Evolving Art: Past, Present and Future

Penousal Machado

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

Keywords:

Evolutionary Computation, Evolutionary Art, Machine Learning, Computational Creativity.

Abstract:

Evolutionary Computation techniques have been applied in recent years to several ﬁelds in the Arts. This talk

overviews how they have been used to create different types of artefacts, along with how past developments

relate to current approaches, trends and challenges. The focus is one of the main challenge in the ﬁeld: ﬁtness

assignment. We analyse this challenge from the perspective of the interplay between the evolutionary system

and the user, and discuss how Machine Learning, Evolutionary Computation and Human-Computer Interaction

techniques can be combined to create Computer-Aided Creativity systems that allow users to express their

artistic and aesthetic intentions.

1 INTRODUCTION

At the very ﬁrst Artiﬁcial Life conference, held in Los

Alamos in September 1987, Richard Dawkins presen-

ted Biomorphs, a system that allowed users to act as

selective breeders of artiﬁcial creatures, guiding the

evolution of their morphology, and thus demonstra-

ting the power of evolution (Dawkins, 1987).

Not long after, Karl Sims demonstrated how user-

guided evolution could be used to evolve abstract ima-

ges, 3D shapes, namely artiﬁcial plants, and animati-

ons (Sims, 1991). William Latham and Peter Todd

were also instrumental for the popularisation of Evo-

lutionary Art by commercialising Evolutionary Art

software (Todd and Latham, 1992).

Eventually, the early work of these, and other, pio-

neers led to the birth of Evolutionary Art as an area of

research. The core goal of Evolutionary Art as a disci-

pline can be deﬁned as the development and applica-

tion of evolutionary techniques for the generation of

computer graphics. An analysis of the research done

throughout the years allows the identiﬁcation of se-

veral core challenges, opportunities, and open issues

(see, e.g., McCormack (2007)). These tend to follow

on two main categories: representation and evalua-

tion.

In what concerns representation, it is important to

differentiate three main types of approaches: declara-

tive, parametric, and procedural (Xiao et al., 2019). In

declarative approaches, the genotype encodes or des-

cribes (i.e. declares) the characteristics of the indivi-

https://orcid.org/0000-0002-6308-6484

dual, i.e. the phenotype. An example of a declara-

tive approach is the work by Baker (1993) who uses a

Genetic Algorithm where each genotype encodes the

coordinates of a set of lines to evolve line drawings. In

parametric approaches, one evolves a set of parame-

ters that inﬂuences the behaviour of a generative art

system. Notable examples of this approach include

the works of Draves (2007), where the genotype is a

set of parameters of a fractal formula, and the work

of Machado et al. (2016), where the genotype is a set

of parameters that deﬁnes the behaviour of artiﬁcial

ants. Finally, in procedural approaches, the genotype

is a program or procedure that, when executed, ge-

nerates the phenotype. The most famous example of

such approach is the seminal work of Sims (1991)

who uses Genetic Programming to evolve symbolic

expressions that, once interpreted, result in colour

images. This expression-based procedural approach

has become the most popular for the evolution of ima-

ges. As such, its theoretical and practical expressive

power becomes of importance. While Machado and

Cardoso (2002) demonstrate that this sort of procedu-

ral representation has the theoretical power to gene-

rate any image, they also point out that in practice the

type of image these systems tend to generate is inti-

mately linked with the primitives they use. As such,

representation has a major impact on the nature and

quality of the results, making it a hot topic of research

since the inception of Evolutionary Art.

We introduce several alternative representation

schemes and analyse the impact of representation in

the outcome of the systems. In particular, we explore

the use of a multi-chromosome Genetic Programming

Machado, P.

Evolving Art: Past, Present and Future.

DOI: 10.5220/0008345700050008

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 5-8

ISBN: 978-989-758-350-6

approach to evolve assemblages of objects (Grac¸a and

Machado, 2015), graphs to evolve non-deterministic

context-free design grammars able to create a fa-

mily of images from a single genotype (Machado

et al., 2015a), behavioural parameters to evolve Non-

Photorealistic Renderings (Machado et al., 2014),

and digital stencil templates (Martins et al., 2018) to

evolve typefaces.

