SIMULA

TING ARTIST AND CRITIC DYNAMICS

An Agent-based Application of an Evolutionary Art System

Gary Greenﬁeld

Department of Mathematics and Computer Science, University of Richmond, Richmond, Virginia, U.S.A.

Penousal Machado

Department of Informatics Engineering, University of Coimbra, 3030 Coimbra, Portugal

Keywords:

Evolutionary art, Quantifying aesthetics, Agent-based simulation.

Abstract:

We describe an agent based artist-critic simulation. Artist agents use a swarm based evolutionary art system

to evolve images that try to match their preferences. Preferred images are submitted to critic agents who then

decide, accordingly to their own criteria, which images should be displayed in a public gallery. The purpose

of our model is to enable the implementation of a variety of behavioral policies which result in different

dynamics. A reward system determines the impact of each critic and the success of each artist, which in turn

leads to behavioral and preference changes. The experimental results indicate the emergence of novel styles

and trends, artist-critic cooperation, and niche exploitation.

1 INTRODUCTION

We describe our implementation of an artists and crit-

ics simulation. The role of artists is to produce art.

The role of critics is to decide what should be exhib-

ited in a public gallery. Artists are software agents

that have recourse to a generative art system to evolve

images and decide whether or not to show them to

critics. Critics are software agents that critique art-

works with reference to examples of “masters” that

they are partial to and previous images they have been

asked to critique.

We wish to make clear at the outset that we ex-

pect every reader to disagree with our assumptions

about how a [public] art gallery actually functions.

This misses the point of our work. Our objective is to

establish a framework for implementing a model sup-

porting constraints and assumptions the reader might

wish to impose, and to do so in such a way that the dy-

namics resulting from the simulation can be archived

and analyzed.

2 BACKGROUND

The impetus for this investigation is three-fold. First,

we reconsider the approach of (Saunders and Gero,

2002) which made use of agents sharing a common

evolutionary art system in order to study artiﬁcial cre-

ativity via the dynamics of artists and critics. Sec-

ond, we invoke a grayscale swarm based evolutionary

art system (Aupetit et al., 2003). Third, we quantify

aesthetics by using image comparison techniques that

rely upon extracting information theoretic measure-

ments from the images under consideration (see, for

example , (Machado and Cardoso, 1998) (Machado

et al., 2007)). Some of our previous work in evolu-

tionary art involving co-evolving agent-critic systems

(e.g. (Greenﬁeld, 2007) (Machado et al., 2007)) pro-

vides further inspiration for this research.

Saunders and Gero. (Saunders and Gero, 2002)

offer a conceptual framework for studying artiﬁcial

creativity supported by an implementation where a

community of software agents, each possessing a

Sims’ style image generating system (Sims, 1991),

attempts to generate a “cultural repository” of shared

aesthetic imagery. Their agents make use of neural

nets and self organizing maps. The alternative frame-

work we present is informed in part by the follow-

ing observations we made about their work: i) an

agent is simultaneously an artist and a critic, ii) agents

share genomes, iii) feature extraction and image eval-

uation use a neural net to evaluate a downsampled

32 ×32 = 1024 bit ﬁltered and thresholded image, iv)

the policy for an image to be added to the cultural

190

Greenﬁeld G. and Machado P. (2009).

SIMULATING ARTIST AND CRITIC DYNAMICS - An Agent-based Application of an Evolutionary Art System.

In Proceedings of the International Joint Conference on Computational Intelligence, pages 189-196

DOI: 10.5220/0002285701890196

 SciTePress

repository is that one agent decides an image is inter-

esting and better than its own work.

Swarm based Generative Art. Swarm art refers

to a generative technique which uses large numbers

of cooperative drawing agents to produce images in

a style that is often reminiscent of Jackson Pollack.

