USING AESTHETIC LEARNING MODEL TO CREATE ART

Yang Li

1∗

, Chang-Jun Hu

and Jing-Qin Pang

School of Computer and Communication Engineering

University of Science and Technology Beijing, Beijing 100083, China

School of Materials Science and Engineering, University of Science and Technology Beijing, Beijing 100083, China

Keywords:

Aesthetic learning, Evolutionary art, Interactive evolutionary computation, Computational aesthetics.

Abstract:

An aesthetic learning model is proposed that applies evolutionary algorithm to generate art. The model is

evaluated using an evolutionary art system by human subjects. The advantages of the model is that it helps

user to automate the process of image evolution by learning the user’s preferences and applying the knowledge

to evolve aesthetical images. This paper implements four categories of aesthetic metrics to establish user’s

criteria. In addition to evolutionary images, external artworks are also included to guide evolutionary process

towards more interesting paths. Then we described an evolutionary art system which adopted the aesthetic

model in detail. Last, we evaluate the aesthetic learning model in several independent experiments to show the

efﬁciency at predicting user’s preferences.

1 INTRODUCTION

Evolutionary Art is originated from the work of

Dawkins (Dawkins, 1986), his Biomorphs program.

Following the earliest ideas of Sims(Sims, 1991) and

Latham(Latham and Todd, 1992), which use lisp ex-

pressions to specify the color of every pixel, a wide

research (see, for example, (Lutton, 2006) (Bentley,

1999) (Machado and Cardoso, 2002)) used Genetic

Programming(GP) to evolve aesthetic images. One of

the signiﬁcant challenges faces such systems is to ﬁnd

an appropriate ﬁtness function for guiding the evolu-

tionary process. Nowadays, most systems rely on In-

teractive Evolutionary Computation (IEC), which the

user take the tedious task to make decisions for every

generation. However, this could cause serious prob-

lems, like premature convergence, user fatigue and

several other disadvantages(Takagi, 1998).

Most of the existing systems take a long time to

ﬁnd interesting images. So users usually lost interest

in exploring design space further. In order to reduce

user fatigue, it is important to prevent stagnation in

evolutionary process, keep user’s interests high and

automate aesthetic judgements. Thus it suffers from

at least two major difﬁculties. First, it is hard to tell

exactly what metrics that inﬂuence individual’s ﬁnal

∗

This work was supported by China Postdoctoral Sci-

ence Foundation Funded Project and the Centre of Excel-

lence for Research in Computational Intelligence and Ap-

plications (CERCIA) at the University of Birmingham, UK.

aesthetic criteria. Second, the searching space is of-

ten ﬁxed by a certain style, which stuck the exploring

paths in a local optimum. This paperwill proposenew

techniques to solve these two issues.

Considering the above issues, our system is aim

at overcoming the drawbacks of existing systems and

automating the aesthetic judgements. This goal is

achieved by (a)devising several metrics and establish-

ing an aesthetic model for measuring user’s aesthetic

values; (b)introducing external images to the training

samples, which could apply external evaluations and

guide the evolutionary process towards more interest-

ing paths.

Three contributions are made by our work. First,

four categories of metrics have been established for

extracting meaningful aesthetic standards. Second,

both internal IEC images and external images col-

lected from internet are included in training the aes-

thetic model. Third, the model based on a classiﬁer is

introduced to automatically distinguish aesthetic im-

ages. To further manipulate the evolutionary process

easily, intuitive mutation parameters are applied.

Our contributions havebeen tested with several in-

dependent experiments by our system. We monitored

how these metrics evolved over iterations, quantita-

tively inﬂuenced the aesthetic model and how the sys-

tem directed the selection process by using the aes-

thetic model. We ﬁnd that the aesthetic model is efﬁ-

cient for most of the users to model their preferences

and properly classifying high, medium and low val-

174

Li Y., Hu C. and Pang J..

USING AESTHETIC LEARNING MODEL TO CREATE ART.

DOI: 10.5220/0003670001740183

In Proceedings of the International Conference on Evolutionary Computation Theory and Applications (ECTA-2011), pages 174-183

ISBN: 978-989-8425-83-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

ued images by learning the interactive run and exter-

nal images selected by user.

The paper is organized as follows: Section 2 be-

gins with an overview of previous work on automatic

ﬁtness assignment and computational aesthetics; in

Section 3 we describe our aesthetic model; follow-

ing with Section 4, which includes a global overview

of the system and the experimental results; ﬁnally, in

Section 5 we draw conclusions and point directions

for future research.

2 STATE OF ART

Evolutionary Algorithm has been successfully ap-

plied in the ﬁelds of design (Bentley, 1999), art

processing (Bachelier, 2008) and imagery generat-

ing (Ventrella, 2008). Most evolutionary art system

use IEC technology, which is based on subjective

human evaluation (Sims, 1991)(Poli and Cagnoni,

1991)(Lutton, 2006)(Rooke, 2002)(Li et al., 2009). A

thorough survey of application and interface research

on IEC can be found in (Takagi, 1998). However,

the user fatigue is one of the biggest problem in this

domain. Therefore, how to automatic the evolution

poses signiﬁcant challenges to this ﬁeld.

