# ppt

AI and Robotics

Dec 1, 2013 (4 years and 5 months ago)

158 views

Cellular Automata

Biologically Inspired Computing

Various credits for these slides, which have in part been adapted
from slides by: Ajit Narayanan, Rod Hunt, Marek Kopicki.

Cellular Automata

A CA is a spatial lattice of N cells, each of
which is one of
k
states

at time
t
.

Each cell follows the same simple rule for updating its state.

The cell's state
s

at time
t
+1 depends on its own state and the
states of some number of neighbouring cells at
t
.

For one
-
dimensional CAs, the neighbourhood of a cell
consists of the cell itself and
r

neighbours on either side.
Hence,
k

and
r

are the parameters of the CA.

CAs are often described as discrete dynamical systems with
the capability to
model various kinds of

natural discrete or
continuous dynamical systems

SIMPLE EXAMPLE

Suppose we are interested in understanding how a forest fire

spreads. We can do this with a CA as follows.

Start by defining a 2D grid of `cells’, e.g.:

This will be a spatial representation of our forest.

SIMPLE EXAMPLE continued

Now we define a suitable set of
states
. In this case, it makes
sense for a cell to be either
empty
,
ok_tree
, or
fire_tree

meaning:

empty: no tree here

ok_tree: there is a tree here, and it’s healthy

fire_tree: there is a tree here, and it’s on fire.

When we visualise the CA, we will use colours to represent

the states. In these cases; white, green and red seem the right

Choices.

A fairly dense forest with a couple of trees

on fire
--

maybe from lightning strikes

SIMPLE EXAMPLE continued

Next we define the
neighbourhood structure

when we run
our CA, cells will change their state under the influence of their
neighbours, so we have to define what counts as a “neighbour”.

You’ll see example neighbourhoods in a later slide, but usually
you just use a cell’s 8 immediately surrounding neighbours.
Let’s do that in this case.

Next we decide what the neighbourhood will be like at the
boundaries of the grid.

CA Rules

Now, the main thing: how do we update the
states at the next time step? We use
sensible rules.

E.g.

If a tree is not on fire, and has
n
neighbours on
fire, it catches fire next step with probabilty
n
/8.

If a tree has been on fire for 3 steps, it dies

Step 0

Step 1

Step 2

Step 3

Step 4 … and so on

CA Rules:

A small number of sensible rules, for any given suitable
application, usually leads to convincing behaviour.

Every CA rule says:

A cell in state X changes to a cell of state Y if certain
neighbourhood conditions are satisfied

What about the “tree on fire dies after three steps rule?”
This can be easily modelled with “pure” CA rules. How?

CAs are increasingly used to simulate a wide number of
complex systems, to see “what would happen if…”, and
generally investigate the effects of various strategies

Modelling HIV infection

See HIV CA demo

Rule

1

-

If

an

H

cell

has

at

least

one

I
1

neighbour,

or

if

has

at

least

2

I
2

neighbours,

then

it

becomes

I
1
.

Otherwise,

it

stays

healthy
.

Rule

2

An

I
1

cell

becomes

I
2

after

4

time

steps

(simulated

weeks)
.

(to

operate

this

the

CA

maintains

a

counter

associated

with

each

I
1

cell)
.

Rule

3

-

An

I
2

cell

becomes

D
.

Rule

4

A

D

cell

becomes

H,

with

probability

;

I
1
,

with

probability

;

otherwise,

it

remains

D

)
1
(
Infec
repl
p
p

Infec
repl
p
p
4 states: Healthy, Infected1, Infected2, Dead

A simple 1D CA to illustrate these points:

States 0 and 1:

Wraparound 2D array of 30 cells

Rules: if both neighbours are 1, become 1;

if both neighbours are 0, become 0;

otherwise, stay the same.

Synchronous update
: most CAs operate this way. Each cell’s new

state for time
t
+1 is worked out in parallel based on the situation at
t
.

Start:
101001010001101000101010010001

T=1 :
110000000001110000010100000001

T=2 :
110000000001110000001000000001

Asynchronous update
:

Sometimes applied in preference

it is arguably a more valid way
to simulate some systems. Here, at each time step, one cell is
chosen at random and updated.