The term is perhaps attributable to Moura (Moura and

Ramos, 2002) who, arguably, along with Jacob (Jacob

et al., 2007), has done the most to popularize the art

form. We are motivated by the formulation of swarm

paintings as ant colony paintings in the manner of

(Aupetit et al., 2003) (Greenﬁeld, 2005b) (Greenﬁeld,

2006) and (Urbano, 2005). Since we are interested

in difﬁcult questions pertaining to aesthetic merit and

artistic style, to simplify matters, we work solely in

grayscale. For a discussion of color considerations

see (Greenﬁeld, 2009).

Image Evaluation and Aesthetics. Methods for

automated image evaluation on the basis of aesthet-

ics are eclectic and varied; the terminology itself

is problematic (Greenﬁeld, 2005c). In the domain

of evolutionary art systems, the use of neural nets

traces back to (Baluja et al., 1994) and is also promi-

nent in (Machado and Cardoso, 1998) (Machado

et al., 2007). Greenﬁeld uses geometric measure-

ments induced by image color organization (Green-

ﬁeld, 2002). Noteworthy evaluation schemes based

on theories of aesthetics include (Greenﬁeld, 2005a)

(Burns, 2006), (Ross et al., 2006), (Jacobsen, 2006),

and (Schmidhuber, 2007). Most of these schemes

make it difﬁcult to account for factors such as culture,

trends, or history which are commonly recognized as

integral to aesthetics.

3 OVERVIEW OF OUR SYSTEM

To set the stage for the details to follow, it is ﬁrst

necessary to provide a brief overview of the agent

based artist and critic simulation discussed in this pa-

per. There are N

artist agents and N

critic agents

prioritized by organizing them into a pyramid with

tiers. The larger the number of artists and crit-

ics, the more difﬁcult it becomes to analyze simula-

tion results, therefore, in this paper we chose as a rea-

sonable compromise N

= 4 artists and N

= 7 crit-

ics organized into N

= 3 tiers where the top tier con-

sists of one critic, the second tier consists of two crit-

ics, and the third tier consists of four critics. At all

times, each artist agent has one target image from a

database of images painted by “masters” while each

critic agent has two images from this database. All

artist agents make use of the same swarm based evo-

lutionary art system. Artist agents access this system

to evolve small populations of art works from which

to select images to submit to the critics. Simulation

proceeds by rounds. At the start of each round every

artist evolves its image population for a ﬁxed number

of generations and then is invited to submit one image

to the critics. Image populations are evolved and eval-

uated by artist agents according to differing criteria. If

an artist agent decides to submit an image, that image

is evaluated by N

critics selected in such a way that

one critic from each tier is represented. If at least one

critic accepts the image, then that image replaces the

oldest image in a public gallery. Artist agents are pro-

vided with the evaluation results of each image that

they submit to the critics. Critic agents receive im-

pact points for accepting images. At the conclusion

of each round, both artist agents and critic agents are

given the opportunity to modify their strategies and

behaviors based on the results, and the critic tier struc-

ture is updated on the basis of impact points. The

policies that govern the various behaviors of artists

and critics are presented as they arise during the more

comprehensive description of our system that follows.

4 THE IMAGES

In this section we discuss how we manage images in

our artist and critic simulation. All our images are

256 × 256 pixel grayscale images. A distinguishing

feature of our system is that it uses a database of 25

images consisting of ﬁve images selected from each

of ﬁve “masters” — Dali, Mondrian, Monet, Picasso,

and van Gogh. This database was culled from on-

line archives and subsequently resized and converted

to grayscale. Examples of four representative images

from this database are shown in Figure 1. For sys-

tem testing we also make use of a database of 13

grayscale “noise” images. All other images in our

system are generated from genomes that are fed to

a swarm-based generative algorithm. Genomes and

pixel maps of images generated from genomes are the

property of artist agents. They are never shared. Fea-

ture vectors extracted from those images are shared.

4.1 Feature Vectors

An image is viewed as organized into six regions —

the entire image, the upper left quadrant, the upper

right quadrant, the lower right quadrant, the lower

left quadrant, and the cental “quadrant” whose area

is one-fourth the size of the entire image. For each

region, along with the mean and standard deviation of

SIMULATING ARTIST AND CRITIC DYNAMICS - An Agent-based Application of an Evolutionary Art System

191

Figure 1: Four images from the “masters” database. Clock-

wise from top left: Wheat Field with Cypresses at the Haute

Galline Near Eygalieres by Van Gogh, The Persistence of

Memory by Dali, Water Lilies by Monet, Tete de Femme

by Picasso.

the luminance values of its pixels, its complexity is

estimated. This yields 6 × 3 = 18 components which

make up the feature vector.