Several systems havebeen successfully performed

automatic evaluation in music composition (Papadou-

los and Wiggins, 1999)(Todd and Werner, 1998)(Ma-

naris et al., 2005)(Manaris et al., 2007). While in the

ﬁeld of evolutionary art, it makes the work more dif-

ﬁcult due to the lack of explicit theory like in music

(for a detailed survey see (Machado et al., 2008)).

Two important issues has been addressed in this

paper to formulate an appropriate ﬁtness function for

aesthetic purposes. First, what kind of metrics or cri-

teria inﬂuence the ﬁnal aesthetic judgement; Second,

the relations among them need to be established to as-

sign aesthetic value.

Taking account of the ﬁrst issue, the automatic

aesthetic judgement falls into the realm of computa-

tional aesthetics, which applies computational meth-

ods that can make applicable aesthetic decisions in

a similar fashion as humans can (Hoenig, 2005). A

brief historical review of the origins of the term could

be found in (Greenﬁeld, 2005). Birkhoff ﬁrst formal-

ized the aesthetic metric is the ratio between order and

complexity. This metric is then quantiﬁedby different

measurements on the following work. The work of

Bense (Bense, 1969) and Jaume (Rigau et al., 2008)

deﬁnes the complexity and order from the Shannon’s

InformationTheory(Shannon,1951). However, it still

needs tests to prove its validity, especially the rela-

tions between complexity and order.

Later, expression-based evolutionary art system,

NEvAr (Machado and Cardoso, 2002) applied an au-

tomated ﬁtness assignment, which consider both visu-

ally complexity and processing complexity are very

important factors. The aesthetic value is estimated

through the division of the complexity of visual stim-

ulus and complexity of the percept. The two measure-

ments are implemented by JPEG compression and

fractal compression for capturing the self-similarities

of the image. This formula is tested by ”Test of Draw-

ing Appreciation”, which is a standardized psycho-

logical test to evaluate art aesthetics (Machado and

Cardoso, 1998). The result of the test were surpris-

ingly good.

An empirical aesthetic model has been used by

Brian et al. for evolutionary image synthesis (Ross

et al., 2006). The main metric used in this model

is proposed by Ralph (Ralph, 2006), which is based

on hundreds of ﬁne art and bell curve distribution has

also been found in the color gradients.

The secondissue in automatic aestheticjudgement

is to incorporate aesthetic criteria and the metrics of

image in the evaluation strategy. This is difﬁcult

to overcome, because different people have different

principles of beauty and appreciations of various met-

rics.

The most simple way is to merge them together

by a weighted sum. The work of (Wannarumon et al.,

2008) introduced a hardwired ﬁtness function which

combined several measurements that reﬂect aesthet-

ics of forms and fractals for jewelry design. This

measurement is based on mathematical foundations

of fractal geometry, chaotic behavior and image pro-

cessing. But ﬁxed ﬁtness function often introduces

bias to the evaluation in evolutionary process.

The ﬁrst published research which uses machine

learning approaches is the work of (Baluja et al.,

1994),which relied on ANN to alleviate the burden

of users. However, the results turned out to be ”some-

what disappointing” (Baluja et al., 1994). (Ross et al.,

2006) use multi-objective ﬁtness testing to evaluate

the candidate textures according to multiple criteria.

The multi-objective approach to evaluate textures is

proposed in (Ross and Zhu, 2004). A Pareto rank-

ing strategy is used to rank the populations, some di-

versity promoting strategies are then provided to gen-

erate a diversity assortment of solutions (Ross et al.,

2006). The results show that the techniques are ide-

ally suited to texture synthesis.

Recently, Machado et al. present an IEC with

a similarity-based approach in their evolutionary art

system (Machado et al., 2008). An Artiﬁcial Art

Critic (AAC) is introduced to distinguish between ex-

ternal images (e.g., paintings) and the internal images

USING AESTHETIC LEARNING MODEL TO CREATE ART

175

created by evolutionary process. Thus it enables new

trends and explorations in stylistic changes.The AAC

includes two models: (i) the feature extractor and (ii)

Artiﬁcial Neural Network (ANN). The evaluator is

used to distinguish between the internal and external

images (Machado et al., 2008). (Ekart et al., 2011)

propose a set of aesthetic measures identiﬁed through

observation of human selection of images and then

use these for automatic evolution of aesthetic images.

3 AESTHETIC LEARNING

MODEL

Three important issues are addressed in our aesthetic

learning model. First, image metrics which we con-

sidered very important factors in aesthetic judgements

should be calculated in the model. Although different

people might disagree on the meaning of beauty, most

metrics are primary characteristics of visual systems

and user’s interest. Second, training populations are

speciﬁcally chosen for the learning classiﬁer. Outer

images are involved to avoid convergency to a spe-

ciﬁc style. Besides, inner images generated using

IEC are also collected for training sets in every itera-

tion. Third, decision tree is used to build the learning

model. Applying populations with assigned ﬁtnesses

using the model in the continuous generation is one

way to reduce the users’ fatigue. This is achieved by

learning the metrics that extracted from images and

applying the model to the following evolutionary pro-

cess.