If Evolutionary Art lives up to its name, then ﬁt-

ness assignment becomes the biggest challenge, since

evolving art requires assessing the artistic quality of

candidate images, which, arguably, requires Artiﬁcial

General Intelligence. Early efforts in the ﬁeld circum-

vented this difﬁculty by resorting to user-guided evo-

lution, i.e. the user selected the images to breed, gui-

ding the evolutionary process. While this approach

has many merits, allowing the generation of remarka-

ble images that match the preferences of the user, it

also has severe limitations. By deﬁnition, this type

of user-guided systems lacks autonomy and ability to

judge their production, which is a major limitation

from an Artiﬁcial Intelligence perspective. Additio-

nally, from a practical perspective, these systems re-

quire constant intervention by the user, which beco-

mes extremely time-consuming, leads to user fatigue

and jeopardises the quality of the results.

Over the years, researchers made several contribu-

tions to the automation of ﬁtness assignment. These

can be classiﬁed in two major categories: the use of

hardwired ﬁtness functions and the use of Machine

Learning techniques.

In the ﬁrst case, the authors create a static function

or program that assesses the quality of the evolved

images. However, it has been proven extremely hard

to formally deﬁne aesthetic criteria and, in most cases,

it is trivial to demonstrate by counterexamples that the

conditions considered by the authors are neither suf-

ﬁcient nor necessary to capture a general notion of

aesthetics. Notwithstanding, several examples exist

that indicate that some aesthetical principles can be

partially formalised, explored and exploited, allowing

the autonomous evolution of images. Among such re-

search efforts we highlight the works of Machado and

Cardoso (2002), who use complexity estimates to as-

sign ﬁtness; Greenﬁeld (2003), who proposes a multi-

objective optimisation approach to evolve images that

satisfy several criteria; Ross et al. (2006), who pro-

motes the evolution of images that show a “natura”

distribution of colour gradients; Romero et al. (2012),

demonstrate how complexity measure can be used in

aesthetic appreciation tasks, later showing how they

relate to humans perception of complexity (Machado

et al., 2015b); and Reed (2013), who revisit Birk-

hoffs work, using aesthetic measures to evolve vase

designs.

In what concerns the use of Machine Learning for

ﬁtness assignment purposes, the seminal work of Ba-

luja et al. (1994) is the ﬁrst effort in the ﬁeld of Evo-

lutionary Art. Interestingly, they propose the use of a

Deep Neural Network to learn user preferences, and

although their results do not live to their own expec-

tations, this work paves the way for future research

in this area. Romero et al. (2003) realise that lear-

ning user preferences is a demanding task and suggest

combining a general purpose Evolutionary Art system

with a Machine Learning classiﬁer, trained to detect

human faces. Roughly ten years later, Machado et al.

(2012) implement this idea, showing that a conven-

tional expression-based Evolutionary Art system gui-

ded by Machine Learning classiﬁers is able to evolve

recognisable faces, ﬂowers, leaves, lips, etc. In la-

ter works, they combine several classiﬁers to evolve

ambiguous images, i.e. images that induce multista-

ble perception, a phenomenon that occurs when the

brain (or the computer) is confronted with visual sti-

mulus that can be interpreted in multiple ways (Ma-

chado et al., 2015c). Furthermore, by changing the

representation, they are able to evolve typefaces (Mar-

tins et al., 2018), photorealistic faces (Correia et al.,

2016), as well as other types of imagery (Assunc¸

et al., 2015), showing the impact of representation on

the outcomes of the evolutionary process.

This type of works highlight the power of Ma-

chine Learning, but also its current limitations. As

Baluja et al. (1994) already indicated, the evolutio-

nary engine tends to ﬁnd ways of exploiting the li-

mitations of the Neural Networks and this way fool

them. For instance, when trying to evolve faces, the

evolutionary engine routinely ﬁnds and converges to

images that, albeit detected as faces by the Machine

Learning classiﬁer, do not resemble faces to the hu-

man eye. The use of false positives generated throug-

hout the evolutionary runs to enrich training data-

sets has been explored by Correia (2018), who shows

that signiﬁcant improvements of performance can be

obtained.

In a different line of research, Machado et al.

(2007) present an adversarial system that promotes

the competition between a Neural Network discri-

minator and an Evolutionary Computation generator.

This approach, which precedes Generative Adversa-

rial Networks (Goodfellow et al., 2014), results in a

continuous pursuit of novelty, style change and re-

invention.

Although the automation of ﬁtness assignment po-

ses many relevant scientiﬁc challenges and questions,

full automation has a cost: users are no longer able to

express themselves through such systems. This rea-

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

lisation led us to propose an approach, meta-level in-

teractive evolution, that overcomes these limitations.

