Searching:

Deterministic
single-agent
Andrew W. Moore
Professor
School of Computer Science
Carnegie Mellon University
www.cs.cmu.edu/~awm
[email protected]
412-268-7599

Note to other teachers and users of these slides. Andrew would be delighted if you found this source
material useful in giving your own lectures. Feel free to use these slides verbatim, or to modify them to fit
your own needs. PowerPoint originals are available. If you make use of a significant portion of these
slides in your own lecture, please include this message, or the following link to the source repository of Slide 1
Andrew’s tutorials: http://www.cs.cmu.edu/~awm/tutorials . Comments and corrections gratefully received.

Overview
• Deterministic, single-agent, search problems
• Breadth First Search
• Optimality, Completeness, Time and Space
complexity
• Search Trees
• Depth First Search
• Iterative Deepening
• Best First “Greedy” Search

Slide 2

1
 A search problem
a GOAL

b c

e
d
f

START h

p r
q

How do we get from S to G? And what’s the

smallest possible number of transitions?
Slide 3

Formalizing a search problem

A search problem has five components:
Q , S , G , succs , cost
• Q is a finite set of states.
• S ⊆ Q is a non-empty set of start states.
• G ⊆ Q is a non-empty set of goal states.
• succs : Q Æ P(Q) is a function which takes a state as
input and returns a set of states as output. succs(s)
means “the set of states you can reach from s in one
step”.
• cost : Q , Q Æ Positive Number is a function which takes
two states, s and s’, as input. It returns the one-step
cost of traveling from s to s’. The cost function is only
defined when s’ is a successor state of s.
Slide 4

2
 Our Search Problem
GOAL
a
b c

e
d f
START
h
p r
q

Q = {START, a , b , c , d , e , f , h , p , q , r , GOAL}
S = { START }
G = { GOAL }
succs(b) = { a }
succs(e) = { h , r }
succs(a) = NULL … etc.
cost(s,s’) = 1 for all transitions
Slide 5

Our Search Problem

GOAL
a
b c

e
d f
START
h
p r
q

at
Q = {START, a , b , c , d , e , f , h , p , q , r , GOAL}
S = { START } h
G = { GOAL } r e ? W is?
succs(b) = { a } e ca ke th
i
succs(e) = { h , r }
do w are l
succs(a) = NULL … etc. y s
cost(s,s’) = 1 for all transitions Wh blem
pro Slide 6

3
 Search Problems

Slide 7

More Search Problems

Scheduling

8-Queens

What next?

Slide 8

4
 More Search Problems
But there are plenty of things which we’d
Scheduling
normally call search problems that don’t fit our
rigid definition… • A search problem has five components:
• Q , S , G , succs , cost 8-Queens
• Q is a finite set of states.
• S ⊆ Q is a non-empty set of start states.
• G ⊆ Q is a non-empty set of goal states.
• succs : Q Æ P(Q) is a function which takes a
state as input and returns a set of states as
output. succs(s) means “the set of states you
can reach from s in one step”.
• cost : Q , Q Æ Positive Number is a function
which takes two states, s and s’, as input. It
What snext?
returns the one-step cost of traveling from to
Can s’. The cost function is only defined when s’ is
y ou t a successor state of s.
hink
of exam
ples
? Slide 9

Our definition excludes…

Slide 10

5
 Our definition excludes…
Hidden State
Game Chance
against
adversary

Continuum (infinite All of the above, plus

number) of states distributed team control

Slide 11

Breadth First Search

GOAL
a
b c

e
d f
START
h
p r
q

Label all states that are reachable from S in 1 step but aren’t reachable
in less than 1 step.
Then label all states that are reachable from S in 2 steps but aren’t
reachable in less than 2 steps.
Then label all states that are reachable from S in 3 steps but aren’t
reachable in less than 3 steps.
Etc… until Goal state reached.
Slide 12