3.1 Aesthetic Metrics

Inspired by the works of (Rigau et al., 2008),

(Schmidhuber, 1997) and (Machado and Cardoso,

1998), we complement the metrics with these theory

by considering these established aesthetic measures.

This work is a continuation of our previous work that

primarily focused on the estimation of image com-

plexity and image order (Li and Hu, 2010). These

metrics that we choose fall into four categories: color

ingredient, image complexity, image order and metric

based on Machado and Cardoso’s work(MC metric).

These categories are shown in Table 1. Additionally,

we divide the original image into ﬁve parts to calcu-

late these metrics, because spatial information is able

to help analyze different parts of the image.

We apply these metrics to different parts of the

image and denote them as m

(1 ≤ i ≤ 14), which are

described as follows.

Table 1: Aesthetic Metrics Summary

Category Description

Color Ingredient Average value and standard deviation

of Hue, Saturation and Lightness

Image Complexity Information entropy of HSL, RGB and Y

709

Image Order Kolmogorov’s complexity based on

JPEG and fractal compressor

MC metric Image complexity(IC) and Processing

complexity(PC) based on MC

3.1.1 Color Ingredient

Our motivation for the choice of color ingredient is

based on HSL color space, which is more intuitive

than RGB to describe perceptual color. Hue speciﬁes

the base color, the other two specify the saturation and

the light of the color, three of which are all important

factors addressed in perception. The color channels

we used for the color ingredient metric is hue, satu-

ration and lightness. We converted the RGB data of

each image to HSL color space, producing three ma-

trices I

, I

, each of which the dimension is M×N.

Then we proceed by calculating the average value and

the standard deviation in each of these three channels.

Although a computer may describe a color using

the amounts of red, green and blue, a human prob-

ably deﬁne it by its attributes of hue, saturation and

lightness. Hue is the angle in the color wheel, the

ﬁrst metric is computed as the average value of hue

∑

M− 1

x=0

∑

N−1

y=0

(x, y). Saturation and light-

ness are also calculated in the same way using I

and

separately to get the metrics m

and m

. The aver-

aged saturation indicator represents the purity of the

color. While light also turns out to be a very impor-

tant factor to discriminate between appealing and un-

appealing images, because sometimes it indicates the

sunlight, shadow or darkness.

In addition, the standard deviation of the three

channels are computed as metrics m

, m

and m

These metrics represents the scale of the color dis-

tribution. In the artistic domain, the range of colors in

paintings that selected by the artist is the fundamental

step to produce a ﬁne art.

3.1.2 Image Complexity

Our measurements of image complexity are based on

the concepts of information theory. The estimation

of the metric has been stated in (Li and Hu, 2010).

The relationship between complexity and aesthetics

has been widely discussed (Birkhoff, 1933)(Machado

and Cardoso, 1998) (Rigau et al., 2008). We use the

image complexity measurement from an information-

theoretic perspective for this metric. The reason we

consider informational aesthetics measurement is in

(Rigau et al., 2008).

ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications

176

HSL, RGB and Y

709

channels are used to calcu-

late the information content of the image. The image

complexity for hue channel is computed as follows:

= N × (−

∑

x∈X

hue

(x)logp

hue

(x)). (1)

where p

hue

is the probability distribution in channel

hue, which is calculated as follows:

hue

(x) =

(0 ≤ x ≤ χ). (2)

where n

is the number of pixels in bin x.The value

of χ is different according to the channel, which rep-

resents the bins of the histogram. In our case, χ for

hue, saturation and lightness channel are 360, 100 and

100 respectively. The image complexity is calculated

by multiplying each pixel’s Shannon entropy in ev-

ery channel and the number of pixels. m

and m

are

also computedin the same manner for the information

channel of saturation and lightness.

The complexity is also calculated from RGB rep-

resentation andY

709

, which is the luminance from lin-

ear red, green and blue. In RGB space, we divide it

into 512 cubes with eight equal partitions along each

axis to reduce the calculation in 256

dimension. The

intensity histogram is then reduced to 512 X

RGB

bins.

The χ for the channel Y

709

is 256. Probability dis-

tributions of the variables X

RGB

and X

709

are used to

calculate the entropy of RGB and Y

709

channel, which

we denote them as the features m

and m

3.1.3 Image Order

The image order is estimated based on the work

of Schmidhuber (Schmidhuber, 1997). (Li and Hu,

2010) have explored the following hypothesis: an im-

age which can be computed by an algorithm in the

shortest description is considered the most beautiful

in a set of images. The coding algorithm is presented

for the description of the human visual system. In

our case, we use the real-world fractal compressor to

achieve this metric. We choose this image compres-

sion method because it is intended to compress in a

visible way for human eye. Training images are as-

signed by three ﬁtness values. Thus the priori dis-

tribution of the images, −logP(i), is assigned to 1,

0.5 or 0 which represents high, medium or low value.

logP(C), the constant when C is given, is disregarded.

