02 Search

Artificial Intelligence:
State Space Search
 Many slides from:

robotics.stanford.edu/~latombe/cs121/2003/home.htm
1
Motivation
Google itinerary
8-puzzle
Rubik’s cube Tetris
2
Today
 State Space Representation Agent Perception
 State Space Search Robotics
 Uninformed search
 Breadth-first and Depth-first Reasoning
 Depth-limited Search Search
 Iterative Deepening Learning
 Uniform Cost
 Informed search KnowledgeConstraint
Planning rep. satisfaction
 Hill climbing
 Best-First
 (Designing Heuristics)
 A* Natural
... Expert
language
 Summary Systems
3
Example: 8-Puzzle
State: Any arrangement of 8 numbered tiles and an empty tile on a 3x3 board
8 2 1 2 3
3 4 7 4 5 6
5 1 6 7 8
Initial state Goal state

there are several standard goals states for the 8-puzzle
1 2 3 1 2 3
4 5 6 8 4 …
7 8 7 6 5
4
(n2-1)-puzzle
....
8 2 1 2 3 4
3 4 7 5 6 7 8
5 1 6 9 10 11 12
13 14 15
8-puzzle
15-puzzle
5
15-Puzzle
Invented in 1874 by Noyes Palmer Chapman
… but Sam Loyd claimed he invented it!
6
15-Puzzle
Sam Loyd even offered $1,000 of his own
money to the first person who would solve the
following problem:
1 2 3 4 1 2 3 4
5 6 7 8 ? 5 6 7 8
9 10 11 12 9 10 11 12
13 14 15 13 15 14
7
But no one ever won the prize…
8
State Space
 Many AI problems, can be expressed in terms of going
from an initial state to a goal state
 Ex: to solve a puzzle, to drive from home to Concordia…
 Often, there is no direct way to find a solution to a

problem
 but we can list the possibilities and search through them
1. Brute force search:

 generate and search all possibilities (but inefficient)
2. Heuristic search:
 only try the possibilities that you think (based on your
current best guess) are more likely to lead to good solutions
9
State Space
 Problem is represented by:
1. Initial State
 starting state
 ex. unsolved puzzle, being at home
2. Set of operators
 actions responsible for transition between states
3. Goal test function
 Applied to a state to determine if it is a goal state
 ex. solved puzzle, being at Concordia
4. Path cost function
 Assigns a cost to a path to tell if a path is preferable to
another
 Search space: the set of all states that can be reached
from the initial state by any sequence of action
 Search algorithm: how the search space is visited
10
Example: The 8-puzzle
8 2 1 2 3
3 4 7 4 5 6
5 1 6 7 8
Initial state Goal state
Set of operators:
blank moves up, blank moves down, blank moves left, blank moves right
Goal test function:

state matches the goal state
Path cost function:

each movement costs 1
so the path cost is the length of the path (the number of moves)
source: G. Luger (2005)

11
8-Puzzle: Successor Function
8 2 7
3 4
5 1 6
8 2 8 2 7 8 2 7
3 4 7 3 4 6 3 4
5 1 6 5 1 5 1 6
Search is about the exploration of alternatives

12
State Graph
 Each state is
represented by a
distinct node
 An arc (or edge)
connects a node s
to a node s’ if
s’ ∈ SUCCESSOR(s)
 The state graph may
contain more than one
connected component
13
Just to make sure we’re clear…
5 8 1 2 3
4 2 1 4 5 6
7 3 6 7 8
Initial state Goal state 14
State Space as a Search Tree
 In graph representation, cycles can prevent

termination
 Blind search without cycle check may never Search tree
terminate
 Use a tree representation, and check for cycles
15
State Space for the 8-puzzle
Right Up Left
Down

16
How large is the state space of the
(n2-1)-puzzle?
 Nb of states:
 8-puzzle --> 9! = 362,880 states
 15-puzzle --> 16! ~ 2.09 x 1013 states
 24-puzzle --> 25! ~ 1025 states
 At 100 millions states/sec:

 8-puzzle --> 0.036 sec
 15-puzzle --> ~ 55 hours
 24-puzzle --> > 109 years
17
Today
 State Space Representation
 State Space Search
 Breadth-first and Depth-first
 Depth-limited Search
 Iterative Deepening
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
18
Uninformed VS Informed Search
 Uninformed search
 We systematically explore the alternatives
 aka: systematic/exhaustive/blind/brute force search
 Breadth-first
 Depth-first
 Uniform-cost
 Depth-limited search
 Iterative deepening search
 Bidirectional search
 …
 Informed search (heuristic search)

 We try to choose smartly
 Hill climbing
 Best-First
 A*
 …
19
Today
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
20
Breadth-first vs Depth-first Search
 Determine order for examining states

 Depth-first:
 visit successors before siblings
 Breadth-first:
 visit siblings before successors
 ie. visit level-by-level

21
Data Structures
 In all search strategies, you need:
 open list (aka the frontier)
 lists generated nodes not yet expanded
 order of nodes controls order of search
 closed list (aka the explored set)

 stores all the nodes that have already been visited (to avoid cycles).
 ex:
Closed = [A, B, C, D, E]
Open = [F, G, H, I, J, K, L]

22
Data Structures
 To trace back the entire path of the solution after
the search, each node in the lists contain:
STATE PARENT-NODE
8 2 REPRESENTATION
3 4 7
5 1 6
BOOKKEEPING
Action Right
CHILDREN
Depth 5
... Path-Cost 5
Depth = length of path from root to node
23
Generic Search Algorithm
1. Initialize the open list with the initial node so (top node)
2. Initialize the closed list to empty
3. Repeat
a) If the open list is empty, then exit with failure.
b) Else, take the first node s from the open list.
c) If s is a goal state, exit with success. Extract the solution path
from s to so
d) Else, insert s in the closed list (s has been visited /expanded)
e) Insert the successors of s in the open list in a certain order if
they are not already in the closed and/or open lists (to avoid
cycles)
Notes:
 The order of the nodes in the open list depends on the search
strategy
24
DFS and BFS
 DFS and BFS differ only in the way they order nodes in the
open list:
 DFS uses a stack:

 nodes are added on the top of the list.
 BFS uses a queue:

 nodes are added at the end of the list.
25
Breadth-First Search

26
Breadth-First Search Example
 BFS: (open is a queue)
Assume U is goal state
1. open = [A-null] closed = []

2. open = [B-A C-A D-A] closed [A]
3. open = [C-A D-A E-B F-B] closed = [B A]
4. → Worksheet #1 (“Breadth-First Search”)
...and so on until either U is found or open = []

27
Snapshot of BFS
 Search graph at
iteration 6 of
breadth-first
search
 States on open
and closed are
highlighted

29
Function Depth-First Search

30
Depth-First Search Example
 DFS: (open is a stack)
Assume U is goal state
1. open = [A-null] closed = []

2. open = [B-A C-A D-A] closed [A]
3. open = [E-B F-B C-A D-A] closed = [B A]
4. open = [K-E L-E F-B C-A D-A] closed = [E B A]
5. open = [S-K L-E F-B C-A D-A] closed = [K E B A]
→ Worksheet #1 (“Depth-First Search”)

31
Snapshot of DFS
 Search graph at
iteration 6 of
depth-first
search
 States on open
and closed are
highlighted

33
Depth-first vs. Breadth-first
 Breadth-first:
 Optimal: will always finds shortest path
But:
 inefficient if branching factor B is very high
 memory requirements high -- exponential space for states
required: Bn
 Depth-first:
 Not optimal (no guarantee to find the shortest path)
But:
 Requires less memory
 But both search are impractical in real applications

because search space is too large!
34
Today
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
35
Depth-Limited Search
Compromise for DFS :

 Do depth-first but with depth cutoff k (depth
at which nodes are not expanded)
 Three possible outcomes:

 Solution
 Failure (no solution)
 Cutoff (no solution within cutoff)
36
Today
 Breath-first and Depth-first
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
37
Iterative Deepening
Compromise between BFS and DFS:
 use depth-first search, but
 with a maximum depth before going to next level
 Repeats depth first search with gradually increasing

depth limits
 Requires little memory (fundamentally, it’s a depth first)
 Finds the shortest path (limited depth)
 Preferred search method when there is a large search

space and the depth of the solution is unknown
38
Iterative Deepening: Example
source: Russel & Norvig (2003)

39

40

41

42
Today
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
43
Uniform Cost Search
 Breadth First Search
 Open is a priority queue sorted using the depth of the nodes from
the root
 guarantees to find the shortest solution path
 But what if all edges/moves do not have the same cost?