6
 Breadth-first Search
a GOAL

b c

0 steps e
from start d
f

START h

p r
q

Slide 13

1 step
Breadth-first Search
from start GOAL
a

b c

0 steps e
from start d
f

START h

p r
q

Slide 14

7
 1 step
Breadth-first Search
from start GOAL
a

b c

0 steps e
from start d
f

START h

p r
q

2 steps
from start

Slide 15

1 step
Breadth-first Search
from start GOAL
a

b c

0 steps e
from start d
f

START h
3 steps
from start
p r
q

2 steps
from start

Slide 16

8
 4 steps

1 step
Breadth-first Search from start

from start GOAL

a

b c

0 steps e
from start d
f

START h
3 steps
from start
p r
q

2 steps
from start

Slide 17

Remember the path!

GOAL
a
b c

e
d f
START
h
p r
q

Also, when you label a state, record the predecessor state. This record
is called a backpointer. The history of predecessors is used to
generate the solution path, once you’ve found the goal:
“I’ve got to the goal. I see I was at f before this. And I was at r before I
was at f. And I was…
…. so solution path is S Æ e Æ r Æ f Æ G”

Slide 18

9
 4 steps

1 step
Backpointers from start

from start GOAL

a

b c

0 steps e
from start d
f

START h
3 steps
from start
p r
q

2 steps
from start

Slide 19

4 steps

1 step
Backpointers from start

from start GOAL

a

b c

0 steps e
from start d
f

START h
3 steps
from start
p r
q

2 steps
from start

Slide 20

10
 Starting Breadth First Search
For any state s that we’ve labeled, we’ll remember:
•previous(s) as the previous state on a shortest path from START state
to s.
On the kth iteration of the algorithm we’ll begin with Vk defined as the
set of those states for which the shortest path from the start costs
exactly k steps
Then, during that iteration, we’ll compute Vk+1, defined as the set of
those states for which the shortest path from the start costs exactly k+1
steps
We begin with k = 0, V0 = {START} and we’ll define, previous(START)
= NULL
Then we’ll add in things one step from the START into V1. And we’ll
keep going.
Slide 21

BFS
a GOAL

b c

e
d
f

START h

V0 p
q
r

Slide 22

11
 BFS
a GOAL

b c

e
d
f

START h

V0 p
q
r

V1

Slide 23

BFS
a GOAL

b c

e
d
f

START h

V0 p
q
r

V1
V2
Slide 24

12
 BFS
a GOAL

b c

e
d
f V3
START h

V0 p
q
r

V1
V2
Slide 25

BFS V4

a GOAL

b c

e
d
f V3
START h

V0 p
q
r

V1
V2
Slide 26

13
 Breadth First Search
V0 := S (the set of start states)
previous(START) := NIL
k := 0
while (no goal state is in Vk and Vk is not empty) do
Vk+1 := empty set
For each state s in Vk
For each state s’ in succs(s)
If s’ has not already been labeled
Set previous(s’) := s
Add s’ into Vk+1
k := k+1
If Vk is empty signal FAILURE
Else build the solution path thus: Let Si be the ith state in the shortest
path. Define Sk = GOAL, and forall i <= k, define Si-1 = previous(Si).
Slide 27

BFS V4

a GOAL

b c

e
ntly
e convenie f te).
d
a rc h s p a c V3
our se ors(sta
START Suppose y to o btain phredecess
u
allowed yo wa y to do BFS
?
fe re nt
p n you th
ink of a dif
V0 • Ca r g
avoid storin
q
e a b le to
ld you b ad to
• And wou a t w e ’d p reviously h
Vs1omething th
store? V2
Slide 28

14
 Another way: Work back
GOAL
a
b c

e
d f
START
h
p r
q

Label all states that can reach G in 1 step but can’t reach it in less than
1 step.
Label all states that can reach G in 2 steps but can’t reach it in less
than 2 steps.
Etc. … until start is reached.
“number of steps to goal” labels determine the shortest path. Don’t
need extra bookkeeping info.
Slide 29