According to the deﬁnition, the order of the image is

calculated as follows:

−logP(i | C) =

CompressionRatio

− t

− logP(i). (3)

and ts are the end-time and start-time of fractal en-

coding. In order to calculate this evaluation, we ﬁrst

average the RGB pixels into grey ones, then the image

is divided into four equal regions. The fractal algo-

rithm encodes each region based on its self-similarity.

Thus the metric m

is approximated by the fractal

compression.

3.1.4 MC Metric

This metric is based on the aesthetic theory of

Machado andCardoso (Machado and Cardoso, 1998).

They stated that the aesthetic value of an artwork is

connected to Image Complexity (IC) and Processing

Complexity (PC). Although, they developed a for-

mula, the

ratio, to evaluate the aesthetic value, we

consider both complexities are signiﬁcant metrics for

the aesthetic model. Therefore, IC and PC are esti-

mated separately as independent metrics.

In order to compute IC, we compress the image

using JPEG compressor. In the compression method,

the image is coded with error, we can specify the

amount of detail kept and the compression ratio.The

IC metric is estimated through the division of the

root mean square error(RMSE) by the compression

ratio resulting from the JPEG compression, m

RMSE

CompressionRatio−JPEG

. Low values of IC indicate sub-

stantially compressible and low complexity of the im-

age. While high values of IC indicate not compress-

ible and more complex image.

PC suppose to reﬂect the compression process of

the image. The fractal image compression is some-

how closer to the way in which humans process the

images. We use fractal image compression for esti-

mating PC. The image varies in perception process as

the time passes, so CP is different in every moments.

In order to compute PC in different time points t

and

, we measure it separately as PC(t

) and PC(t

). We

argue that PC can be substituted by the following for-

mula:

= (PC(t

) × PC(t

)) × (

PC(t

) − PC(t

)

PC(t

)

). (4)

This process yields a total of 14 metrics for the

whole image. However, these global metrics could

be incorrect. To better capture metrics in different

regions of the image, we segment each image into

ﬁve different partitions: four quadrants and an cen-

tral square with the same size. Then we applied the

same 14 metrics to each of partitions and the global

image. Thus a total of 84 features are extracted for

each image.

3.2 Training Set

In order to build the aesthetic model, a training set

of 64× 64 pixel outer and inner images was used.

USING AESTHETIC LEARNING MODEL TO CREATE ART

177

The process used to obtain these images was as fol-

lows. For each generation, one image that the subject

choose as the parent of the next generation, as well

as other 66 images present on the screen, became the

inner candidates for the training set. The subject rank

these images into three categories: low, medium and

high, which are assigned to values 0.0,0.5 and 1.0 re-

spectively. Although collecting the inner images gen-

erated using IEC as the candidates would have been

easier than including the real world artworks, it leads

to speciﬁc stylistic images that belong to the EC-

generated class. Therefore, in order to provide the

training set with external evaluations, images which

the subject selected from paintings, photographs, in-

ternet, etc. were required. These outer images are

assigned to 1.0 as they were regarded as the interest-

ing images. Through several interactive evolutions,

hundreds of images were collected in total.

An image library is built to save different artworks

that we collected from internet. During the evolu-

tionary process, the subject can choose any pleasing

images from our image library to adjust the model.

To train the model, images were scaled to range

60× 60. About 30 paintings were collected from dif-

ferent sources. The paintings are from the artists Ed-

vard Munch, Marc Chagall and Vincent Van Gogh.

Paintings from one artist can be chosen to allow a bias

towards a particular artist’s style, or a set of abstract

paintings can be selected to guide the model towards

a speciﬁc type.

The inner populationsare gathered during the evo-

lutionary process. The population size of each itera-

tion is 67. The total number of the inner populations

depends on the number of iterations before the evolu-

tionary process stops.

3.3 Learning Aesthetic

As we stated, this model is based on a classiﬁer, the

input of which is the 84 metrics that we introduced

and the ﬁtnesses. The classiﬁer is chosen by the fol-

lowing reasons:(1) we could explore the metrics of

image that are shown correlations with user’s aes-

thetic judgement; (2) images could be rendered with

ﬁtness assigned by the model in each generation in

order to help the user to evaluate populations.

ANN is the mostly used machine learning ap-

proach to learn evaluating aesthetic judgments. Al-

though this approach is elegant, the ANN training

process is usually time-consuming and not easy to un-

derstand. In (Machado et al., 2008), it needs 6 sec-

onds per image to extract features, and ANN train-

ing takes hours to train the model. And in the work

of Baluja, it becomes a very difﬁcult undertaking to

determine what the ANN has ’learned’, and an even

harder task to translate the knowledge embedded in

the ANN into understandable rules (Baluja et al.,

1994). These are the two reasons why we use deci-

sion trees to build the classiﬁer. By looking at the

decision tree, it is easy to see which variables split the

data into aesthetic category. This information is very

important for us to understand the nature of beauty

according to the user’s behavior.