The core idea is to allow users to become designers of

the ﬁtness function, by allowing them to specify their

preferences and goals through the use of a responsive

user interface, which implicitly deﬁnes ﬁtness. This

approach frees users from the need of evaluating thou-

sands of images, as is the case of user-guided evolu-

tion, while still engaging the users, allowing them to

inﬂuence the result of the system and, above all, gi-

ving them a sense of authorship. Following this line

of research, Photogrowth (Machado et al., 2016), a

system that relies on the simulation of artiﬁcial ant

species to produce Non-Photorealistic Renderings, al-

lows the user to design ﬁtness functions by specifying

features pertaining the desired behaviour of the ants,

as well as features related with the output image.

A ﬁnal word goes to recent advancements in the

ﬁeld of Machine Learning. Considering the success

of Generative Adversarial Networks and Style Trans-

fer approaches, which set new expectations for the ap-

plication of Artiﬁcial Intelligence to artistic domains,

we analyse their strengths and limitations, identifying

opportunities for research.

REFERENCES

Assunc¸

ao, F., Correia, J., Martins, P., and Machado, P.

(2015). Evolving families of shapes. In Yang, Q. and

Wooldridge, M., editors, Proceedings of the Twenty-

Fourth International Joint Conference on Artiﬁcial In-

telligence, IJCAI 2015, Buenos Aires, Argentina, July

25-31, 2015, pages 4134–4135. AAAI Press.

Baker, E. (1993). Evolving line drawings. Technical Report

TR-21-93, Harvard University Center for Research in

Computing Technology.

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

automated artiﬁcial evolution for computer-generated

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

Correia, J., Martins, T., Martins, P., and Machado, P. (2016).

X-faces: The exploit is out there. In Pachet, F., Car-

doso, A., Corruble, V., and Ghedini, F., editors, Pro-

ceedings of the Seventh International Conference on

Computational Creativity (ICCC 2016), pages 164–

182. Sony CSL Paris, France.

Correia, J. a. (2018). Evolutionary Computation for Clas-

siﬁer Assessment and Improvement. PhD thesis, Uni-

versity of Coimbra.

Dawkins, R. (1987). The blind watchmaker: why the evi-

dence of evolution reveals a universe without design.

W.W. Norton and Company, Inc., New York.

Draves, S. (2007). Evolution and collective intelligence of

the electric sheep. In Romero, J. and Machado, P., edi-

tors, The Art of Artiﬁcial Evolution: A Handbook on

Evolutionary Art and Music, pages 63–78. Springer

Berlin Heidelberg.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Courville, A., and Ben-

gio, Y. (2014). Generative adversarial nets. In Ghahra-

mani, Z., Welling, M., Cortes, C., Lawrence, N. D.,

and Weinberger, K. Q., editors, Advances in Neu-

ral Information Processing Systems 27, pages 2672–

2680. Curran Associates, Inc.

Grac¸a, F. and Machado, P. (2015). Evolving assemblages.

IJART, 8(2):167–184.

Greenﬁeld, G. R. (2003). Evolving aesthetic images using

multiobjective optimization. In Sarker, R., Reynolds,

R., Abbass, H., Tan, K. C., McKay, B., Essam, D., and

Gedeon, T., editors, Proceedings of the 2003 Congress

on Evolutionary Computation CEC2003, pages 1903–

1909, Canberra. IEEE Press.

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

NEvAr. Applied Intelligence, Special Issue on Crea-

tive Systems, 16(2):101–119.

Machado, P., Correia, J., and Assunc¸

ao, F. (2015a). Graph-

based evolutionary art. In Gandomi, A. H., Alavi,

A. H., and Ryan, C., editors, Handbook of Genetic

Programming Applications, pages 3–36. Springer In-

ternational Publishing, Cham.

Machado, P., Correia, J., and Romero, J. (2012).

Expression-based evolution of faces. In Evolutionary

and Biologically Inspired Music, Sound, Art and De-

sign - First International Conference, EvoMUSART

2012, M

alaga, Spain, April 11-13, 2012. Proceedings,

volume 7247 of Lecture Notes in Computer Science,

pages 187–198. Springer.

Machado, P., Martins, T., Amaro, H., and Abreu, P. H.

(2014). An interface for ﬁtness function design. In

Romero, J., McDermott, J., and Correia, J., editors,

Evolutionary and Biologically Inspired Music, Sound,

Art and Design - Third International Conference, Evo-

MUSART 2014, Granada, Spain, April 23-25, 2014.

Proceedings, volume 8601 of Lecture Notes in Com-

puter Science. Springer.

Machado, P., Martins, T., Amaro, H., and Abreu, P. H.

(2016). Beyond interactive evolution: Expres-

sing intentions through ﬁtness functions. Leonardo,

49(3):251 – 256.