Breadth First Details

• It is fine for there to be more than one goal state.
• It is fine for there to be more than one start state.
• This algorithm works forwards from the start. Any
algorithm which works forwards from the start is
said to be forward chaining.
• You can also work backwards from the goal. This
algorithm is very similar to Dijkstra’s algorithm.
• Any algorithm which works backwards from the
goal is said to be backward chaining.
• Backward versus forward. Which is better?
Slide 30

15
 Costs on transitions
a GOAL
2 2
b c
5
5
1 8
2 e
3 d f
9 1 9
START h
1 4 5
p 4 3 r
1 q
5

Notice that BFS finds the shortest path in terms of number of

transitions. It does not find the least-cost path.
We will quickly review an algorithm which does find the least-cost path.
On the kth iteration, for any state S, write g(s) as the least-cost path to
S in k or fewer steps.

Slide 31

Least Cost Breadth First

Vk = the set of states which can be reached in exactly k steps, and for which the least-
cost k-step path is less cost than any path of length less than k. In other words, Vk = the
set of states whose values changed on the previous iteration.
V0 := S (the set of start states)
previous(START) := NIL
g(START) = 0
k := 0
while (Vk is not empty) do
Vk+1 := empty set
For each state s in Vk
For each state s’ in succs(s)
If s’ has not already been labeled
OR if g(s) + Cost(s,s’) < g(s’)
Set previous(s’) := s
Set g(s’) := g(s) + Cost(s,s’)
Add s’ into Vk+1
k := k+1
If GOAL not labeled, exit signaling FAILURE
Else build the solution path thus: Let Sk be the kth state in the shortest path.
Define Sk = GOAL, and forall i <= k, define Si-1 = previous(Si).
Slide 32

16
 Uniform-Cost Search
• A conceptually simple BFS approach
when there are costs on transitions
• It uses priority queues

Slide 33

Priority Queue Refresher

A priority queue is a data structure
in which you can insert and
retrieve (thing, value) pairs with
the following operations:
Init-PriQueue(PQ) initializes the PQ to be empty.

Insert-PriQueue(PQ, thing, value) inserts (thing, value) into the queue.

Pop-least(PQ) returns the (thing, value) pair with the lowest

value, and removes it from the queue.

Slide 34

17
 Priority Queue Refresher
For mo
re
A priority queue is a data structure or Sedg details, see Kn
wick or uth
any boo ba
in which you can insert and k with th sically
“algorith e word
m
appeari s” prominently
retrieve (thing, value) pairs with ng in th
e title.
the following operations:
Init-PriQueue(PQ) initializes the PQ to be empty.

Insert-PriQueue(PQ, thing, value) inserts (thing, value) into the queue.

Pop-least(PQ) returns the (thing, value) pair with the lowest

value, and removes it from the queue.

Priority Queues can be Very cheap (though

not absolutely,
implemented in such a way that incredibly cheap!)
the cost of the insert and pop
operations are
O(log(number of things in priority queue)) Slide 35

Uniform-Cost Search
• A conceptually simple BFS approach when
there are costs on transitions
• It uses a priority queue

PQ = Set of states that have been

expanded or are awaiting expansion
Priority of state s = g(s) = cost of
getting to s using path implied by
backpointers.
Slide 36

18
Starting UCS
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q

PQ = { (S,0) }
Slide 37

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost state from PQ
PQ = { (S,0) } 2. Add successors
Slide 38

19
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost state from PQ
PQ = { (p,1), (d,3) , (e,9) } 2. Add successors
Slide 39

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (d,3) , (e,9) , (q,16) } 2. Add successors
Slide 40

20
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (b,4) , (e,5) , (c,11) , (q,16) } 2. Add successors
Slide 41

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
h d he
r e: s
d5 wa
START
p e n e ia
a p 4 v n
h at h g to e est-know
1
t e w e t t in
No 4 at g 3 ly b r
p 15
a liz ed th previous
e q e
• d r r th a n th Iteration: d
be t t e to e ngeleast-cost
o g et s c1.haPop
t a
way ority
w state from PQ
PQ = { (b,4) , (e,5) , (c,11)so, e(q,16) ’s pri } 2. Add successors
• and Slide 42