Therefore, decision tree is one utility that could

help us better understand the inﬂuence of different

metrics and criterias directly. The structured rules that

learned from the model would be applied on the im-

ages to assist in making aesthetic decisions. In order

to build the classify model, we focus on C4.5 (Quin-

lan, 1993) decision tree to analyze the users’ behavior

using WEKA J48 (Witten and Frank, 2000) imple-

mentation.

In our paper, we convert our data collected from

evolutionary process into a ﬁle, which is the input

for the algorithm. 84 metrics are the attributes with a

set of different real values in several evolutionaryrun.

Then the algorithm of the decision tree recursively di-

vides the dataset into smaller subsets by selecting one

of the most informative attribute until it has been la-

beled as terminal. The model is then constructed by

the decision tree, which will be applied in the new

evolutionary process. It is used to classify new popu-

lations into three categories to help users accomplish

the primary evaluation.

4 EXPERIMENTS

In this section, we deﬁne what kind of evolutionary

art system we use for the interactive evolution, how

we conduct our experiments to evaluate the aesthetic

model, and what subjects participated in our exper-

iments. After that, we present an individual interac-

tive evolutionexperiment to better understand the aes-

thetic judgement process, use the automatic model to

compare the results, and discuss the effectiveness of

the aesthetic model by performing a number of runs

for different subjects.

4.1 System Description

Our system is comprised of two modules, image gen-

erator which uses GP to produce images and aesthetic

model. The overall architecture of our system is given

in Figure 1. The method we employed in image gen-

erator is described in Section 4.1.2. And the aesthetic

model is brieﬂy introduced in Section 3.

ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications

178

Figure 2: Main window of our system.

Figure 1: Overall architecture of our system.

4.1.1 Overall Architecture

Our evolutionary art system has four distinctive char-

acteristics. First, we setup four intuitive mutation pa-

rameters for user to manipulate the evolutionary pro-

cess. Second,training samples for the aesthetic model

are collected from every generation performed by the

user, while images that didn’t participate in the evo-

lutionary process are also included. This approach

not only enables more stylistic training images intro-

duced in learning process, but also make it more eas-

ier to train the model by applying the domestic im-

ages. Third, four categories of metrics are extracted

from the training samples. Finally, the decision tree is

introduced to build the model, and further assist user

in assigning ﬁtnesses for the continuous generation.

The main window of the system is shown in Fig-

ure 2. Initial populations generated randomly by GP

are displayed on the screen. From the displayed set

the user can assign ﬁtness by choosing one of the

three icons below each image, which represents high,

medium or low value. Besides the IEC images, outer

images also can be chosen from our painting library

during the evolutionary process, which are automati-

cally assigned high ﬁtness values. Four mutation pa-

rameters can be set manually to adjust rate of stylistic

changing. Evolutionary process stops when the user

meets his/her ﬁnal criteria or he/she loses interest in

exploring further populations. AutoFitness checkbox

is used to collect the data extracted from the previ-

ous process, learn the user’s aesthetic judgement and

build the aesthetic model.

4.1.2 Image Generator

Genetic Programming

Like most of the evolutionaryart system, the individu-

als in our system are represented by trees, this shares

many similarities with the application developed by

Sims. As such, the genotypes are mathematical func-

tions represented by trees, which are constructed by

a lexicon of functions and terminals. The functions

we use are a set of simple functions, such as tran-

scendental functions, arithmetic operators and logic

operations. The terminals are variables x, y, or con-

stants. And then it is normalized into the proper range

to specify the color of the pixels.

The symbolic expression is called for every pixel

of the image to calculate the color. There are two

kinds of root nodes to map the values into color. One

is RGB node, which is applied by three channels, red,

green and blue to determine the ﬁnal color. Another

one is color map node. The numeric value is called by

this node, which is then looked up in a color map to

transfer into the real color. In Figure 3 we present

some examples of genotype and its corresponding

phenotypes.

Mutation Operators

Mutation operators are one of the most important ge-

netic operators for the stylistic variations in images.

USING AESTHETIC LEARNING MODEL TO CREATE ART

179

Figure 3: Some examples of genotypes and their corre-

sponding phenotypes.

Figure 4: Some examples of generations according to dif-

ferent mutation operators.

Mutation is performed by changing values in a leaf

node or functions in an internal node, deleting the

subtree at any point or replacing the pruning part

with a new random subtree (Li and Hu, 2010). Four

kinds of mutation operators are applied in our system,

which are coarse mutation rate, color mutation rate,

pattern mutation rate and ﬁne mutation rate. They are

used to control the rate of changing the subtree, the

internal node, the root node and the leaf node. In Fig-

ure 4, the parent is the image on the top of the ﬁg-

ure, examples of mutation populations generated by

different mutation operators with corresponding mu-

tation rates are shown below. Although populations

are not precisely transformed by using each mutation

operator, they can lead to speciﬁc exploration accord-

ing to users’ preference.