4.2 Complexity Measure of a Region

Our approach to measuring the complexity of a re-

gion is founded upon two assumptions: i) region com-

plexity is an aesthetically relevant characteristic, and

ii) perceived complexity can be estimated using frac-

tal image compression. To estimate the perceived

complexity of a region, we use quad-tree fractal im-

age compression (Fisher, 1995). Our rationale is that

complex images are harder to compress, resulting in

larger ﬁles than simple images, therefore we assume

that the compression ratio is negatively correlated

with image complexity.

The fractal image compression scheme we use for

each region is lossy so there will be compression er-

ror, i.e., the compressed image won’t exactly match

the original. All other factors being equal, complex

images will tend toward higher compression errors

and simple images will tend toward lower compres-

sion errors. Thus, the compression error is positively

correlated with image complexity.

We estimate image complexity of region i from

image I using the following formula (Machado and

Cardoso, 1998):

Complexity(i) = RMSE(i,FC(i))×

s(FC(i))

s(i)

(1)

where RMSE stands for the root mean square error,

FC is the fractal compression transformation, and s is

the ﬁle size function.

Fractal image compression can provide a compact

encoding to images with high apparent visual com-

plexity, in fact this characteristic of is the basis for the

aesthetic judgement scheme described in (Machado

and Cardoso, 1998). Although the estimate used has

its shortcomings, previous work (Machado and Car-

doso, 2002) indicates that images with similar com-

plexity estimates tend to have similar visual complex-

ity. In the experiments described here we use the set

of parameters for our compression scheme given in

Table 1. Note that letting the minimum partition level

be 3 implies that the selected region is always ﬁrst

partitioned into 64 blocks. Subsequently, at each step,

for each block, if one ﬁnds a transformation that gives

a good enough pixel by pixel match, then that trans-

formation is stored and the image block isn’t further

partitioned. (Here, pixel by pixel match is with re-

spect to the usual 0 to 255 grayscale interval encod-

ing.) If the pixel by pixel match error is more than 8

for at least one of the pixels of the block in the par-

tition, that image block is further partitioned into 4

sub-blocks, the level increases, and the process is re-

peated. Since the maximum partition level is 6, when

that level is reached the best transformation found is

stored even if the pixel by pixel match error for the

block exceeds 8.

Table 1: Image Compression Parameters.

Compression scheme Quadratic tree fractal

Image size 256 × 256 pixels

Minimum partition level 3

Maximum partition level 6

Maximum error per pixel 8

4.3 Comparing Two Images

Given an image I with feature vector X, and a re-

gion i of I, we let c

(i) be the complexity component

and m

(i) be the mean luminance component from

X of region i. We deﬁne an image comparison met-

ric C(A,B) for comparing two feature vectors A and

B by adding a complexity term together with a mean

luminance term as follows. Our complexity term is a

weighted, relative absolute difference of region com-

plexity terms obtained by summing over regions:

100|c

(i) − c

(i)|/max(c

(i),c

(i)). (2)

For our luminance term, we take into account both the

ordering of the mean luminances of regions within the

two vectors and the mean luminances between corre-

sponding regions of the two vectors. More precisely,

IJCCI 2009 - International Joint Conference on Computational Intelligence

192

we let r

(i) denote the rank of the mean luminance

of region i within its feature vector X , and obtain by

summing over regions:

10|r

(i) − r

(i)| + Σ

(i) − m

(i)|/5. (3)

Note that we have a perfect match between A and B

when the comparison value C(A, B) is zero.