21
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (e,5) , (a,6) , (c,11) , (q,16) } 2. Add successors
Slide 43

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (a,6),(h,6),(c,11),(r,14),(q,16) } 2. Add successors
Slide 44

22
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (h,6),(c,11),(r,14),(q,16) } 2. Add successors
Slide 45

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (q,10), (c,11),(r,14) } 2. Add successors
Slide 46

23
 UCS Iterations
Note what happen
ed here:
way to get to p
• h found a 2new a ay
GOAL
ly 2 the best known w
than
as m or e co st
• but it w c
b
io rit y w as un ch anged 2
• and so p’s pr 1 8
5

2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (q,10), (c,11),(r,14) } 2. Add successors
Slide 47

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (c,11),(r,13) } 2. Add successors
Slide 48

24
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (r,13) } 2. Add successors
Slide 49

UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (f,18) } 2. Add successors
Slide 50

25
 UCS Iterations
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (G,23) } 2. Add successors
Slide 51

u discover
UCS
Que st ion: Iterations
Is “t erm in ate as soon as yo
criterion?
stopping
the goal” the right
a GOAL
2 2

b c
2
5
1 8
2 e
3 d
f
9 1 9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { (G,23) } 2. Add successors
Slide 52

26
UCS terminates
a GOAL
2 2

b c
2
5
1 8
the
ly once the go al is2 popped efrom
Terminate on s a shorter path.
3 eue. Elsde we may mis
qu
priorit y 9 1
f
9
START
h
1 4 5
4 3
p 15 r
q
Iteration:
1. Pop least-cost
state from PQ
PQ = { } 2. Add successors
Slide 53

Judging a search algorithm

• Completeness: is the algorithm guaranteed to find a solution
if a solution exists?
• Guaranteed to find optimal? (will it find the least cost path?)
• Algorithmic time complexity
• Space complexity (memory use)
Variables:
N number of states in the problem
B the average branching factor (the average
number of successors) (B>1)
L the length of the path from start to goal with the
shortest number of steps

How would we judge our algorithms?

Slide 54

27
 Judging a search algorithm
N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))

Slide 55

Judging a search algorithm

N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))

Slide 56

28
 Search Tree Representation
GOAL
a
b c
e
d f
START
h
p q r

thr ough
go
e r d o we h BFS?
t o rd e wit
Wha arch tre
e
the s Slide 57

Depth First Search

a GOAL
2 2
b c
5
5
1 8
2 e
3 d
f
9 1 9
START h
1 4 5
p 4 3 r
15 q

An alternative to BFS. Always expand from the most-

recently-expanded node, if it has any untried successors.
Else backup to the previous node on the current path.
Slide 58

29
 DFS in action

a GOAL

b c
START e
START d d f
START d b START h
START d b a
START d c p r
q
START d c a
START d e
START d e r
START d e r f
START d e r f c
START d e r f c a
START d e r f GOAL

Slide 59

DFS Search tree traversal

GOAL
a
b c
e
d f
Can you draw in START
h
the order in which p q r
the search-tree
nodes are visited?

Slide 60

30
 DFS Algorithm
We use a data structure we’ll call a Path to represent the , er, path from the
START to the current state.
E.G. Path P = <START, d, e, r >
Along with each node on the path, we must remember which successors we
still have available to expand. E.G. at the following point, we’ll have

P = <START (expand=e , p) ,
d (expand = NULL) ,
e (expand = h) ,
r (expand = f) >

Slide 61

DFS Algorithm
Let P = <START (expand = succs(START))>
While (P not empty and top(P) not a goal)
if expand of top(P) is empty
then
remove top(P) (“pop the stack”)
else
let s be a member of expand of top(P)
remove s from expand of top(P)
make a new item on the top of path P:
s (expand = succs(s))
This algorithm can be
If P is empty written neatly with
return FAILURE recursion, i.e. using the
program stack to
Else implement P.
return the path consisting of states in P

