Automatic Selection in Evolutionary Art

1
Automatic Selection in Evolutionary Art

• Neale Samways
• And
• Thorsten Schnier

School of Computer Science, University of
Birmingham
2
Introduction

• The findings of an investigation into
  evolutionary art are presented
• Problematic issues with the current systems are
  identified
• The key issue of the automation of evolutionary
  art was focussed on
• A broad theoretical model is proposed for future
  development of automated evolutionary art
• Potential benefits and limitations of the
  theoretical model are considered

3
Evolutionary Art

• Several different systems explored genetic art,
  organic art, creative design systems
• Varying levels of difference, but all possess a
  common theme of interactive evolution
• Interactive evolution has many benefits but can
  represent a severe bottleneck in the system

4
Interactive Evolution

• Requires the user to determine the fitness of
  individuals within the population
• Excellent for avoiding the encoding of complex
  fitness criteria
• Allows for a variation in personal preference
• Cannot exploit the monotonous iterative
  processing capabilities of the computer
• Constrained number of images to be considered
  within selection
• Subject to user fatigue and boredom

5
Interactive Evolution - Consequences

• Extended runs from EA systems are unlikely,
  avoiding potentially interesting images
• Fatigued users likely to be inconsistent and
  contradictory in judgement

6
Overcoming Fatigue

• Possible solutions
• Farming out the task (e.g. on the web)each
  user contributes a fraction of the run
• Automation of the selection procedurefitness
  apportioned by the computer
• Both approaches previously attempted, we focus on
  automation of the procedure

7
Automation of the Evolution of Art

• We aim to understand (and hopefully mimic) the
  performance of the user
• This requires a consideration of the
  psychological and aesthetic behaviour of the user
  and can be guided by the observation of aesthetic
  choice
• Several experiments have been previously
  conducted in an attempt to isolate the basic
  features of aesthetic preference

8
Investigation of Aesthetic Preference

• We propose that an understanding of the low-level
  influences can lead to guidance of an
  evolutionary art system, avoiding issues of
  cultural influence and higher level, subjective
  factors
• In the consideration of abstract pieces, an
  emphasis is placed on form rather than content
  experiments determining potentially measurable
  elements are of interest (symmetry, regularity of
  elements etc.)

9
Factors of Aesthetic Judgement (1)

• Evidence for the preference of specific features
  and factors was examined
• Previous studies implicated several factors in
  the judgement of the aesthetic merit of form
• Principles of Pictorial Composition ( Pickford
  1972)(perception and organisation)
• Complexity factors (Berlyne 1969)
• Amount of material, regularity, symmetry, number
  of separate units, distribution.

10
Factors of Aesthetic Judgement (2)

• Certainty (Dorfman and Mc Kenna 1969)
• Preference for simple lines / shapes (Barnhart
  1960)
• Balance (Arnheim 1974)
• Possibility of conducting bespoke aesthetic
  preference experiments to determine a more
  specific selection guidance metric

11
Previous Attempts

• Neural network modelling (Balluja et. al 1994)
• NN taught to rate images, based on user data
• Results disappointing
• What was the network trying to learn? Reducing
  the factors to a single measure likely to have
  been problematic
• NEvAr (Machado et. al 2002)
• Partial automation of images
• Fitness function uses calculation of suggested
  image complexity and processing complexity
• Interesting results, however behaviour on
  extended runs unknown
• Model of aesthetic judgement interesting, but
  perhaps too simplistic how much variation would
  be present in the images created?

12
A Novel Proposal (1)

• Three stage aesthetic judgement system

• Breaking down of image in to separate components
  (emulating the discrimination of objects in the
  scene)
• Isolated features are explored, with the presence
  of particular elements determined (number of
  separate objects, regularity, size etc.) each
  value is represented as a global vector
• Presence of the derived elements leads to the
  determination of the goodness of the image the
  distance between the derived vector, and others
  representing good form give an indication of
  fitness. Alternatively a multi-objective GA could
  be employed, with the Pareto front containing the
  optimal combination of features.

13
A Novel Proposal (2)
14
Potential Benefits of Approach

• Alleviates user fatigue, allowing for extended
  runs and larger population sizes
• Not restricted to any particular method of EA
  creation (genetic art, organic art)
• Based on low-level premises, hence avoids higher
  level issues such as cultural influence

15
Potential Limitations

• Complex to build - Hard to implement crisp rules
  describing fuzzy concepts, such as certainty
• Concentration on low level features might result
  in uninteresting images sidestepping subjective
  and higher level factors might also sidestep the
  more interesting images

16
Conclusions

• A novel system for automating selection in
  evolutionary art is presented. Successful
  implementation would alleviate the troublesome
  issue of user fatigue currently experienced with
  contemporary systems.
• Ideally the system would be capable of selecting
  interesting images for reproduction. More
  realistically it is hoped that the system would
  be able to avoid universally bad images.
  Potentially this would allow extended runs.
• Several issues must be addressed before
  implementation, however, with the design based on
  the results of empirical studies the approach
  holds the promise of producing some potentially
  interesting results.

17
Nature Inspired Creative Design

•       "Designing for the 21st Century" cluster
  -- Network of scientists, artists, designers,
  and industrialists- focuses on taking ideas,
  methods, paradigms and algorithms from nature
  and introducing them into the design
  process.Nature inspired approaches have the
  potential to- create better designs- better
  design processes and- better tools.First
  workshop coming soonSee www.cercia.ac.uk/design2
  1 or talk to Thorsten