 Uniform Cost Search
4 3
 uses a priority queue sorted using 2
the cost from the root to node n –
later called g(n) 6 3
 guarantees to find the lowest cost
solution path
44
Today
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
45
Informed Search (aka heuristic search)
 Most of the time, it is not feasible to do an exhaustive search,
search space is too large
 e.g. state space of all possible moves in chess = 10120
 1075 = nb of molecules in the universe
 1026 = nb of nanoseconds since the “big bang”
 so far, all search algorithms have been uninformed (general

search)
 so need an informed/heuristic search
 Idea:
 choose "best" next node to expand
 according to a selection function (i.e. a heuristic function h(n))
 But: heuristic might fail
46
Heuristic - Heureka!
 Heuristic:
 a rule of thumb, a good bet
 but has no guarantee to be correct whatsoever!
 Heuristic search:
 A technique that improves the efficiency of search,
possibly sacrificing on completeness
 Focus on paths that seem most promising according to
some function
 Need an evaluation function (heuristic function) to
score a node in the search tree
 Heuristic function h(n) = an approximation of the
lowest cost from node n to the goal
47
g(n)
4 3
2
6 3
2 1
 g(n) = cost of current path from start to node n
48
h(n)
3 6
 h(n) = estimate of the lowest cost from n to goal
49
Today
 Informed search
 Hill Climbing
 Best-First
 A*
 Summary
50
Example: Hill Climbing with Blocks World
 Operators:
 pickup&putOnTable(Block)
 pickup&stack(Block1,Block2)
 Heuristic:
 0pt if a block is sitting where it is supposed to sit
 +1pt if a block is NOT sitting where it is supposed to sit
 so lower h(n) is better

 h(initial) = 2
 h(goal) = 0
source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991.

51
h(n) = 2
pickup&putOnTable(A)
h(n) = 1
pickup&stack(A,H) pickup&putOnTable(H)
pickup&stack(H,A)
h(n) = 2 h(n) = 2 h(n) = 2

52
Hill Climbing
 General hill climbing strategy:
 as soon as you find a position that is better than the
current one, select it.
 Does not maintain a list of next nodes to visit (an open list)
 Similar to climbing a mountain in the fog with amnesia …
always go higher than where you are now, help!
but never go back…
 Steepest ascent hill climbing:

 instead of moving to the first position that
is better than the current one
 pick the best position out of all the next possible moves
53
Steepest Ascent Hill Climbing
currentNode = startNode;
loop do
L = CHILDREN(currentNode);
nextEval = INFINITY;
nextNode = NULL;
for all c in L
if (HEURISTIC-VALUE(c) < nextEval) // lower h is better
nextNode = c;
nextEval = HEURISTIC-VALUE(c);
if nextEval >= HEURISTIC-VALUE(currentNode)

// Return current node since no better child state exists
return currentNode;
currentNode = nextNode;
Adapted from https://en.wikipedia.org/wiki/Hill_climbing

54
h(n) = 2
hill-climbing will stop, pickup&putOnTable(A)

because all children have
higher h(n) than the
parent… --> local minimum h(n) = 1
pickup&stack(A,H) pickup&putOnTable(H)
pickup&stack(H,A)
Don’t be confused…
a lower h(n) is better…
h(n) = 2 h(n) = 2 h(n) = 2

55
Problems with Hill Climbing
 Foothills (or local maxima)
 reached a local maximum, not the global maximum
 a state that is better than all its neighbors but is not better
than some other states farther away.
 at a local maximum, all moves appear to make things worse.
 ex: 8-puzzle: we may need to move tiles temporarily out of goal
position in order to place another tile in goal position
 ex: TSP: "nearest neighbour" heuristic
n
ent
es
h(n) e pr
or
rt
te
ame
ar
rp
t he
m eo
So
Some parameter to represent n
56
Problems with Hill Climbing
 Plateau
 a flat area of the search space in which the next states
have the same value.
 it is not possible to determine the best direction in which
to move by making local comparisons.
source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991. 57

Some Solutions to Hill-Climbing
 Random-restart hill-climbing
 random initial states are generated
 run each until it halts or makes no significant progress.
 the best result is then chosen.
 keep going even if the best successor has the same value as
current node
 works well on a "shoulder"
 but could lead to infinite loop
on a plateau
source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991. & Russel & Norvig (2003)
58
Today
 Informed search
 Hill Climbing
 Best-First
 A*
 Summary
59
Best-First Search
 problem with hill-climbing:
 one move is selected and all others are forgotten.
 solution to hill-climbing:
 use "open" as a priority queue
 this is called best-first search
 Best-first search:
 Insert nodes in open list so that the nodes are sorted in
ascending h(n)
 Always choose the next node to visit to be the one with the
best h(n) -- regardless of where it is in the search space
60
Best-First: Example
Lower h(n) is better
source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991.