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

ments in computational aesthetics: An iterative appro-

ach to stylistic change in evolutionary art. In Romero,

J. and Machado, P., editors, The Art of Artiﬁcial Evo-

lution: A Handbook on Evolutionary Art and Music,

pages 381–415. Springer Berlin Heidelberg.

Machado, P., Romero, J., Nadal, M., Santos, A., ao Correia,

J., and Carballal, A. (2015b). Computerized measures

of visual complexity. Acta Psychologica, 160(1):43 –

57.

Machado, P., Vinhas, A., Correia, J., and Ek

art, A. (2015c).

Evolving ambiguous images. In Yang, Q. and Wool-

dridge, M., editors, Proceedings of the Twenty-Fourth

International Joint Conference on Artiﬁcial Intelli-

gence, IJCAI 2015, Buenos Aires, Argentina, July 25-

31, 2015, pages 2473–2479. AAAI Press.

Martins, T., Correia, J., Costa, E., and Machado, P. (2018).

Evotype: Towards the evolution of type stencils. In

Romero, J. and Liapis, A., editors, Evolutionary and

Evolving Art: Past, Present and Future

Biologically Inspired Music, Sound, Art and Design.

Springer International Publishing. (To Appear).

McCormack, J. (2007). Facing the future: Evolutionary

possibilities for human-machine creativity. In Ro-

mero, J. and Machado, P., editors, The Art of Artiﬁ-

cial Evolution: A Handbook on Evolutionary Art and

Music, pages 417–451. Springer Berlin Heidelberg.

Reed, K. (2013). Aesthetic measures for evolutionary vase

design. In Machado, P., McDermott, J., and Carbal-

lal, A., editors, Evolutionary and Biologically Inspi-

red Music, Sound, Art and Design - Second Internati-

onal Conference, EvoMUSART 2013, Vienna, Austria,

April 3-5, 2013. Proceedings, volume 7834 of Lecture

Notes in Computer Science, pages 59–71. Springer.

Romero, J., Machado, P., Carballal, A., and Santos, A.

(2012). Using complexity estimates in aesthetic image

classiﬁcation. Journal of Mathematics and the Arts,

6(2-3):125–136.

Romero, J., Machado, P., Santos, A., and Cardoso, A.

(2003). On the development of critics in evolutio-

nary computation artists. In G

unther, R. et al., editors,

Applications of Evolutionary Computing, EvoWorks-

hops 2003: EvoBIO, EvoCOMNET, EvoHOT, EvoI-

ASP, EvoMUSART, EvoSTOC, volume 2611 of LNCS,

Essex, UK. Springer.

Ross, B. J., Ralph, W., and Hai, Z. (2006). Evolutionary

image synthesis using a model of aesthetics. In Yen,

G. G., Lucas, S. M., Fogel, G., Kendall, G., Salo-

mon, R., Zhang, B.-T., Coello, C. A. C., and Ru-

narsson, T. P., editors, Proceedings of the 2006 IEEE

Congress on Evolutionary Computation, pages 1087–

1094, Vancouver, BC, Canada. IEEE Press.

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

ACM Computer Graphics, 25:319–328.

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

Computers. Academic Press, Winchester, UK.

Xiao, P., Toivonen, H., Gross, O., Cardoso, A., Correia,

J., Machado, P., Martins, P., Oliveira, H. G., Sharma,

R., Pinto, A. M., D

ıaz, A., Francisco, V., Gerv

as, P.,

Herv

as, R., Le

on, C., Forth, J., Purver, M., Wiggins,

G. A., Miljkovi

c, D., Podpe

can, V., Pollak, S., Kralj,

J.,

Znidar

c, M., Bohanec, M., Lavra

c, N., Urban

T., van der Velde, F., and Battersby, S. (2019). Con-

ceptual representations for concept creation. ACM

Computing Surveys.

BRIEF BIOGRAPHY

Penousal Machado leads the Cognitive and Media Sy-

stems at the University of Coimbra. His research inte-

rests include Evolutionary Computation, Computati-

onal Creativity, and Evolutionary Machine Learning.

In addition to the numerous scientiﬁc papers in these

areas, he is the recipient of scientiﬁc awards, inclu-

ding the EvoStar Award for Outstanding Contribu-

tion to Evolutionary Computation in Europe and the

award for Excellence and Merit in Artiﬁcial Intelli-

gence granted by the Portuguese Association for Ar-

tiﬁcial Intelligence. His works have been presented in

venues such as the National Museum of Contempo-

rary Art (Portugal) and the Talk to me exhibition of

the Museum of Modern Art, NY (MoMA).

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence