Chapter 3

Solving Problems by Searching

Jane Hsu
National Taiwan University

Acknowledgements: This presentation is created by Jane hsu based

on the lecture slides from The Artificial Intelligence: A Modern
Approach by Russell & Norvig, a PowerPoint version by Min-Yen Kan,
as well as various materials from the web.
Outline
Informed (Heuristic) Search Strategies

p Best-first search
n Greedy best-first search
n A* search

p Heuristic Functions
n Relaxed problems
n Pattern database

Heuristic Search Jane Hsu 2

Best-First Search
p Idea: use an evaluation function f(n) for each
node
n estimate of "desirability"
n Expand most desirable unexpanded node

p Implementation:
p Order the nodes in the fringe in decreasing order
of desirability

p Special cases:
n greedy best-first search
n A* search

Heuristic Search Jane Hsu 3

Romania with Step Costs

Heuristic Search Jane Hsu 4

Greedy Best-First Search
p Evaluation function f(n) = h(n) (heuristic)
= estimate of cost from n to goal

e.g., hSLD(n) = straight-line distance from

n to Bucharest

p Greedy best-first search expands the node

that appears to be closest to goal

Heuristic Search Jane Hsu 5

Greedy Best-First Search Example

Heuristic Search Jane Hsu 6

Greedy Best-First Search Example

Heuristic Search Jane Hsu 7

Greedy Best-First Search Example

Heuristic Search Jane Hsu 8

Greedy Best-First Search Example

Heuristic Search Jane Hsu 9

Greedy Best-First Search Example

Heuristic Search Jane Hsu 10

Greedy Best-First Search Example

Heuristic Search Jane Hsu 11

 Map of Manhattan
p How would you find a path from S to G?

52nd St
S

51st St G

50th St
10th Ave

9th Ave

8th Ave

2nd Ave
3rd Ave
7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 12

Best-First Search
p The Manhattan distance (D x+ D y) is an
estimate of the distance to the goal
n It is a heuristic function
p Best-First Search
n Order nodes in priority queue to minimize
estimated distance to the goal h(n)
p Compare w/ Dijkstra
n Order nodes in priority queue to minimize
actual distance from the start

Heuristic Search Jane Hsu 13

 Best First in Action
p How would you find a path from S to G?

52nd St
S

51st St G

50th St
10th Ave

9th Ave

8th Ave

2nd Ave
3rd Ave
7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 14

 Problem 1: Led Astray
p Eventually will expand vertex to get back
on the right track

52nd St
S G

51st St

50th St
10th Ave

9th Ave

8th Ave

2nd Ave
3rd Ave
7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 15

Problem 2: Optimality
p With Best-First Search, are you
guaranteed a shortest path is found when
n goal is first seen?
n when goal is removed from priority queue (as
with Dijkstra?)

Heuristic Search Jane Hsu 16

Sub-Optimal Solution
p No! Goal is by definition at distance 0:
will be removed from priority queue
immediately, even if a shorter path
exists! (5 blocks)

52nd St S h=2
h=5

51st St h=4 G h=1

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 17

Properties of Greedy Best-First Search
p Complete? No – can get stuck in loops,
e.g., Iasi à Neamt à Iasi à Neamt à
p Time? O(bm), but a good heuristic can give
dramatic improvement
p Space? O(bm) -- keeps all nodes in
memory
p Optimal? No

Heuristic Search Jane Hsu 18

Synergy?
p Dijkstra / Breadth First guaranteed to find
optimal solution
p Best First often visits far fewer vertices,
but may not provide optimal solution

n Can we get the best of both?

Heuristic Search Jane Hsu 19

A* Search
p Idea: avoid expanding paths that are
already expensive
p
p Evaluation function f(n) = g(n) + h(n)
p
n g(n) = cost so far to reach n
n h(n) = estimated cost from n to goal
n f(n) = estimated total cost of path through n
to goal