61
Notes on Best-first
 If you have a good h(n), best-first can find the

solution very quickly
 The first solution may not be the best,

but there is a good chance of finding it quickly
 It is an exhaustive search …
 will eventually try all possible paths
62
Best-First Search: Example
1. open = [A-null-5] closed = []
2. open = [B-A-4 C-A-4 D-A-6] (arbitrary choice) closed [A]
3. open = [C-A-4 E-B-5 F-B-5 D-A-6] closed = [B A]
4. → Worksheet #1 (“Best-First Search”)
Lower h(n) is better P-0
source: adapted from G. Luger (2005)

63
Today
 Informed search
 Hill Climbing
 Best-First
 A*
 Summary
65
Designing Heuristics
 Heuristic evaluation functions are highly

dependent on the search domain
 In general: the more informed a heuristic is,
the better the search performance
 Bad heuristics lead to frequent backtracking
 So how do we design a “good” heuristic?
66
Example: 8-Puzzle – Heuristic 1
 h1: Simplest heuristic
 Hamming distance : count
number of tiles out of place

when compared with goal
5 8 1 2 3
4 2 1 4 5 6
7 3 6 7 8
STATE n Goal state
 h1(n) = 6
 does not consider the
distance tiles have to be

moved

67
 h2: Better heuristic
 Manhattan distance: sum up
all the distances by which

tiles are out of place
5 8 1 2 3
4 2 1 4 5 6
7 3 6 7 8
STATE n Goal state
 h2(n) = 2+3+0+1+3+0+3+1
= 13

68
 h3: Even Better
 sum of permutation
inversions
 See next slide…

69
h3(N) = sum of permutation inversions
5 8 5 8 4 2 1 7 3 6
4 2 1
7 3 6
STATE n
 For each numbered tile, count how many tiles on its right
should be on its left in the goal state.
1 2 3
 h3(n) = n5 + n8 + n4 + n2 + n1 + n7 + n3 + n6 4 5 6
=4 +6 +3 +1 +0 +2 +0 +0 7 8
= 16
Goal state
70
Heuristics for the 8-Puzzle
5 8 1 2 3
4 2 1 4 5 6
7 3 6 7 8
STATE n Goal state
 h1(n) = misplaced numbered tiles
=6
 h2(n) = Manhattan distance

= 2 + 3 + 0 + 1 + 3 + 0 + 3 + 1 = 13
 h3(n) = sum of permutation inversions

= n5 + n8 + n4 + n2 + n1 + n7 + n3 + n6
= 4 + 6 + 3 + 1 + 0 + 2 + 0 + 0 = 16
71
g(n), h(n) and f(n)
 Evaluation function f(n) = g(n) + h(n) for node n:
 g(n) current cost from start to node n
 h(n) estimate of the lowest cost from n to goal
 f(n) estimate of the lowest cost of the solution
path (from start to goal passing through n)
 Now consider f*(n) = g*(n) + h*(n):

 g*(n) cost of lowest cost path from start to
node n
 h*(n) actual lowest cost from n to goal
 f*(n) actual cost of lowest cost of the solution
path (from start to goal passing through n)
72
Evaluating Heuristics
1. Admissibility:
 “optimistic”
 never overestimates the actual cost of reaching the goal
 guarantees to find the lowest cost solution path to the goal (if it
exists)
2. Monotonicity:
 “local admissibility”
 guarantees to find the lowest cost path to each state n encountered
in the search
3. Informedness:
 measure for the “quality” of a heuristic
 the more informed, the better
73
Admissibility
 A heuristic is admissible if it never overestimates the
cost of reaching the goal
 i.e.:
 h(n) ≤ h*(n) for all n
 guarantees to find the lowest cost solution path to the

goal (if it exists)
 Note: does not guarantee to find the lowest cost search
path.
 e.g.: breadth-first is admissible -- it uses f(n) = g(n) + 0
74
Example: 8-Puzzle
5 8 1 2 3
4 2 1 4 5 6
7 3 6 7 8
n goal
 h1(n) = Hamming distance = number of misplaced tiles = 6