Start:
101001010001101000101010010001

T=1 :
1010010
0
0001101000101010010001

T=2 :
101001000001
1
01000101010010001

Clearly if there are
n

cells, then
n
timesteps in an asynchronous CA

corresponds to the 1 timestep of a synchronous CA.

T=3 :
1
1
1001000001101000101010010001

T=4 :
1110010000011010001010100
0
0001

T=5 :
etc ...

Boundary conditions

Rules for a cell’s state transitions are usually defined in terms of

the cell’s neighbourhood. E.g. this is the Moore neighbourhood:

But what about cells on the edge?

The common approach in 2D is to

treat the CA surface as a
Toroid

This just means wraparound in

the way indicated by the

blue and green neighbourhoods

illustrated

Types of neighbourhood

Many more
neighbourhood
techniques exist
-

see
http://cell
-
auto.com

and
‘neighbourhood survey’

Classes of cellular automata (Wolfram)

Class 1: after a finite number of time steps, the CA tends to
achieve a unique state from nearly all possible starting
conditions (limit points)

Class 2: the CA creates patterns that repeat periodically or are
stable (limit cycles)

probably equivalent to a regular
grammar/finite state automaton

Class 3: from nearly all starting conditions, the CA leads to
aperiodic
-
chaotic patterns, where the statistical properties of
these patterns are almost identical (after a sufficient period of
time) to the starting patterns (self
-
similar fractal curves)

computes ‘irregular problems’

Class 4: after a finite number of steps, the CA usually dies, but
there are a few stable (periodic) patterns possible (e.g. Game of
Life)
-

Class 4 CA are believed to be capable of universal
computation

John Conway’s Game of Life

2D cellular automata system.

Each cell has 8 neighbors
-

orthogonally, 4 adjacent diagonally. This is
called the Moore Neighborhood.

Simple rules, executed at each
time step:

A live cell with 2 or 3 live neighbors survives
to the next round.

A live cell with 4 or more neighbors dies of
overpopulation.

A live cell with 1 or 0 neighbors dies of
isolation.

An empty cell with exactly 3 neighbors
becomes a live cell in the next round.

Is it alive?

http://www.bitstorm.org/gameoflife/

Compare it to the definitions…

Glider

CA are a main part of the research area “Artificial
Life”. A common definition of “life” involves that the
living organism(s) must be capable of self
-
reproduction. Langton’s “Loops” achieve that.

Characteristics

8 states, 2D Cellular automata

Needed CA grid of 100 cells

Self Reproduction into identical copy

A simple set of rules produces self
-
reproducing
“organism”

a deep connection between Life and
Computation.

Langton’s Loops

Langton’s Loop

0

Background cell state

3, 5, 6

Phases of reproduction

1

Core cell state

4

Turning arm left by 90 degrees

2

Sheath cell state

state

7

Arm extending forward cell state

Langton’s Loops

There remains debate and interest
about the `essentials of life’ issue
with CAs, but their main BIC
value is as modelling techniques.

Modelling Sharks and Fish:

Predator/Prey Relationships

Bill Madden, Nancy Ricca and Jonathan Rizzo

Research Project using Department’s 20
-
CPU Cluster

We’ve seen HIV

here are some more examples.

This project modeled a predator/prey relationship

Begins with a randomly distributed population of
fish, sharks, and empty cells in a 1000x2000 cell
grid (2 million cells)

Initially,

50% of the cells are occupied by fish

25% are occupied by sharks

25% are empty

Here’s the number 2 million

Fish: red; sharks: yellow; empty: black

Rules

A dozen or so rules describe life in each cell:

birth, longevity and death of a fish or shark

breeding of fish and sharks

over
-

and under
-
population

fish/shark interaction

Important: what happens in each cell is
determined
only

by rules that apply locally, yet
which often yield long
-
term large
-
scale patterns.

Do a LOT of computation!

Apply a dozen rules to each cell

Do this for 2 million cells in the grid

Do this for 20,000 generations

Well over a trillion calculations per run!