5 THE GENERATIVE SYSTEM

Our generative system is a swarm system modeled af-

ter a colony of ants. It most closely follows (Green-

ﬁeld, 2006). Each virtual ant, or organism, is tracked

by maintaining a location and a compass heading for

it. At each time step, each organism senses the three

cells immediately in front of it in order to ascer-

tain their color luminances and their concentrations

of each of two organism produced pheromones. Each

organism then deposits color in the cell it currently

occupies and advances between 1 and 4 cells in a di-

rection determined by one of these three sensed cells.

5.1 Swarm Genomes

Each swarm painting is generated from a swarm

genome. A swarm genome has four global attributes:

a pseudorandom number generator seed value, the

number of organisms, the time to execute the painting

(viz. number of execution cycles), and the organism

footprint size. Since organism behavior is stochas-

tic, to make swarm paintings reproducible each paint-

ing must be associated with an integer seed to initial-

ize the pseudorandom number generator. The most

signiﬁcant attribute is footprint size. Consistent with

an organism’s current heading, when it is depositing

color, footprint size determines the stroke width.

For each swarm painting, all its organisms are

assumed to be of the same species. Thus a swarm

genome must also include the following species at-

tributes: the number of color scents to recognize or

deposit (here six), the various scent detection thresh-

olds, the probability of following a color scent if one

is detected, the probability of following (or avoiding)

a virtual scent if one is detected, and the probabilities

for which direction (forward, left, or right) to follow

when either no scent is detected or scent is being ig-

nored. In support of color scents, a swarm genome

also includes a color scheme for the painting; here a

palette consisting of six shades of gray.

Which species attributes will actually come in to

play is determined by the caste an individual organ-

ism belongs to. There are three castes representing

three different behaviors. The explorer caste tries to

ﬁnd unvisited cells. The color caste is sensitive to

detection of either the color it tries to deposit or the

color it is trying to ﬁnd i.e., if the deposited color is

sensed then the ant tries to avoid it, but if the color it

seeks is sensed then (almost always) the ant follows

it. Similarly, the pheromone caste is sensitive to the

presence of virtual scent: one of the scents is an at-

tracting scent, the other is a repelling scent.

Finally, a swarm genome also includes attributes

for each of the individual organisms. Organisms are

differentiated by caste, color to deposit, color to fol-

low, and number of cells traversed at each time step.

Crossover and mutation operators needed to breed

swarm painting genomes are consistent with those de-

scribed in (Greenﬁeld, 2006).

5.2 Generating a Swarm Painting

The number of organisms per swarm painting is re-

stricted to be between 20 and 80, in multiples of ﬁve.

Because we wish to evolve swarm paintings, it is nec-

essary to be consistent in the initial placement of the

organisms on the canvas. Rather than initially placing

all of the organisms at pseudorandom locations on the

grid, we cluster them into ﬁve equally sized groups.

One group is placed at the center of our canvas, while

the other four groups are placed at the centers of each

of the four quadrants. Further, we use identical initial

heading orientations for organisms within clusters.

5.3 Behavior Statistics

When the sense-decide-color-move cycle is per-

formed for each organism according to the number

of times determined by the time limit parameter in

the swarm genome, ﬁve statistics B

,. ..,B

are col-

lected for future use. Respectively, they are the num-

ber of instances where an organism: i) visited a previ-

ously unvisited cell, ii) was in “wandering” mode, iii)

pursued (or avoided) color, iv) pursued (or avoided)

scent, v) blended the color it deposited. Regarding

this last statistic, there is some duplication of effort. If

color is being deposited in a previously unvisited cell,

then it replaces the background color (black), other-

wise it blends with the existing color.

5.4 System Capabilities

From our cursory description of the generative sys-

tem, in light of master images such as those in Fig-

ure 1 that the artists and critics are assigned, it should

be clear that the generative system will never be able

to approximate such images closely! This is our inten-

tion. We treat images in the database done by “mas-

SIMULATING ARTIST AND CRITIC DYNAMICS - An Agent-based Application of an Evolutionary Art System

193

ters” either as if they were created by a process that

has been lost or forgotten, or as if they were retrieved

from an alien extraterrestrial site. Either way, they

have now become the inspiration for artists and crit-

ics alike.