4.1.3 Evolutionary and Learning Process

The evolutionary and learning process is brieﬂy de-

scribed as following steps:

1. A set of initial GP images are displayed on our

system.

2. The subject choose his/her favorite images from

the image library that we now collected from fa-

mous paintings on internet.

3. Fitness values of current populations are assigned

by selecting one of the three icons below each im-

age according to his/her preference. The one with

high ﬁtness is selected as the parent of next gener-

Table 2: Parameters of GP for representation.

Parameter Setting

Population size per generation 67

Mutation operator coarse, color, pattern and ﬁne mutation

Mutation Rate manually set every generation, in the range 0-100

Unary function set sin, cos, tan, abs, ﬂoor, ceil, sqr,

square, cube, log, exp, negative, spiral, circle

Binary function set +,-,*,/, max, min, pow, average

Ternary function set if, lerp

Terminal set X,Y, scalar and random constants

Initial maximum tree depth 6

ation. Four mutation parameters are also set man-

ually.

4. Aesthetic metrics are extracted from the cur-

rent populations and favorite paintings, and then

stored in a measure list.

5. The evolutionary process stops when a termina-

tion criteria is met. In our case, the number of

generations is ﬁxed or the subject choose to train

the aesthetic model.

6. The training result is shown on screen, according

to which the subject can choose continue training,

start over or use the model to continue generating.

7. The model is used to assign ﬁtnesses in the fol-

lowing generations.

4.2 Experimental Setup

In this section we present some of the experimental

results applied with our aesthetic model.

Our intention in implementing this system is not

to exclusively reduce errors in learning aesthetic but

to decrease the tedious work in IEC process, and try

to assist user to interactively evolve images that they

ﬁnd aesthetically pleasing. For this purpose, we are

mainly interested in analyzing the metrics relevant to

aesthetic judgement, reducing user fatigue and cate-

gorizing the populations to help preliminary evalua-

tion. Therefore, we ﬁrst conducted one single exper-

iment, from the ﬁrst IEC generation till the last itera-

tion that performed by the model we built. Then we

completed 42 runs on our system by 21 different sub-

jects, two experiments by each subject.

In the experiments we use the settings presented in

Table 2. 67 images are displayed in every generation,

all of which are rendered at 64 by 64 resolution. The

functions that we use in GP fall into three categories:

unary, binary and ternary functions. The terminals

are variable x, y position or constants. The maximum

initial tree depth is set to 6.

4.3 Analysis of the Model

We conducted one experiment from random popula-

tions to the last generation evolved by our aesthetic

ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications

180

Figure 5: Initial populations of the experiment. The icon

below each image indicate the individual’s ﬁtness value.

model. In this single experiment, the subject manip-

ulated 10 generations. IEC images generated from it-

eration to iteration, while eight paintings chosen from

our image library remains the same. The number of

the outer images still need to be adjusted in the future

work to solve the unbalanced classes problem.

Initial Populations

The initial populations of our system is shown in Fig-

ure 5. The population size of the initial random im-

ages is 67. These populations were created using ran-

dom tree depths to avoid monotony of the images.

Eight external paintings were also selected by user

which we collected from internet. The mutation rates

for the next generation were then set separately. The

third image from the upper-left corner of the ﬁgure

was selected as the parent. Parent from every itera-

tion is kept in the next generation, thus genotype with

high ﬁtness value occurs more than once in the evo-

lutionary process. In order to remove the repetition

in the aesthetic model, we delete the same population

that appears in the previous iteration in the training

set.

IEC Results

Considering the main goal of our work, preventing

stagnation in evolution and keeping users’ interests

high is one way to reduce user fatigue. Users usu-

ally lost interest in exploring the design space fur-

ther, because most of the systems produce imagesthat

are quite similar to each other or the process leads to

blind exploration. Therefore, we applied four muta-

tion operatorsto better manipulate the exploring paths

in search space.

In the experiment the subject performed the ex-

periment as we described before. The mutation rates

were set manually, as shown in Figure 6. We present

the experimental results attained in the 5th and the last

generation in Figure 7 and Figure 8. The results show

that from the ﬁrst iteration till the last, user could

easily control the convergence of optimum or main-

tain sufﬁcient diversity in exploring the design space.

0 2 4 6 8 10 12

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Mutation Rate

Iteration t

Mutation Rate

Coarse Mutation

Color Mutation

Pattern Mutation

Fine Mutation

Figure 6: Mutation Rate values along with iterations.

Figure 7: IEC populations from the 5th iteration.

Figure 8: IEC populations from the 10th iteration.

Most of the populations in 5th iteration generates the

pattern of a ﬂower. In the following iterations, the

user mainly focused on exploring the color and slight

variation in shape.