Heuristic Search Jane Hsu 20

A* Search Example

Heuristic Search Jane Hsu 21

A* Search Example

Heuristic Search Jane Hsu 22

A* Search Example

Heuristic Search Jane Hsu 23

A* Search Example

Heuristic Search Jane Hsu 24

A* Search Example

Heuristic Search Jane Hsu 25

A* Search Example

Heuristic Search Jane Hsu 26

A* Search Example

Heuristic Search Jane Hsu 27

A* Search Example

Heuristic Search Jane Hsu 28

A* Search Example

Heuristic Search Jane Hsu 29

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 0 5 5
9th
51st St G

50th St
9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 30

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
51st & 9th 1 4 5

50th St
9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 31

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
51st & 8th 2 3 5

50th St 50th & 9th 2 5 7

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 32

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
51st & 7th 3 2 5

50th St 50th & 9th 2 5 7

50th & 8th 3 4 7

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

Heuristic Search Jane Hsu 33

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
51st & 6th 4 1 5

50th St 50th & 9th 2 5 7

50th & 8th 3 4 7

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave 50th & 7th 4 3 7

Heuristic Search Jane Hsu 34

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
51st & 5th 5 0 5

50th St 50th & 9th 2 5 7

50th & 8th 3 4 7

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave 50th & 7th 4 3 7

Heuristic Search Jane Hsu 35

A* in Action

(5 blocks) vertex g(n) h(n) f(n)

52nd St S
52nd & 5 2 7
4th
51st St G
50th & 9th 2 5 7

50th St 50th & 8th 3 4 7

50th & 7th 4 3 7

9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave

DONE!
Heuristic Search Jane Hsu 36
What Would Dijkstra Have Done?
(5 blocks)
52nd St S

51st St G

50th St

49th St

48th St

47th St
9th Ave

8th Ave

7th Ave

6th Ave

5th Ave

4th Ave
Heuristic Search Jane Hsu 37
Admissible Heuristics
p A heuristic h(n) is admissible if for every node n,
h(n) ≤ h*(n), where h*(n) is the true cost to
reach the goal state from n.

p An admissible heuristic never overestimates the

cost to reach the goal, i.e., it is optimistic
p Example: hSLD(n) (never overestimates the actual
road distance)
p Theorem: If h(n) is admissible, A* using TREE-
SEARCH is optimal.

Heuristic Search Jane Hsu 38

Optimality of A* (proof)
p Suppose some suboptimal goal G2 has been generated and is in
the fringe. Let n be an unexpanded node in the fringe such that n
is on a shortest path to an optimal goal G.
p

p f(G2) = g(G2) since h(G2) = 0

p g(G2) > g(G) since G2 is suboptimal
p f(G) = g(G) since h(G) = 0
p f(G2) > f(G) from above
Heuristic Search Jane Hsu 39
Optimality of A* (proof)
p Suppose some suboptimal goal G2 has been generated and is in
the fringe. Let n be an unexpanded node in the fringe such that n
is on a shortest path to an optimal goal G.
p

p f(G2) > f(G) from above

p h(n) ≤ h* (n) since h is admissible
p g(n) + h(n) ≤ g(n) + h*(n)
p f(n) ≤ f(G)
Hence f(G2) > f(n), and A* will never select G2 for expansion
Heuristic Search Jane Hsu 40
Optimality of A*
p A* expands nodes in order of increasing f value
n Gradually adds "f-contours" of nodes
n Contour i has all nodes with f=fi, where fi < fi+1
n All nodes with f(n) > C* (optimal solution) are pruned.

Heuristic Search Jane Hsu 41

Properties of A*
p Complete? Yes (unless there are infinitely
many nodes with f ≤ f(G) )
p
p Time? Exponential
p
p Space? Keeps all nodes in memory
p
p Optimal? Yes

Heuristic Search Jane Hsu 42

Properties of A*
p Complete? Yes – assume finitely many nodes with f(n) ≤
f(G)
p Optimal? Yes
p Efficient?
n A* is optimally efficient for any given heuristic function. i.e. no
other optimal algorithm is guaranteed to expand fewer nodes
than A*.
p Time? Exponential
p Space? Exponential (all nodes in memory)

p Theorem: The search space of A* grows exponentially

unless the error in the heuristic function grows no faster
than the logarithm of the actual path cost. [p.101]

Heuristic Search Jane Hsu 43

 Question?

Heuristic Search Jane Hsu 44

8-Puzzle
p Convert the initial configuration into the
goal configuration by moving the tiles.
p Legal moves:
n Move any tile to adjacent empty square

Heuristic Search Jane Hsu 45

Admissible Heuristics
E.g., for the 8-puzzle:

p h1(n) = number of misplaced tiles

p h2(n) = total Manhattan distance
(i.e., no. of squares from desired location of each tile)

p h1(S) = ?
p h2(S) = ?
p
Heuristic Search Jane Hsu 46
Admissible Heuristics
E.g., for the 8-puzzle:

p h1(n) = number of misplaced tiles

p h2(n) = total Manhattan distance
(i.e., no. of squares from desired location of each tile)

p h1(S) = ? 8
p h2(S) = ? 3+1+2+2+2+3+3+2 = 18
Heuristic Search Jane Hsu 47
Dominance
p If h2(n) ≥ h1(n) for all n (both admissible)
p then h2 dominates h1
p h2 is better for search
p

p Typical search costs (average number of nodes expanded):

p

p d=12 IDS = 3,644,035 nodes

A*(h1) = 227 nodes
A*(h2) = 73 nodes
p d=24 IDS = too many nodes
A*(h1) = 39,135 nodes
A*(h2) = 1,641 nodes
p