6 THE ARTISTS

Each of our four artists has a population of swarm

genomes. Genome population size is 10, with 6

genomes culled after each generation and the remain-

ing 4 used to repopulate via standard evolutionary art

methods. In all the experiments run here, when an

artist agent invokes the generative system, its genome

population is evolved for 6 generations according to

the criteria it provides before examining the results.

At all times, each artist has a target image from the

database of masters, and a binary coefﬁcient vec-

tor (c

,. ..,c

) for use with the behavior statistics

,. ..,B

that are gathered during the image gener-

ation procedure.

Image Fitness Calculation. When an artist calcu-

lates image ﬁtness it compares its current target image

feature vector T with the feature vector X extracted

from the image under consideration I and obtains ﬁt-

ness value F by letting

F = C(T,X) + Σ

, (4)

where the w

’s are scaling weights that are constant

for the simulation.

Submitting an Image to the Critics. Our policy is:

When queried, an artist submits an image to the crit-

ics by submitting its feature vector if and only if its

generative system has produced a more ﬁt image than

the last time it was queried. In response, the simula-

tion provides the ranks (between 0 and 9) that were

assigned by the three critics — the tier 1 critic, a tier

2 critic, and a tier 3 critic — that were selected to cri-

tique it, and the decision about whether the image was

accepted for the gallery or not.

Changing an Artist’s Target. The policy we use to

change an artist’s target is: If the number of consecu-

tive submissions rejected reaches the threshold value

ATAR then it is time to change one’s target!

Changing an Artist’s Behavior. Because of the

way artist’s calculate image ﬁtness, artists can choose

not to be slaves to the pursuit of trying to match their

target image’s characteristics. This means they can

pursue their own unique style, or further develop suc-

cessful styles, or change the evolutionary pressure

currently being exerted by a target that is not produc-

ing gallery acceptances. Artist behavior is changed

by ﬂipping one or more of the bits in the binary co-

efﬁcient vector (c

,. ..,c

). This is implemented by

making the probability of each bit being ﬂipped 0.2.

The policy we use to change an artist’s behavior is:

Within the last ALAG rounds, if either no submission

has been made, or every submission has been rejected,

then it is time to change one’s behavior!

7 THE CRITICS

As previously stated, we use seven critics. At all

times, each critic has two target images from the

database of masters. In addition it has a rank-ordered

list of its preferences, or favorites, resulting from the

top ten evaluations it has given to all the images it has

critiqued so far.

Critiquing an Image. This favorites list makes the

actual act of critiquing an image simple. With tar-

get feature vectors T

and T

, given the feature vector

X from an image I to critique, our policy is: Take

the smaller of C(T

,X ) and C(T

,X ), insert I in the

favorites list by replacing the value of the highest

ranked image in the list of favorites with that value,

sort the favorites list in increasing order, and returns

the rank that I receives as a result.

Managing the Preferences List. During initializa-

tion we instantiate a “random artist” and generate an

initial population of images. We use that set of fea-

ture vectors to initialize all the critic preference lists.

Because critics have different targets, the way they

rank these images is different. Thus critic diversity is

present right from the start. In our model, critic pref-

erences are not ﬁxed. Their tastes change over time.

The way we chose to implement this notion was to age

each preference list by artiﬁcially inﬂating the rank-

ing of the oldest critiqued image after every 3 rounds.

This causes the next image the critic is asked to cri-

tique to force this image to be dropped from the list.

Changing a Critic’s Target. If a change in target is

triggered, one of the two target images the critic cur-

rently has is replaced by an image chosen at random

from the masters database. Note that as a side effect,

some of the images in the preferences list must be re-

evaluated — in particular, those that were present be-

cause their comparison values were based on an im-

IJCCI 2009 - International Joint Conference on Computational Intelligence

194

age that is now gone — and the preferences list must

be re-ranked. The policy we use for triggering a critic

target change is: If a critic was unsuccessful in ﬁnd-

ing an image to recommend for the gallery in the last

CTAR rounds it is time to change one of its targets!