--> admissible
 h2(n) = Manhattan distance = 13

--> admissible
75
Monotonicity (aka consistent)
 An admissible heuristics may temporarily reach non-goal states
along a suboptimal path
 A heuristic is monotonic if it always finds the optimal path to
each state the 1st time it is encountered !
 guarantees to find the lowest cost path to each state n
encountered in the search
 h is monotonic if for every node n and every successor n’ of n:

 h(n) ≤ c(n,n') + h(n')
 i.e. f(n) is non-decreasing along any path
76
Informedness
 Intuition: number of misplaced tiles is less informed than
Manhattan distance
 For two admissible heuristics h1 and h2
 if h1(n) ≤ h2(n), for all states n
 then h2 is more informed than h1
 h1(n) ≤ h2(n) ≤ h*(n)
More informed heuristics search smaller space to find the
solution path
 However, you need to consider the computational cost of

evaluating the heuristic…
 The time spent computing heuristics must be recovered
by a better search
77
Today
 Informed search
 Hill Climbing
 Best-First
 A*
 Summary
78
Algorithm A
 Heuristics might be wrong:
 so search could continue down a wrong path
 Solution:
 Maintain depth/cost count, i.e., give preference to
shorter/least expensive paths
 Modified evaluation function f:
f(n) = g(n) + h(n)
 f(n) estimate of total cost along path through n
 g(n) actual cost of path from start to node n
 h(n) estimate of cost to reach goal from node n
79
Algorithm A on the 8-puzzle
→ Worksheet #1 (“Algorithm A”)

80

82

83
BFS vs.
Heuristic search
(tiles out of place)

Algorithm A vs Algorithm A*
 if g(n) ≥ g*(n) for all n
 best-first used with such a g(n) is called “algorithm A”
 if h(n) ≤ h*(n) for all n

 i.e. h(n) never overestimates the true cost from n to a goal
 algorithm A used with such an h(n) is called “algorithm A*”
 an A* algorithm is admissible
 i.e. it guarantees to find the lowest cost solution path
from the initial state to the goal
85
Today
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
86
Summary
Search Uses h(n)? Open is a…
Breadth-first No Queue
Depth-first No Stack
Depth-limited No Stack
Iterative Deepening No Stack
Uniform Cost No Priority queue sorted by g(n)
Hill Climbing Yes none
Best-First Yes Priority queue sorted by h(n)
Algorithm A Yes Priority queue sorted by f(n)
- no constraints on h(n) f(n) = g(n) + h(n)
Algorithm A* Yes Priority queue sorted by f(n)
- same as A, but h(n) must be f(n) = g(n) + h(n)
admissible
- guarantees to find the lowest
cost solution path 87
Today
 Uniform Cost
 Informed search
 Hill climbing
 Best-First
 A*
 Summary
88

02 Search

Uploaded by

Copyright:

Available Formats

02 Search

Uploaded by

Document Information

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

02 Search

Uploaded by

Copyright:

Available Formats

Artificial Intelligence:

State Space Search

 Many slides from:

Rubik’s cube Tetris

Initial state Goal state

 Often, there is no direct way to find a solution to a

1. Brute force search:

Goal test function:

Path cost function:

source: G. Luger (2005)

Search is about the exploration of alternatives

 In graph representation, cycles can prevent

source: G. Luger (2005)

 At 100 millions states/sec:

 Informed search (heuristic search)

 Determine order for examining states

source: G. Luger (2005)

 order of nodes controls order of search

 closed list (aka the explored set)

source: G. Luger (2005)

Depth = length of path from root to node

 DFS uses a stack:

 BFS uses a queue:

source: G. Luger (2005)

1. open = [A-null] closed = []

...and so on until either U is found or open = []

source: G. Luger (2005)

source: G. Luger (2005)

source: G. Luger (2005)

1. open = [A-null] closed = []

→ Worksheet #1 (“Depth-First Search”)

source: G. Luger (2005)

source: G. Luger (2005)

 memory requirements high -- exponential space for states

 But both search are impractical in real applications

Compromise for DFS :

 Three possible outcomes:

 with a maximum depth before going to next level

 Repeats depth first search with gradually increasing

 Preferred search method when there is a large search

source: Russel & Norvig (2003)

source: Russel & Norvig (2003)

source: Russel & Norvig (2003)

source: Russel & Norvig (2003)

 But what if all edges/moves do not have the same cost?

 so far, all search algorithms have been uninformed (general

 But: heuristic might fail

 g(n) = cost of current path from start to node n

 h(n) = estimate of the lowest cost from n to goal

 so lower h(n) is better

source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991.

h(n) = 2 h(n) = 2 h(n) = 2

 Steepest ascent hill climbing:

if nextEval >= HEURISTIC-VALUE(currentNode)

Adapted from https://en.wikipedia.org/wiki/Hill_climbing

hill-climbing will stop, pickup&putOnTable(A)

h(n) = 2 h(n) = 2 h(n) = 2

Some parameter to represent n

source: Rich & Knight, Artificial Intelligence, McGraw-Hill College 1991. 57

 Now consider f(n) = g(n) + h*(n):