# Fighting Intelligent Fires

AI and Robotics

Oct 29, 2013 (4 years and 6 months ago)

95 views

Fighting Intelligent Fires
Anthony Bonato

1

Fighting Intelligent Fires

Anthony Bonato

Ryerson University

CMS Summer Meeting

June 5, 2010

Fighting Intelligent Fires
Anthony Bonato

2

Firefighter

G

simple, undirected, connected graph

fire spreads from a vertex over discrete time
-
steps or
rounds

vertices are
burned
,
saved
, or
available

firefighter can save one vertex in each round

(
Hartnell
, 95)
introduced Firefighter

simplified model for the spread of a fire/disease/virus
in a network

Fighting Intelligent Fires
Anthony Bonato

3

Saving vertices

one
-
player game

firefighter aims to maximize the number of
saved vertices

sn(G,v)
= maximum number of saved

vertices in
G

if a fire starts at
v

Fighting Intelligent Fires
Anthony Bonato

4

Examples

sn(P
n
,v) = n
-
1
, if
v

is an end
-
vertex

= n
-
2
, else

sn(K
n
,v) = 1

(MacGillivray, P. Wang, 03)
:

sn(Q
n
,v) = n

Fighting Intelligent Fires
Anthony Bonato

5

Previous Work

(MacGillivray, P. Wang, 03), (
Messinger
, 04),
(
Devlin,Hartke
, 07), (Fogarty,07), (
Cai
, W.
Wang, 09):
finite and infinite grids: Cartesian,
strong, triangular, higher dimensions

(
Hartnell
, Li, 00), (
Cai

et al, 10):
trees

(
Finbow

et al, 10), (King, MacGillivray, 10):
algorithms and complexity

(
Cai
, W. Wang, 07)
,
(
Finbow
, P. Wang,
W. Wang, 10), (Prałat, 10):

surviving rate

(
Finbow
, MacGillivray, 10)
: survey

Fighting Intelligent Fires
Anthony Bonato

6

Complexity

(
Finbow

et al, 10)

“Is
sn(G,v) ≥ k
?”
NP
-
complete, if
G

is a tree
with maximum degree
3

Fighting Intelligent Fires
Anthony Bonato

7

Surviving rate

(
Cai
, W. Wang, 09)
surviving rate
of
G
,

ρ
(G)
= expected percentage of vertices

saved if fire starts at a random

vertex

Fighting Intelligent Fires
Anthony Bonato

8

Example: path

Fighting Intelligent Fires
Anthony Bonato

9

Results on
ρ
(G)

(
Cai
, W. Wang, 10):
ρ
(G) ≥ 1

Θ
(log n /n)
if
G

is
outerplanar

(
Finbow
, P. Wang, W. Wang, 10):

if
G

has size at most
(4/3

ε
)n
, then
ρ
(G) ≥ 6/5
ε
,
where
0 <
ε

< 5/27

(Prałat, 10):

if
G

has size at most
(15/11

ε
)n
, then
ρ
(G) ≥
1/60
ε
, where
0 <
ε

< 1/2
(
15/11

best possible
)

Fighting Intelligent Fires
Anthony Bonato

10

Intelligent fires

fire now
chooses

k

vertices to burn in each
round

Fighting Intelligent Fires
Anthony Bonato

11

k=1

x

y

a

b

K
100

burns
51

vertices…

Intelligent fires

fire now
chooses

k

vertices to burn in each
round

Fighting Intelligent Fires
Anthony Bonato

12

k=1

x

y

a

b

K
100

burns two vertices

k
-
Firefighter
(B,
Messinger
, Prałat, 10)

played similar to Firefighter, except now
fire chooses at most
k

nodes to burn

two
-
player game
: fire has strategy for
optimal burning

proposed by
(Devlin, Hartke,07)

k
-
surviving rate
ρ
(
G,k
)
defined
analogously

Fighting Intelligent Fires
Anthony Bonato

13

Bounds

Theorem
(BMP, 10)

ρ
(
G,k
) ≥ (1/(k+1))(1
-
1/n)

Theorem
(BMP,10)

ρ
(
G,k
) ≤ 1

2/n + 1/n
2

+ 1/n
2
(n
-
1)/(k+1)

Fighting Intelligent Fires
Anthony Bonato

14

Eg:
Wheels

and
prisms

Fighting Intelligent Fires
Anthony Bonato

15

Questions

what is the value of
ρ
(
G,k
)
in “typical”
sparse graphs?

expect k
-
surviving rate to be high

ρ
(
G,k
)
in infinite graphs?

what is
ρ
(
G,k
)
for the infinite random graph?

Fighting Intelligent Fires
Anthony Bonato

16

Random regular graphs

random d
-
regular graph, G(
n,d
)

d
-
regular graphs with uniform probability
distribution

pairing model
(
Bollobás,Wormald
)

G(
n,d
)

is
flammable

for all
k
: for large

n
, high probability that a sizeable part

of graph burns

Fighting Intelligent Fires
Anthony Bonato

17

Main result

Theorem
(BMP,10)
A.a.s.

