Artificial Intelligence 3451

UNIT: 05
Planning and Reasoning
Lecture: 12

Farhad Muhammad Riaz

From Problem Solving to Planning
 Planners can
 Use FOL (for example) for representing states, goals, and
actions
 Break hard problems into manageable subproblems
 Solve some subproblems independently of others
Planning
 Could use standard FOL inference (e.g., resolution) to find
plan
 But this isn’t practical: Huge search space
 Instead, use special-purpose (=efficient) algorithms, each with
its own restricted (= smaller search space) language
STRIPS Representation for Planning

 STRIPS language
 States = conjunctions of function-free ground literals
 Goals = same thing, but can also have variables (assumed to be
existentially quantified)
STRIPS Representation for Actions

 Actions have three components:

 Action description = just the name
 Precondition = conjunction of positive literals that says what
must be true to make the action applicable
 Effect = conjunction of positive or negative literals that says
what changes in the state when the action is applied
 Can be separated into an add list and a delete list
Representations for Plans
 Relative ordering of plan steps
 Total order (linearization): all steps in order
 Partial order: some steps in order
Plan Solutions

 Problems with linearized, totally ordered plans

 Which linearization to choose?
 What if agent can perform some steps in parallel?
 When combining plans (e.g., for subproblems), partial
ordering gives flexibility
 Solution = an executable plan that guarantees
reaching the goal
 Complete plans: all preconditions of all steps are achieved
by prior steps
 Consistent plans: no inconsistencies in ordering
constraints or causal links
Plan Space

 Searching through plan space

 Each node is a partial plan, which represents a set of complete
plans
 Search operators modify plans: add steps, reorder steps, etc.
 Progression (forwards) vs. Regression (backwards)
planning
The Plan Data Structure

 A Plan data structure contains:

 A set of plan steps, Si
 A set of step ordering constraints
 A set of open preconditions
 A set of causal links
 Initial plan: Steps = Start and Finish
 Start’s effects = initial state. No preconditions.
 Finish’s preconditions = goal state. No effects.
Clobbering
Partial-Order Planning

 Start with the minimal partial plan

 Each point in planning adds a step (and causal links) that
achieves some precondition
 Can use standard search algorithms
 The key is the formulation: Nodes aren’t world states, they’re
partial plans
Partial-Order Planning Example

 Start at home
 Goal = return home with milk, bananas, and drill
 Only add steps that achieve an unachieved precondition ->
Reduces search space
 Causal links help prune bad plans quickly, before going
too far down the dead end
Example
Example
Example
Situation Calculus
 Formalization of actions in first-order logic (McCarthy & Hayes, 1969).
Search performed by logical inference.
 Situations are logical terms denoting states of the world.
 Actions and facts are represented as logical terms called fluents.
 puton(A,B): The action of putting block A on block B.
 on(A,B): The proposition that block A is on block B.
 Propositional fluents are asserted to be true in a particular
 state by using the predicate: holds
 holds(on(A,B), s): A is on B in situation s
 Situations resulting from performing an action in another situation are
represented using the function: result
 result(puton(A,B), s): The situation resulting from putting A on B in
situation s.
An example – Blocks world
 Blocks on a table
 Can be stacked, but only one block on top of another
 A robot arm can pick up a block and movie to another
position
 On the table
 On another block
 Arm can pick up only one block at a time
 Cannot pick up a block that has another one on it

19/34 AI Lecture 17
Representation language
 STRIPS (Fikes & Nilsson, 1971)
 One of the first planning systems
 Robotics
 See it in action at
http://www.ai.sri.com/movies/Shakey.ram
 Main contribution is its language description formalism
 Many variants/extensions

20/34 AI Lecture 17
STRIPS
 State is a conjunction of positive ground literals
On(B, Table) Λ Clear (A)
 Goal is a conjunction of positive ground literals
Clear(A) Λ On(A,B) Λ On(B, Table)
 Action schema
 Conjunction of positive literals as preconditions
 Conjunction of positive and negative literals as effects

21/34 AI Lecture 17
More on action schema

 Example: Move (b, x, y)

 Precondition:
Block(b) Λ Clear(b) Λ Clear(y) Λ On(b,x) Λ (b ≠ x) Λ (b ≠ y) Λ (y ≠ x)
 Effect:
¬Clear(y) Λ ¬On(b,x) Λ Clear(x) Λ On(b,y)

Delete list Add list

 An action is applicable in any state that satisfies its

precondition
22/34 AI Lecture 17
STRIPS assumptions
 Closed World assumption
 Unmentioned literals are false (no need to explicitly list out)
 w/an open world assumption unmentioned literals are unknown
 STRIPS assumption
 Every literal not mentioned in the “effect” of an action schema
remains unchanged

23/34 AI Lecture 17
STRIPS expressiveness
 Literals are function free: Move (Block(x), y, z)
 Any action scheme can be propositionalized
Move(b,x,y) and 3 blocks and table can be expressed as 48 purely
propositional actions
 No disjunctive goals: On(B, Table) V On(B, C)
 No conditional effects: On(B, Table) if ¬On(A, Table)
 In original STRIPS, no equality: x ≠ y
 Ramification – esp. in more complex domain

24/34 AI Lecture 17
Forward state-space search (1)
 Progression
 Initial state: initial state of the problem
 Actions:
 Applied to a state if all the preconditions are satisfied
 Succesor state is built by updating KB with add and delete lists
 Goal test: state satisfies the goal of the problem

25/34 AI Lecture 17
Forward search in the Blocks world

26/34 AI Lecture 17
Forward state-space search (2)
 Advantages
 No functions in the declarations of goals 
search state is finite
 Sound
 Complete (if algorithm used to do the search is complete)
 Limitations
 Irrelevant actions  not efficient
 Need heuristic or prunning procedure

27/34 AI Lecture 17
Backward state-space search (1)
 Regression
 Initial state: goal state of the problem
 Actions:
 Choose an action that
 Is relevant; has one of the goal literals in its effect set
 Is consistent; does not negate another literal
 Construct new search state
 Remove all positive effects of A that appear in goal
 Add all preconditions, unless already appears
 Goal test: state is the initial world state

28/34 AI Lecture 17
Backward state-space search (2)
 Possible because of STRIPS-like language
 Goals are listed
 Predecessors are listed for each action/state
 Advantages
 Consider only relevant actions  much smaller branching
factor
 Ways to reduce even more the branching factors
 Limitations
 Still need heuristic to be more efficient

29/34 AI Lecture 17
Heuristics for state-space search (1)
 Valid both for forward and backward searches
 Valid for many planning problems
 Possible approaches
 Divide and conquer
 Derive a relaxed problem
 Combine both

30/34 AI Lecture 17
Heuristics for state-space search (2)
 Divide and conquer
 Subgoal independence assumption
 What if there are negative interactions between the subgoals of the
problems?
 What if there are redundant actions in the subgoals?
 Derive a relaxed problem
 Remove all preconditions from the actions
 Remove all negative effects from the actions
(empty delete list)

31/34 AI Lecture 17
Limitation of state-space search
 Linear planning or Total order planning
 Example
 Initial state: all the blocks are clear and on the table
 Goal: On(A,B) ∧ On(B,C)
 If search achieves On(A,B) first, then needs to undo it in order
to achieve On(B,C)
 Have to go through all the possible permutations of the
subgoals

32/34 AI Lecture 17