Slide 62

31
 Judging a search algorithm
N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
DFS Depth First N N
Search N/A N/A

Slide 63

Judging a search algorithm

N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
DFS Depth First N N
Search N/A N/A

Slide 64

32
 Judging a search algorithm
N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions
O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
Depth First
DFS**
Search
Y N O(BLMAX) O(LMAX)
Assuming Acyclic
Slide 65
Search Space

Judging a search algorithm

N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions
O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
Depth First
DFS**
Search
Y N O(BLMAX) O(LMAX)
Assuming Acyclic
Slide 66
Search Space

33
 Questions to
ponder
• How would you
prevent DFS from
looping?

• How could you

force it to give an
optimal solution?

Slide 67

Questions to Answer 1:
ponder PC-DFS (Path Checking DFS):
• How would you Don’t recurse on a state
prevent DFS from if that state is already in
looping? the current path

Answer 2:
• How could you
MEMDFS (Memoizing DFS):
force it to give an
optimal solution? Remember all states
expanded so far. Never
expand anything twice.

Slide 68

34
 Questions to Answer 1:
ponder PC-DFS (Path Checking DFS):
• How would you Don’t recurse on a state
prevent DFS from if that state is already in
looping? the current path

Answer 2:
• How could you
MEMDFS (Memoizing DFS):
force it to give an
optimal solution? Remember all states
expanded so far. Never
expand anything twice.

Slide 69

Questions to Answer 1:
ponder PC-DFS (Path Checking DFS):
• How would youFS is Don’t recurse on a state
CD
prevent DFS n Pfrom
he S?
if that state is already in
looping? ns w MDF the current path
o
i ME
as c an S
oc DF
e re ter th EM
th bet M
Are hen S? Answer 2:
• How could w DF
ns Cyou
asio an P MEMDFS (Memoizing DFS):
c
force
ere ettit
oc er tto
h give an
th sb Remember all states
Are optimal
i solution?
expanded so far. Never
expand anything twice.

Slide 70

35
 Judging a search algorithm
N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions
O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Slide 71

Judging a search algorithm

N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions
O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Slide 72

36
 Judging a search algorithm
N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions
O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Slide 73

Maze example
Imagine states are cells in a maze, you can move N, E, S, W. What
would plain DFS do, assuming it always expanded the E successor
first, then N, then W, then S?

Expansion order E, N, W, S

S
What would BFS do?
Other questions: What would PCDFS do?
What would MEMDFS do? Slide 74

37
 Two other DFS examples

G
Order: N, E, S, W?

G Order: N, E, S, W
with loops prevented

S
Slide 75

Forward DFSearch or Backward

DFSearch
If you have a predecessors() function as well
as a successors() function you can begin at
the goal and depth-first-search backwards
until you hit a start.

Why/When might this be a good idea?

Slide 76

38
 Invent An Algorithm Time!
Here’s a way to dramatically decrease costs
sometimes. Bidirectional Search. Can you
guess what this algorithm is, and why it can
be a huge cost-saver?

Slide 77

N number of states in the problem

B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Bidirection
BIBFS
BF Search
Y Y O(min(N,2BL/2)) O(min(N,2BL/2))

Slide 78

39
 N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Bidirection All trans
BIBFS
BF Search
Y
same cost O(min(N,2BL/2)) O(min(N,2BL/2))

Slide 79

Iterative Deepening
Iterative deepening is a simple algorithm which
uses DFS as a subroutine:

1. Do a DFS which only searches for paths of

length 1 or less. (DFS gives up any path of
r o h lar

length 2)
be uc u

s.
m m reg

te

2. If “1” failed, do a DFS which only searches

ta
nu be an

fs

paths of length 2 or less.

th an th
n t c ter

3. If “2” failed, do a DFS which only searches

ha s et
r t co b

paths of length 3 or less.

te ut ch

e
ea B u
gr S. e m

….and so on until success

D F an b
C

Cost is
O(b1 + b2 + b3 + b4 … + bL) = O(bL) Slide 80

40
 Maze example
Imagine states are cells in a maze, you can move N, E, S, W. What
would Iterative Deepening do, assuming it always expanded the E
successor first, then N, then W, then S?