Heuristic Search Jane Hsu 48

 7 2 3
4 1 6
Importance of Heuristics 8 5
p h1 = number of tiles in the wrong place
p h2 = sum of distances of tiles from correct location

D IDS A*(h1) A*(h2)

2 10 6 6
4 112 13 12
6 680 20 18
8 6384 39 25
10 47127 93 39
12 364404 227 73
14 3473941 539 113
18 3056 363
24
Heuristic Search Jane39135
Hsu 1641 49
Inventing Heuristic Functions
p Relaxed problems: the cost of an exact solution to a relaxed
problem is often a good heuristic for the original problem. e.g.
n A if B and C
n A if B
n A if C
n A
p Composite heuristics
n h(n) = max (h1(n),…,hm(n))
p Weighted evaluation function
n fw(n) = (1-w)g(n) + w h(n)
p Learn the coefficients for features of a state
n h(n) = c1x1(n) + … + ckxk(n)
p Statistical information
p Search cost
n Good heuristics should be efficiently computable.

Heuristic Search Jane Hsu 50

Relaxed Problems
p A problem with fewer restrictions on the actions is called a
relaxed problem
p
p The cost of an optimal solution to a relaxed problem is an
admissible heuristic for the original problem
p Original: move any tile to adjacent empty square
p Relaxed problems:
n Move from A to B, if A is adjacent to B.
p Manhattan distance
n Move from A to B, if B is empty.
p Gaschnig’s heuristic (1979)
n Move from A to B. e.g. a tile can be moved to anywhere
p Misplaced tiles

Heuristic Search Jane Hsu 51

Composite Heuristics
p Given a collection of admissible heuristics
h1… hn for a problem, and none of them
dominates any of the others, which should
we choose?

Heuristic Search Jane Hsu 52

Finding Optimal Solutions
p Input: A random solvable initial state
p Output: A shortest sequence of moves
that maps the initial state to the goal state

p Generalized sliding-tile puzzle is NP

Complete (Ratner and Warmuth, 1986)
n People can’t find optimal solutions.
n Progress measured by size of problems that
can be solved optimally.

Heuristic Search Jane Hsu 53

Fifteen Puzzle
p Invented by Sam Loyd in 1870s
p “...engaged the attention of nine
1 2 3
out of ten persons of both 4 5 6 7
sexes and of all ages and
conditions of the community.” 8 9 10 11
p $1000 prize to swap positions of 12131415
two tiles

Ø Adapted from “Recent Progress in the

Design and Analysis of Admissible
Heuristic Functions” by R. Korf.

Heuristic Search Jane Hsu 54

Linear Conflict Heuristic
3 1 1 3

p Hansson, Mayer, and Yung, 1991

p Given two tiles in their goal row, but reversed in
position, additional vertical moves can be added
to Manhattan distance.
p Still not accurate enough to solve 24-Puzzle
p We can generalize this idea further.
Heuristic Search Jane Hsu 55
Sub-Problems
p Admissible heuristics can also be derived
from the solution cost of a sub-problem of
a given problem.

Heuristic Search Jane Hsu 56

Pattern Database Heuristics
p A pattern database is a complete set of
such positions, with associated number of
moves.
n e.g. a 8-tile pattern database for the Fifteen
Puzzle contains 519 million entries.
p On 15 puzzle, IDA* with pattern database
heuristics is about 10 times faster than
with Manhattan distance
n Culberson and Schaeffer, 1996
p Pattern databases can also be applied to
Rubik’s Cube.
Heuristic Search Jane Hsu 57
Additive Databases
5 10 14 7 1 2 3
8 3 6 1 4 5 6 7
15 12 9 8 9 10 11
2 11 4 13 12 13 14 15
p The 7-tile database contains 58 million entries.
n 20 moves needed to solve red tiles
p The 8-tile database contains 519 million entries.
n 25 moves needed to solve blue tiles
p Overall heuristic is 20+25=45 moves
Heuristic Search Jane Hsu 58
Rubik’s Cube
p Invented in 1974 by
Erno Rubik of Hungary
p Over 100 million sold
worldwide
p Most famous
combinatorial puzzle
ever

Heuristic Search Jane Hsu 59