(i.e. with
probability tending to
1

as
n →∞
)

ρ
(
G,k
) ≤ (1 + O(d
-
1/2
))/ k+1

→ 1/(k+1)
as
d →∞

eg: for
k=1
½

of graph!

Fighting Intelligent Fires
Anthony Bonato

18

Sketch of proof

short cycle
: length
L = log
d
-
1
log
d
-
1

n

a.a.s. most nodes not in a short cycle

wlog

focus on such nodes:
U

fire starts at
u

in
U
stages:

Fighting Intelligent Fires
Anthony Bonato

19

Stages

Stage I
: fire can spread only to
< k

nodes

up to round
t
0

=
Θ

(log k/d)
(constant)

Stage II
: no short cycles up to round

t
1
= 1/2L

burned subgraph is a tree order
(1+o(1))k t
1

Stage III
: there are short cycles, but many
nodes burning

firefighter cannot contain fire effectively

spectral bounds
(Friedman, 10)
,

expander mixing lemma
(
Alon
, Chung,88)

Fighting Intelligent Fires
Anthony Bonato

20

Numerical bounds

Fighting Intelligent Fires
Anthony Bonato

21

Infinite graphs

P

“most” vertices saved

Fighting Intelligent Fires
Anthony Bonato

22

Limiting surviving rate

ρ
(P

,k) = lim
n→∞

ρ
(
P
n
,k
) = 1

similarly,

ρ
(
K

,k
) = 1/(k+1)

how to define
ρ
(
G,k
)
for an infinite graph?

Fighting Intelligent Fires
Anthony Bonato

23

Chains

countably

infinite
G
, express as limit of
chain
C = (
G
n
: n ≥ 1),
where each
G
n

is
connected

ρ
C
(
G,k
)
= lim
n→∞

ρ
(
G
n
,k
)

real number in
[0,1],
when it exists

does not depend on
C

for paths, cliques, will

depend on
C
, in general

Fighting Intelligent Fires
Anthony Bonato

24

Fighting Intelligent Fires
Anthony Bonato

25

Theorem
(Erdős,Rényi, 63):
With probability
1
, any two graphs sampled from
G(
N
,p)
are
isomorphic.

isotype
R

unique with the
e.c. property
:

The infinite random graph

For all finite
A

B

there exists
z

Fighting Intelligent Fires
Anthony Bonato

26

Aside: cop
density

c(G)
=
cop number
of
G

if
C = (
G
n
: n ≥ 1)
is a chain of graphs with limit G,
then define the
cop density

D
C
(G
)
=
lim

c(
G
n
)/|V(
G
n
)|

(
B,Hahn,Wang
, 08):
There are chains
C

such
D
C
(R)

is any fixed real in
[0,1].

(Frankl,84):
if
G

connected, then
c(G) = o(|V(G)|)

D
C
(G) =0
for all chains
C

n→∞

Limiting surviving rate of R

Theorem
(BMP,10):
For every real
number
r

in
[1/k+1,1],
there is a chain
C

such that

ρ
C
(
R,k
) = r.

Proof ideas:

use e.c. property

many extensions lower
k
-
surviving rate

add long path to increase it

Fighting Intelligent Fires
Anthony Bonato

27

Future research

ρ
(
G,k
)
in graph classes
, products

k
-
Firefighter as a combinatorial game?

grids? …

Fighting Intelligent Fires
Anthony Bonato

28

29

Fighting Intelligent Fires
Anthony Bonato

k = ∞
,
Cartesian
grid P
7

P
7

30

Fighting Intelligent Fires
Anthony Bonato

k = ∞
,
Cartesian grid
P
7

P
7

MW Strategy

(MacGillivray, Wang, 03):
MW Strategy
:

If fire breaks out at
(
r,c
), 1≤r≤c≤n/2
, save vertices in
following order:

(r + 1, c), (r + 1, c + 1), (r + 2, c
-

1), (r + 2, c + 2),

(r + 3, c
-
2),(r + 3, c
-

3), ..., (r + c, 1), (r + c, 2c),

(r + c, 2c + 1), ..., (r + c, n)

MW strategy saves
n(n
-
r)
-
(c
-
1)(n
-
c)
vertices

MW is optimal strategy assuming fire breaks out in
columns (rows)
1,2, n
-
1, n

Fighting Intelligent Fires
Anthony Bonato

31

¼

-
conjecture

32

Fighting Intelligent Fires
Anthony Bonato

Infinite hexagonal grid

can one cop
contain

the fire?

Fighting Intelligent Fires
Anthony Bonato

33

Fighting Intelligent Fires
Anthony Bonato

34

preprints, reprints, contact
:

Anthony Bonato

Graphs at Ryerson (
G@R
)

Fighting Intelligent Fires
Anthony Bonato

35

New Book

Cops and Robbers on Graphs

with Richard Nowakowski

expected release 2011…

Fighting Intelligent Fires
Anthony Bonato

36