Training Results

Another method to reduce the evaluation work is to

build the aesthetic model to learn the users’ prefer-

ences. Our aesthetic model ﬁrst extracted metrics that

we stated before from IEC populations and outer im-

ages, then the training results is shown on screen (see

Figure 9). All the trainingresults, includes roc curves,

decision tree and classiﬁed rate are displayed. The

user decide whether to continue training, start over or

apply the model to the following evolutionary runs.

Table 3 shows the accuracy of the correctly classi-

ﬁed images with high, medium and low value, which

USING AESTHETIC LEARNING MODEL TO CREATE ART

181

Table 3: Number of image and correctly classiﬁed images.

Number of image in training set Correctly classiﬁed Rate

Low 30 73.33%

Medium 426 80.75%

High 345 83.19%

Figure 9: Training results after building aesthetic model.

Figure 10: Populations from 30th iteration using the aes-

thetic model.

we used a 10-fold cross validation. The results show

that our system is capable of capturing user’s prefer-

ence from evolutionary process. This is very impor-

tant for trying to reduce the user’s fatigue. The model

is applied to help ﬁlter out most of the uninteresting

populations, thus brings images satisﬁed users.

The results generated by the model for the follow-

ing evolutionary runs are shown in Figure10. We ﬁnd

that the aesthetic model is able to distinguish three

categories of images. Most of the low value images

are correctly classiﬁed by the model. And the popula-

tions in ﬁnal iteration share some similarities in stylis-

tic and color with the results generated by IEC.

4.4 Subjects Evaluation

We request 21 students in twenties to operate our sys-

tem, each of which runs twice, with and without aes-

thetic model. In the ﬁrst experiment, the subjects per-

formed our system like using most of the evolutionary

art system. In the second experiment, the subject used

the same populations in the 10th iteration generated

in previous experiments (we request subjects perform

more than 10 iterations in IEC), then used the training

model to generate next 30 iterations and report the ﬁt-

ness categories of the ﬁnal images for comparison.

Table 4: average test results of 21 subjects on the aesthetic

model.

Mode Total populations Number of Time Percentage of

populations generations consumption high valued images

IEC 884.4 13.2 10’48“ 43.07%

IEC

(Aesthetic Model) 3189.2 47.6 25’21“ 63.5%

Table 4 shows the average time consumption,

number of populations, number of generations and

percentage of high valued images made by human

subjects. Clearly, it took 2/3 of the time used by

IEC in each iteration using aesthetic model. In other

words, more images are generatedin certaintime with

aesthetic model. For IEC with aesthetic model, we

also ﬁnd that the percentage of high valued images

is increased although the total number of generation

is almost ﬁve times more than IEC process. These

results show that our aesthetic model is sufﬁcient to

predict the user’s preference and reduce user’s fatigue

in the evolutionary process.

5 CONCLUSIONS AND FUTURE

WORK

This paper introduce an aesthetic model in exploring

user’s behavior of evaluation in evolutionary art sys-

tem. The model has incorporated several new ideas

in reducing user’s fatigue. The metrics we extracted

from images are fall into four categories, color ingre-

dient, image complexity, image order and MC met-

ric. Four new features has signiﬁcantly improved our

system for stylistic changing and better exploration.

External evaluations of real world paintings or pho-

tographs help the model involving outside values of

metrics not only limited in IEC space. The four muta-

tion operators can help the system to avid stagnation

and blind exploration.

Although more rigorous and more comprehensive

evaluation of our model is needed, our preliminary

study here does illustrate the efﬁciency of out sys-

tem. The future work of this research includes sev-

eral tasks. First, comparison of different metrics help

us to better understand aesthetic criterias in evolution-

ary process. Second, exploring the relations between

the inner IEC evaluations and external evaluations for

outer images may also help to analyze the aesthetic

model.

ECTA 2011 - International Conference on Evolutionary Computation Theory and Applications

182

REFERENCES

Bachelier, G. (2008). The Art of Artiﬁcial Evolution:A

Handbook on Evolutionary Art and Music, chapter

Embedding of Pixel-Based Evolutionary Algorithms

in My Global Art Process, pages 249–268. Springer

Berlin Heidelberg New York.

Baluja, S., Pomerlau, D., and Todd, J. (1994). Towards

automated artiﬁcial evolution for computer-generated

images. Connection Science, 6(2):325–354.

Bense, M. (1969). Einf´uhrung in die informationstheoretis-

che Asthetik, chapter Graundlegung und Anwendung

in der Texttheorie (Introduction to the Information-

theoretical Aesthetics. Foundation and Application to

the Text Theory). Rowohlt Taschenbuch Verlag.

Bentley, P. (1999). Aspects of evolutionary design by com-

puters. In Proc. of Advances in Soft Computing-

Engineering Design and Manufacturing, pages 99–

118. London.

Birkhoff, G. D. (1933). Aesthetic Measure. Harvard Uni-

versity Press.

Dawkins, R. (1986). The Blind Watchmaker. Harlow Long-

man.

Ekart, A., Sharma, D., and Chalakov, S. (2011). Modelling

human preference in evolutionary art. In Evo Applica-

tions 2011, Part II, LNCS 6625, pages 303–312.