Do this as quickly as you can

Rules in detail: Initial Conditions

Initially cells contain fish, sharks or are empty

Empty cells = 0 (black pixel)

Fish = 1 (red pixel)

Sharks =

1 (yellow pixel)

Rules in detail: Breeding Rule

Breeding rule: if the current cell is empty

If there are >= 4 neighbors of one species, and >=
3 of them are of breeding age,

»
Fish breeding age >= 2,

»
Shark breeding age >=3,

and there are <4 of the other species:

then create a species of that type

»
+1= baby fish (age = 1 at birth)

»
-
1 = baby shark (age = |
-
1| at birth)

Breeding Rule: Before

EMPTY

Breeding Rule: After

Rules in Detail: Fish Rules

If the current cell contains a fish:

Fish live for 10 generations

If >=5 neighbors are sharks, fish dies (shark
food)

If all 8 neighbors are fish, fish dies
(overpopulation)

If a fish does not die, increment age

Rules in Detail: Shark Rules

If the current cell contains a shark:

Sharks live for 20 generations

If >=6 neighbors are sharks and fish
neighbors =0, the shark dies (starvation)

A shark has a 1/32 (.031) chance of dying
due to random causes

If a shark does not die, increment age

Shark Random Death: Before

I Sure Hope that the

random number

chosen is >.031

Shark Random Death: After

YES IT IS!!!

I LIVE

Results

Next several screens show behavior over a
span of 10,000+ generations

Spring 2005

BM

42

Generation: 0

Spring 2005

BM

43

Generation: 100

Generation: 500

Generation: 1,000

Generation: 2,000

Generation: 4,000

Generation: 8,000

Generation: 10,500

Long
-
term trends

Borders tended to ‘harden’ along vertical,
horizontal and diagonal lines

Borders of empty cells form between like
species

Clumps of fish tend to coalesce and form
convex shapes or ‘communities’

What can be discovered by
simulating very small populations

Fish can live in stable isolated communities
as small as 20
-
30

A community of less than 200 sharks tends
not to be viable

Forest Fire Model (FFM)

During each time step the system is updated according to
the rules:

Forest Fire Model is a stochastic
3
-
state cellular automaton
defined on a
d
-
dimensional lattice with
L
d

sites.

Each site is occupied by a tree, a burning tree, or is empty.

1.
empty site

tree

with the growth rate probability
p

2.
tree

burning tree

with the lightning rate probability
f
, if
no nearest neighbour is burning

3.
tree

burning tree

with the probability
1
-
g
, if at least one
nearest neighbour is burning, where
g

defines immunity.

4.
burning tree

empty site

The application

Eventually

After some time forest
reaches

in which the mean
number of growing trees
equals the mean number
of burning trees.

Modelling brain tumour growth

Kansal et al, 2000, Journal of Theoretical Biology

Incidence of primary malignant brain tumours is 8/100,000 p.a.

3D CA, modelling brain tumour growth

Shows that
Macroscopic

tumour behaviour can be predicted via
microscopic

parameters

Uses only 4 parameters

Makes predictions that match the biological reality

MRI scan showing a

tumour; the white area

Represents blood leakage

around the tumour

Kansal et al use the
Delaunay Tesselation

as their lattice

on the right

we see blackened cells representing the tumour, in a simplified 2D version

States and Rules

Not easy to glean from the
paper, but: cells are either
healthy (empty lattice site)
or tumour.

Tumour cells are either
proliferative

(they divide
cells) or not. When a
proliferative tumour cell
wants to divide, it fills a
healthy space with a new
tumour cell if it can find
one within
delta_p

of its
position. If it can’t find
one, it becomes non
-
proliferative.

1.5M lattice sites

Initial tumour is 1000 proliferative cells at centre of lattice

Result seems realistic

Very good fit to real data;

The lines are the CA model predictions of tumour radius and

volume against time

The plotted points are measurements from real cases of untreated tumours

for yourself.

See the www site for the:

Influenza CA paper

Tumour CA paper

A Traffic Simulation CA paper

Historic urban growth in the San Francisco bay area CA

Not
examinable

recommended

Next week

‘revision’ lecture

I’ll go quickly over the examinable material from

my lectures (not Patricia’s)