8 THE SIMULATION

The overarching simulator for our artists and critics

is straightforward. After initialization, for each of N

rounds the artists are queried and invited to make sub-

missions. If a submission is made, the simulator se-

lects and asks three critics to evaluate it. If one or

more decide to accept it to the gallery, then the simu-

lator handles all the record keeping as well as enforc-

ing the various policies that apply to artists and critics

that we have previously described.

Managing the Gallery. The gallery of eight im-

ages is initialized by selecting image from the mas-

ters database. As artist generated swarm images are

accepted the oldest image in the gallery is replaced.

Selecting the Critics. Each critic has an impact rat-

ing. This rating increases when images are accepted

to the gallery by the critic. At the start of each round

all critics are sorted on the basis of this rating to de-

termine the tier 1 critic, the two tier 2 critics, and the

remaining four tier 3 critics. When a submission is re-

ceived, one critic from each tier is selected and asked

to evaluate the image. Our policy is: An image is con-

sidered to be accepted by the tier 1 critic if it places in

the top three of its preferences list, by the tier 2 critic

if it places in the top two of its preferences list; and

by the tier 3 critic if it places at the top of its pref-

erences list. For breaking acceptance ties our policy

is: In case of a tie the lower ranking critic gets credit

for having accepted the image. This last policy has

profound implications with respect to the dynamics

of impact ratings.

Assessing Artists and Critics. For artists, we keep

a running tally of the number of submissions and the

number of acceptances. For critics, in addition to

keeping a running tally of the number of critiques and

the number of acceptances, as previously mentioned,

we maintain an impact rating. We add a further twist

to the calculation of the impact rating by rewarding

critics for introducing novelty into the gallery as fol-

lows. The critic’s impact rating increases by 1 if it

accepts a submission to the gallery, but by 2 if, in

addition, the submission is sufﬁciently different from

the images currently in the gallery. Here “different” is

determined by considering the region that consists of

the entire image and checking to see if either its com-

plexity or average luminance is more than one stan-

dard deviation away from either of those quantities

averaged across the entire gallery.

9 SIMULATION RESULTS

To examine the dynamics resulting from our model,

we made several runs of our simulation lasting N = 50

submission-evaluation rounds using as default set-

tings ATAR = 5, ALAG = 3 and CTAR = 10. Fig-

ure 2 shows image diversity from one of these runs.

Figure 3 shows the style development of an artist that

had 8 of 29 submissions accepted during this same

run. The graphs in Figure 4 show the rank ordering

of artists on the basis of number of acceptances, and

of critics on the basis of impact rating, as well as the

cumulative totals of these quantities by rounds during

the course of the run. Note that by round #25 critic

rankings have stabilized. This is consistent with most

of the other experiments we performed. Under our

policies it is difﬁcult for lower ranked critics to dis-

place higher ranked ones. Consequently, as a run pro-

gresses there is tendency for critic ranking to stabilize

and tier changes to become less frequent. Although

tier changes are uncommon, they still do sometimes

occur, and in such cases may even trigger dramatic

changes in both critic and artist rankings as well as in

the type of imagery accepted for the gallery.

10 FURTHER EXPERIMENTS

Our artist-critic simulation has several parameters and

many features. To further examine its capabilities

and help ascertain its limitations we performed sev-

eral tests. These tests included runs with:

• N = 30, ATAR = 10, and ALAG = 1,3, 5, and 7;

• N = 30, and ATAR = 5, 10, and 15;

• N = 30, and CTAR = 5,10, and 15;

• N = 30, where one master image is selected and

assigned to be the same target for all artists;

• N = 30, where one master is ﬁrst selected and then

each critic is assigned two (not necessarily dis-

tinct) target images from the ﬁve that are available

in the master’s database by that master;

• N = 20, where one artist functions as a “ran-

dom artist” by submitting its highest ranked image

SIMULATING ARTIST AND CRITIC DYNAMICS - An Agent-based Application of an Evolutionary Art System

195

Figure 2: Artist diversity as evidenced by an accepted sub-

missions of each of the artists during rounds #7 through #10

of an N = 50 round run.