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

tional aesthetics. pages 9–12, Aire-la-Ville, Switzer-

land.

Hoenig, F. (2005). Deﬁning computational aesthetics. In

Neumann L., Sbert M., Gooch B., Purgathofer W.,

(Eds.), Eurographics Association.

Latham, W. and Todd, S. (1992). Evolutionary Art and

Computers. Academic Press, Winchester, UK.

Li, Y. and Hu, C. J. (2010). Aesthetic learning in an in-

teractive evolutionary art system. In Evo Applications

2010, Part II, LNCS 6025, pages 301–310.

Li, Y., Hu, C. J., and Yao, X. (2009). Innovative batik design

with an interactive evolutionary art system. Journal of

Computer Science and Technology, 24(6):1035–1047.

Lutton, E. (2006). Evolution of fractal shapes for artists and

designers. International Journal on Artiﬁcial Intelli-

gence Tools, 15(4):651–672.

Machado, P. and Cardoso, A. (1998). Computing aesthetics.

In Proceedings of XIVth Brazilian Symposium on Ar-

tiﬁcial Intelligence(SBIA ’98), LNCS, pages 219–229,

Porto Alegre, Brazil.

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

nevar. Applied Intelligence, Special issue on Creative

Systems, 16(2):101–119.

Machado, P., Romero, J., and Manaris, B. (2008). In The

Art of Artiﬁcial Evolution: A Handbook on Evolution-

ary Art and Music, chapter P. Machado:Experiments

in Computational Aesthetics: An Iterative Approach

to Stylistic Change in Evolutionary Art, pages 381–

415. SPRINGER.

Manaris, B., Romero, J., Machado, P., Krehbiel, D., Hirzel,

T., Pharr, W., , and Davis, R. (2005). Zipf’s law, music

classiﬁcation and aesthetics. Computer Music Jour-

nal, 29(1):55–69.

Manaris, B., Roos, P., Machado, P., Krehbiel, D., Pellicoro,

L., and Romero, J. (2007). A corpus-based bybrid ap-

proach to music analysis and composition. In Proc. of

the 22nd Conference on Artiﬁcial Intelligence (AAAI

07).

Papadoulos, G. and Wiggins, G. (1999). Ai methods for

algorithmic composition: A survey, a critical view and

future prospects. In Proc. of AISB’99 Symposium on

Musical Creativity, pages 110–117, Edinburgh, UK.

Poli, R. and Cagnoni, S. (1991). Genetic programming with

user-driven selection: Experiments on the evolutiona

of algorithms for image enhancement. In Proc. 2nd

Annual Conf. on Genetic Programming, pages 269–

277. Morgan Kaufmann.

Quinlan, J. R. (1993). C4.5: Programs for Machine Learn-

ing. Morgan Kaufmann Publishers.

Ralph, W. (2006). Painting the bell curve: The occurrence

of the normal distribution in ﬁne art. In preparation.

Rigau, J., Feixas, M., and Sbert, M. (2008). Informational

aesthetics measures. Computer Graphics and Appli-

cations, 28(2):24–34.

Rooke, S. (2002). Eons of Genetically Evolved Algorithmic

Images, pages 330–365. Morgan Kaufmann.

Ross, B. and Zhu, H. (2004). Procedural texture evolution

using multiobjective optimization. New Generation

Computing, 22(3):271–293.

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

image synthesis using a model of aethetics. In Proc.

2006 IEEE Congress on Evolutionary Computation,

pages 1087–1094, Vancouver, BC, Canada.

Schmidhuber, J. (1997). Low-complesity art. Leonardo.

Journal of the International Society for the Arts, Sci-

ences, and Technology, 30(2):93–103.

Shannon, C. (1951). Prediction and entropy of printed en-

glish. Bell System Technical Journal, 30:50–64.

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

In Proc. of the 18th Annual Conference on Computer

Graphics and Interactive Techniques SIGGRAPH ’91,

pages 319–328. New York, NY: ACM Press.

Takagi, H. (1998). Interactive evolutionary computation.

In Proc. of the 5th International Conference on Soft

Computing and Information Intelligent Systems, pages

41–50, Iizuka, Japan.

Todd, P. M. and Werner, G. (1998). Frankensteinian Meth-

ods for Evolutionary Music Composition, pages 313–

339. MIT Press/Bradford Books.

Ventrella, J. (2008). Evolving the mandelbrot set to imitate

ﬁgurative art. Innovations in Evolutionary Design.

Wannarumon, S., Bohez, E., and Annanon, K. (2008). Aes-

thetic evolutionary algorithm for fractal-based user-

centered jewelry design. Artiﬁcial Intelligence for

Engineering Design, Analysis and Manufacturing (AI

EDAM), 22(1):19–39.

Witten, I. and Frank, E. (2000). Data Mining: Practi-

cal Machine Learning Tools and Techniques with Java

Implementations. Morgan Kaufmann Publishers.

USING AESTHETIC LEARNING MODEL TO CREATE ART

183