partial function from states to

A deterministic action is a partial function from states to

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 5 / 11

A deterministic action is a partial function from states to

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 5 / 11

State Action Resulting State

State Action Resulting State

The state is a function from features into values. So it can be

The state is a function from features into values. So it can be

The state is a function from features into values. So it can be

Pick-up coffee (puc):

Pick-up coffee (puc):

Pick-up coffee (puc):

Pick-up coffee (puc):

© D . L . Poole and A.K. Mackworth 2010-2020 Artificial Intelligence, Lecture 6.1 9 / 11

Artificial Intelligence, Lecture 6.2 1/ 6

Idea: search in the state-space graph.

Artificial Intelligence, Lecture 6.2 2/ 6

Idea: search in the state-space graph.

Artificial Intelligence, Lecture 6.2 2/ 6

Idea: search in the state-space graph.

Artificial Intelligence, Lecture 6.2 2/ 6

〈lab, ¬rhc, swc, mw, ¬rhm〉 Locations:

Artificial Intelligence, Lecture 6.2 3/ 6

The search graph can be constructed on demand: you only

Artificial Intelligence, Lecture 6.2 4/ 6

The search graph can be constructed on demand: you only

Artificial Intelligence, Lecture 6.2 4/ 6

The search graph can be constructed on demand: you only

Artificial Intelligence, Lecture 6.2 4/ 6

The search graph can be constructed on demand: you only

Artificial Intelligence, Lecture 6.2 4/ 6

The search graph can be constructed on demand: you only

Artificial Intelligence, Lecture 6.2 4/ 6

Forward search can use domain-specific knowledge specified as:

Artificial Intelligence, Lecture 6.2 5/ 6

Forward search can use domain-specific knowledge specified as:

Artificial Intelligence, Lecture 6.2 5/ 6

Forward search can use domain-specific knowledge specified as:

Artificial Intelligence, Lecture 6.2 5/ 6

Forward search can use domain-specific knowledge specified as:

Artificial Intelligence, Lecture 6.2 5/ 6

Idea: search backwards from the goal description: nodes

Artificial Intelligence, Lecture 6.3 1/ 8

Idea: search backwards from the goal description: nodes

Artificial Intelligence, Lecture 6.3 1/ 8

Idea: search backwards from the goal description: nodes

Artificial Intelligence, Lecture 6.3 1/ 8

Idea: search backwards from the goal description: nodes