Figure 3: Artist style development is observed by consider-

ing an artist’s accepted images during the course of a run.

For the virtual artist chosen here, its acceptances are shown

(clockwise from top left) for rounds #3, #19, #23 and #46

of an N = 50 round run. Its accepted image from round #7

is shown at the top left in the previous ﬁgure.

from a random population each round (Reassur-

ingly, no image from this artist was ever accepted

by the critics beyond round 5);

• N = 20, where the master’s database is replaced

by “noise” images, the number of critics is re-

duced to 3, one master image is chosen, and both

targets for all critics are assigned to be this image

(This tests how successful the generative system

0

1

2

3

4

5

6

0 10 20 30 40 50

critic0

critic1

critic2

critic3

critic4

critic5

critic6

artist0

artist1

artist2

artist3

1

2

3

0 10 20 30 40 50

0

2

4

6

8

10

12

0 10 20 30 40 50

0

1

2

3

4

5

6

0 10 20 30 40 50

0

1

2

3

4

5

6

7

8

9

10

0 10 20 30 40 50

Figure 4: Graphs showing artist and critic dynamics from a

sample run lasting 50 rounds. The top row is for artists. The

bottom row is for critics. The ﬁrst column shows rankings

(0 always represents highest rank) on the basis of accep-

tances for artists and impact points for critics. The second

column shows the cumulative totals of the number of accep-

tances for artists and the number of impact points for critics

by rounds.

is at “solving” the image duplication problem).

Due to space limitations we are unable to discuss

all of these experiments. We focus on two runs where

a single master was ﬁrst identiﬁed; the two assigned

targets for each critic were selected from that mas-

ter; and critic targets did not change during the course

of an N = 30 round run. Because there were only ﬁve

targets available for critics, most critics shared at least

one target with another critic. However, if a critic had

an unshared target this created a niche opportunity

such that one artist could develop an exploitative rela-

tionship supporting a sustained period of submissions

and acceptances that catapulted both artist and critic

to higher rankings. These runs produced the greatest

turmoil within the tier system we used for critics (see

Figure 5). Interestingly, for both these runs the critic

that came to dominate had only one target image i.e.,

its two targets were the same. Evidently, for artists,

the existence of a critic with an unambiguous prefer-

ence was useful.

11 CONCLUSIONS

We have described a ﬂexible artist-critic agent based

simulation. It is based on a model whereby different

policies can be implemented in order to test the conse-

quences of what one might expect to occur. The poli-

cies we chose supported phenomena such as the de-

velopment of artistic styles; artist-critic cooperation;

IJCCI 2009 - International Joint Conference on Computational Intelligence

196

0

1

2

3

4

5

6

0 10 20 30 40 50

critic0

critic1

critic2

critic3

critic4

critic5

critic6

0

1

2

3

4

5

6

0 10 20 30

0

1

2

3

4

5

6

0 10 20 30

Figure 5: The results from two N = 30 round runs show-

ing dramatic changes in critic rankings during the course of

the run. For both runs critics were forced to choose their

two targets from a restricted database consisting of only 5

images all by the same “master”.

critic diversity (because critics have different prefer-

ences and these preferences may change over time);

and artistic freedom (because artists can exhibit “free-

dom of expression” by deciding not to blindly follow

the critics). Also noteworthy is that our model es-

tablishes a critic hierarchy such that over time some

critics become more equal than others. A distinguish-

ing feature of our model is that it values the discov-

ery of novel imagery by requiring that in addition to

being new, images must also be “ﬁt”. More impor-

tantly, it appears our policies allow for artistic trends

to arise via a strategy whereby a lower ranked critic

locates a niche to exploit and teams with an artist to

furnish images so that the resulting ﬂurry of accep-

tances cause images in this artist’s style to populate

the public gallery

REFERENCES

Aupetit, S., Bordeau, V., Monmanrch

e, N., Slimane, M.,

and Venturini, G. (2003). Interactive evolution of ant

paintings. In Congress on Evolutionary Computation,

volume 2, pages 1376–1383.

