Artificial Ecosystems for Creative Discovery

Transcription

Artificial Ecosystems for Creative Discovery
Artificial Ecosystems
for Creative Discovery
Jon McCormack
Centre for Electronic Media Art
Monash University
Clayton 3800, Australia
www.csse.monash.edu.au/~jonmc
[email protected]
1
Computational Creative Discovery
‣ Evolutionary synthesis is creative, able to discover (for example) prokaryotes,
eukaryotes, higher multi-cellularity and language through a non-teleological
process of replication and selection.
‣ We would like to adapt the processes of evolutionary adaptation to problems
of creative discovery (either machine initiated or human-machine).
‣ The aim is to structure these artificial ecosystems in such a way that they
exhibit novel discovery in a creative context rather than a biological one.
‣ In the past, many EC algorithms have downplayed the role of environment
and interactions between biotic and abiotic components.
2
Biomorphs
Richard Dawkins, The Blind Watchmaker, 1986
3
Evolved Image
Steven Rooke, 1998
Evolved Image
Karl Sims, 1991
4
Interactive Genetic Algorithm
‣human selection bottleneck
‣genetic search is poorly balanced
‣most ‘discoveries’ depend on initial conditions
‣method doesn’t scale
generative system
user
fitness
selection
phenotypes
5
Artificial Ecosystems
‣ A new kind of Evolutionary Computing algorithm, particularly for creative
discovery
‣ No explicit fitness function: agents evolve to fit their environment
‣ Gives more consideration to the role of the environment and the interactions
between biotic and abiotic components
‣ Macro properties emerge as a simulation outcome, from the interaction of
(specified) micro components
‣ Knowledge is encoded in the environment (a different knowledge
representation scheme)
‣ Environment acts as a kind of ‘memory’
‣ Heterogeneous environments
‣ Homeostasis, mutualism, symbiosis, parasitism...
6
Essential Properties and Processes
‣ The concept of genotype and phenotype produced through enaction of the
genotype;
‣ Groups of individuals represent species, multiple species are possible;
‣ Spatial distribution and (optionally) movement of individuals;
‣ The ability of individuals to modify and change their environment (either
directly or indirectly as a result of their development within, and interaction
with, the environment);
‣ the concept of individual health as an abstract scalar measure of an
individual's success in surviving within its environment over its lifetime;
‣ the concept of an individual life-cycle, in that an individual undergoes
stages of development that may affect its properties, physical interaction
and behaviour;
‣ the concept of an environment as a physical model with consistent physical
rules on interaction and causality between the elements of the environment;
‣ An energy-metabolism resource model, which describes the process for
converting energy into resources that may be utilised by species in the
environment to perform actions (including the production of resources).
7
Example: Colourfield
•
An ecosystem of colour, operating in a one-dimensional world
!
ri = wi2 k0 + k1 log
Rt = f (Ht )
!
d||Ci ||
dt
""
+ k2
dwi
dt
histogram
resources
added based
on histogram
distribution
1
2
3
H
4
5
S
6
7
L
9 10 11 12 13
wl
wr
ws
colour weights
left, right, self
colour
age health
8
R
width
H
S
...
...
...
wg
individuals
n-3 n-2 n-1 n
cells
resources
genome
growth
weight
V
state
individual
8
Example: Resource Maximisation
600
500
400
Total
Available
Resources
300
200
100
500
1000
1500
2000
2500
3000
Iterations
9
Colourfield
Ecosystem development over 5000 time steps
(warm colour bias function)
t = 10
t = 500
t = 1000
t = 5000
10
Horizontal Stripe Painting
Patrick Heron, 1957–58, 274.3 x 154.8 cm
Tate Modern Collection, London UK
11
Example: Eden
‣ An ecosystem of sonic agents, operating in a two-dimensional world
‣ Rocks, Biomass, Agents
‣ Agents learn about and adapt to their environment, based on a modified
version of Wilson’s XCS
‣ Those agents that best fit the environment come to dominate the
population
‣ The virtual and real worlds are connected: human interest drives resource
production, giving rise to selection pressure based on maintaining human
interest
‣ Agents use changing, interesting sound to maintain their food supply, hence
ensure their survival
12
Eden architecture
sunlight
AGENT
sound-related
sound-related
XCS learning system
biomass
sensors
actuators
agent
human
interest
13
14
Causal Mechanisms: Colourfield and Eden
f(H)
f(Aa + Ab, day)
energy
source
energy
source
chroma and intensity
histogram (H)
radiant
energy
albedo
(Ab)
albedo (Aa)
growth
radiant
energy
hue, saturation, lightness
entities
animals
death
eat /
death
death
biomass
eat
people
COLOURFIELD
EDEN
15
Buy the book!
Impossible Nature: the art of Jon McCormack
Jon McCormack, Jon Bird, Annemarie Jonson, Alan Dorin.
Australian Centre for the Moving Image, 2004
www.acmi.net.au
16