On Evolving MultiPheromone Ant Paintings

1
On Evolving Multi-Pheromone Ant Paintings

• Gary Greenfield
• University of Richmond
• CEC Evolved Art Music July, 2006

2
Outline

• Background.
• Objectives.
• Multiple-Pheromone Model.
• (Virtual) Ant Model.
• Evolutionary Framework.
• Results.
• Conclusions.

3
I. Background

• Ramos et al. (ANTS 2000) ACO for image
  processing.
• Monmarche et al. (CEC 2003) Interactive
  evolution of ant paintings.

4

• Greenfield (EvoMUSART 2005) Non-interactive ant
  paintings.

5

• Urbano (EvoMUSART 2005) Ant paintings using one
  (environmental) pheromone.

6

• Greenfield (BRIDGES 2006) Ant paintings using
  two pheromones.

7
Model Comparison

• Monmarche Ants leave color trails while seeking
  luminance.
• Greenfield (a) Ants leave color trails while
  seeking and avoiding tristimulus colors.
• Urbano Ants seek scent exuded by grid cells.
  Cells are re-colored by first ant to visit.
• Greenfield (b) Ants seek scent exuded by grid
  cells and avoid scent exuded by ants. Cells are
  re-colored by first ant to visit but may be
  subsequently re-colored.

8
II. Objectives

• Refine ants subsequent re-coloring ability
  i.e. improve ant mark making capability.
• Evolve ant paintings on the basis of a single
  trait, the relative locations, or cluster points,
  of two species of ants.

9
Cluster Point Test (Urbano Model)
10
Remarks

• Non-interactive (image) evolution is one of
  McCormacks five open problems in evolutionary
  music and art.
• Ant paintings can be considered from the point of
  view of the creativity problem
  Q Why
  should ants be able to create paintings? A
  Stigmergy - individual ants are rule-based, but
  collectively their efforts appear to be goal
  oriented and organized.

11
III. Multiple-Pheromone Model

• Each never visited grid cell emits Pc units of
  cell pheromone at each time step.
• Each ant emits Pa units of ant pheromone at each
  time step.
• E percent of each type pheromone evaporates at
  each time step.
• D percent of each type of pheromone is diffused
  to the eight neighboring grid cells at each time
  step.

12
IV. (Virtual) Ant Model

• Ant deposits background color b (white or black
  according to species) whenever it is first to
  visit a cell.
• Ant maintains current position and current
  compass heading N, NE, E, SE, S, SW, W, NW.
• Ant senses pheromone levels in each of the three
  cells in its forward field of vision.
• Ant moves to the sensed cell with maximum cell
  pheromone S, if S gt T, otherwise ant moves to the
  sensed cell with minimum ant pheromone s, AND
  leaves trail.

13
Ant Mark Making

• The trail is made by blending a time varying
  percentage of the ants foreground color f with
  the current cell.
• viz. Over L time steps an ant may diffuse and
  blend its foreground color (modulated from, say,
  f/2 to f ) thereby simulating a stroke being
  painted on a background that was initially
  re-colored white and black a la Urbano.

14
Example

• 500 ants
• 600 x 600 grid
• Two species, each initially clustered
• Time series after 500, 1000, 1500, 2000, 2500,
  and 3000 time steps

15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
V. Evolutionary Framework

• M x M grid (M 200 or 600).
• 500 Ants.
• Na ants use black for background color
• Nb ants use white for background color
• Na Nb
• Each ant has randomly generated initial offset
  (Ox,Oy) relative to its species cluster point --
  Ca (Ax, Ay) or Cb (Bx,By) -- and randomly
  generated initial direction.

22
Genetics

• Genome Ca Cb (Ax,Ay,Bx,By).
• Recombination Operator Uniform crossover.
• Mutation Operator Genetic drift.
• Population Size P 8.
• Number of Generations G 9.
• Replacement Entire population using random pairs
  formed from P/2 most fit genomes.

23
Fitness

• Ant Painting Termination Condition Time t, where
  t is the smaller of 500 times steps or the number
  of steps until all grid cells have been visited
  at least once.
• Foreground Painting Measure Let s be the number
  of times ants made foreground marks during
  completion of the painting.
• Fitness Function F(Ca Cb) st.

24
Remarks

• Fitness is minimized because the objective is to
  locate the two species colonies in such a way
  that all grid squares are visited in the least
  amount of time with the least amount of
  overpainting.
• Because of complete replacement the best
  painting may appear in any generation.

25
VI. Results

• These two evolved images contrast the aesthetic
  result we are trying to achieve with the result
  we are trying to prevent.

26

• From best initial fitness to best fitness over
  the course of an evolutionary run.

27

• The image with the lowest fitness ever
  recorded.

28

• The best image from the run that had the most
  difficulty meeting the fitness objective, and the
  best image (in an initial population) whose gene
  line went extinct.

29

• Two images with the same numeric fitness
  values (their cluster points are parallel
  translates) but with different aesthetic fitness
  values.

30
From the Design and Testing Phase to Show the
Potential of the Model
31
More Evolved Examples
32
(No Transcript)
33
VII. Conclusions

• The multiple pheromone model improved the quality
  of (our) ant paintings.
• Non-interactive evolution was able to achieve the
  design objective.
• As a theoretical point, ants could autonomously
  perform the fitness calculation themselves
  (artificial creativity implications?)

34
Future Work

• Need to validate the genetics in these kind of
  evolutionary art schemes.
• Need to better understand how to design fitness
  functions to extract desired local and global
  image characteristics.
• Need more diverse ant behaviors in the model.
• Preference testing. (a slippery slope?)

35

• Thank-you!!
• Questions??