S
Expansion order E, N, W, S

Slide 81

N number of states in the problem

B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Bidirection All trans
BIBFS
BF Search
Y
same cost O(min(N,2BL/2)) O(min(N,2BL/2))
Iterative if all
ID Deepening
Y
transitions
O(BL) O(L)
Slide 82
same cost

41
 N number of states in the problem
B the average branching factor (the average number of successors) (B>1)
L the length of the path from start to goal with the shortest number of steps
LMAX Length of longest cycle-free path from start to anywhere
Q the average size of the priority queue

Algorithm Comp Optimal Time Space

lete
Breadth First Y if all
BFS
Search transitions O(min(N,BL)) O(min(N,BL))
same cost
Least Cost
LCBFS
BFS
Y Y O(min(N,BL)) O(min(N,BL))
Uniform
UCS
Cost Search
Y Y O(log(Q) * min(N,BL)) O(min(N,BL))
PCDFS Path Check Y N O(BLMAX) O(LMAX)
DFS
MEMDFS Memoizing Y N O(min(N,BLMAX)) O(min(N,BLMAX))
DFS
Bidirection All trans
BIBFS
BF Search
Y
same cost O(min(N,2BL/2)) O(min(N,2BL/2))
Iterative if all
ID Deepening
Y
transitions
O(BL) O(L)
Slide 83
same cost

Best First “Greedy” Search

Needs a heuristic. A heuristic function maps a state onto
an estimate of the cost to the goal from that state.

Can you think of examples of heuristics?

E.G. for the 8-puzzle?

E.G. for planning a path through a maze?

Denote the heuristic by a function h(s) from states to a cost

value.
Slide 84

42
 Heuristic Search
Suppose in addition to the standard search
specification we also have a heuristic.
A heuristic function maps a state
onto an estimate of the cost to the
goal from that state.
Can you think of examples of heuristics?
• E.G. for the 8-puzzle?
• E.G. for planning a path through a maze?

Denote the heuristic by a function h(s) from states

to a cost value.
Slide 85

Euclidian Heuristic
a GOAL
2 2
h=0
h=8 c
b 2
5
1 8 h=5 h=4
h=11
2 e
3 d
f
h=8 9 1 9
h=4
START
h
1 4 h=6 5
h=12
4 3
p 15 r
q
h=11 h=6
h=9

Slide 86

43
 Euclidian Heuristic
a GOAL
2 2
h=0
h=8 c
b 2 .
alg orithm 5
1 8
yq ueueh=4
h=5
ic
eurist
h=11
t h er priorit 2 et he h
3 • Anod e , prior
ity is
t
this 9 im f
•h=8But .
1 9
h=4
START value h
1 4 h=6 5
h=12
4 3
p 15 r
q
h=11 h=6
h=9

Slide 87

Best First “Greedy” Search

Init-PriQueue(PQ)
Insert-PriQueue(PQ,START,h(START))
while (PQ is not empty and PQ does not contain a goal state)
(s , h ) := Pop-least(PQ)
foreach s’ in succs(s)
if s’ is not already in PQ and s’ never previously been visited
Insert-PriQueue(PQ,s’,h(s’))

Algorithm Comp Optimal Time Space

lete
Best First
BestFS
Search
Y N O(min(N,BLMAX)) O(min(N,BLMAX))

A few improvements to this algorithm can make things

much better. It’s a little thing we like to call: A*….
…to be continued!
Slide 88

44
 What you should know
• Thorough understanding of BFS, LCBFS,
UCS. PCDFS, MEMDFS
• Understand the concepts of whether a
search is complete, optimal, its time and
space complexity
• Understand the ideas behind iterative
deepening and bidirectional search
• Be able to discuss at cocktail parties the
pros and cons of the above searches
Slide 89

45