Baluja, S., Pomerleau, D., and Jochem, T. (1994). Towards

automated artiﬁcial evolution for computer-generated

images. Connection Science, 6:325–354.

Burns, K. (2006). Atoms of eve: a bayesian basis for aes-

thetics of style in sketching. Artiﬁcial Intelligence

for Engineering, Design, Analysis and Manufactur-

ing, 20:185–199.

Fisher, Y., editor (1995). Fractal Image Compression: The-

ory and Application. Springer, London.

Greenﬁeld, G. (2002). Color dependent computational aes-

thetics for evolving expressions. In BRIDGES 2002

Conference Proceedings, pages 9–16.

Greenﬁeld, G. (2005a). Designing metrics for the purpose

of aesthetically evaluating images. In Computational

Aesthetics 2005, pages 151–158. Eurographics Asso-

ciation.

Greenﬁeld, G. (2005b). Evolutionary methods for ant

colony paintings. In EvoWorkshops 2005 Proceed-

ings, volume LNCS 3449, pages 478–487.

Greenﬁeld, G. (2005c). On the origins of the term compu-

tational aesthetics. In Computational Aesthetics 2005,

pages 9–12. Eurographics Association.

Greenﬁeld, G. (2006). On evolving multi-pheromone ant

paintings. In 2006 IEEE World Congress on Compu-

tational Intelligence Conference Proceedings, pages

7425–7431. IEEE Press.

Greenﬁeld, G. (2007). Co-evolutionary methods in evolu-

tionary art. In The Art of Artiﬁcial Evolution: A Hand-

book on Evolutionary Art and Music, pages 357–380.

Springer.

Greenﬁeld, G. (2009). On variation within swarm paintings.

In Hyperseeing — Proceedings of ISAMA 2009, pages

5–12.

Jacob, C., Huslak, G., Boyd, J., Sayles, M., Nuytten, P.,

and Pilat, M. (2007). Swarmart: interactive art from

swarm intelligence. Leonardo, 40:248–254.

Jacobsen, T. (2006). Bridging the arts and sciences:

a framework for the psychology of aesthetics.

Leonardo, 39:155–162.

Machado, P. and Cardoso, A. (1998). Computing aesthet-

ics. In Proceedings XIV-th Brazilian Symposium on

Artiﬁcial Intelligence, pages 219–229.

Machado, P. and Cardoso, A. (2002). All the truth about

NEvAr. Applied Intelligence, Special Issue on Cre-

ative Systems, 16(2):101–119.

Machado, P., Romero, J., and Manaris, B. (2007). Exper-

iments in computational aesthetics: an iterative ap-

proach to stylistic change in evolutionary art. In The

Art of Artiﬁcial Evolution: A Handbook on Evolution-

ary Art and Music, pages 381–415. Springer.

Moura, L. and Ramos, V. (2002). Swarm paintings - nonhu-

man art. In Architopia: Book, Art, Architecture, and

Science, pages 5–24. Institut d’Art Contemporain.

Ross, B., Ralph, W., and Zong, H. (2006). Evolutionary

image synthesis using a model of aesthetics. In 2006

IEEE Congress on Evolutionary Computation, pages

3832–3839. IEEE Press.

Saunders, R. and Gero, J. (2002). Artiﬁcial creativity: a

synthetic approach to the study of creative behavior. In

Creative Design V. Key Centre of Design Computing

and Cognition.

Schmidhuber, J. (2007). Simple algorithmic principles

of discovery, subjective beauty, selective attention,

curosity and creativity. In Proceedings 10-th Inter-

national Conference on Discovery Science, volume

LNAI 4755, pages 26–38.

Sims, K. (1991). Artiﬁcial evolution for computer graphics.

In Proc. SIGGRAPH ’91, volume 25, pages 319–328.

Urbano, P. (2005). Playing in the pheromone playground:

experiences in swarm painting. In EvoWorkshops

2005 Proceedings, volume LNCS 3449, pages 527–

532.

SIMULATING ARTIST AND CRITIC DYNAMICS - An Agent-based Application of an Evolutionary Art System

197