Finding aesthetic pleasure on the edge of chaos: A proposal for robotic creativity

1
Finding aesthetic pleasure on the edge of
chaosA proposal for robotic creativity

• Ron Chrisley
• COGS
• Department of Informatics
• University of Sussex

Workshop on Computational Models of Creativity in
the Arts Goldsmiths College, May 16th-17th 2006
2
Background

• Goal Design a robot/environment system likely
  to exhibit creative behaviour
• Novel (at least for the robot)
• Of (aesthetic) value (for humans, if possible)
• Engineering approach
• No direct modelling of human creativity
• But exploit what is known about creativity in
  humans (and animals?), when expedient
• Allow for possibility that insights into the
  human case may accrue anyway
• Manifesto only No implementation yet
• Set of "axioms"
• Assume case of musical output for examples

3
Underlying architecture
4
Underlying architecture

• CNM
• Recurrent neural network
• Forward model of environment
• Learns to anticipate/predict the sensory input it
  will receive if it performs a given action in a
  given context
• In conjunction with motivators can enable the
  robot to select actions that carry an expectation
  of "pleasure"

5
Main idea

• Add new motivators, corresponding to two
  dimensions of creativity
• Value
• Novelty
• Axiom 1 If you make your robot
  pleasure-seeking, and make creativity
  pleasurable, you'll make your robot creative

6
Value Appreciation

• Axiom 2 To be a good creator, it helps to be an
  appreciator
• The CNM should evaluate the output of itself and
  others
• That is, it should be able to feel pleasure upon
  experiencing outputs
• Use this to guide its creative process (action
  selection)

7
Value Reality

• Axiom 3 Let the robot experience output in the
  real world, as we do
• Avoids the input bottleneck
• Robot can learn all the time
• Learns reality, not our edited version of it
• Increases likelihood of consonance between what
  we value and what it values

8
Value In our image

• Axiom 4 We wont like what it likes unless it
  likes what we like
• Built-in motivators should resemble ours
• E.g., a preference for integer frequency ratios

9
Value Sociability

• Axiom 5 An important motivator is the approval
  or attention of others
• Indirect Preference for human proximity/input
• Direct Buttons on robot that allow listeners to
  provide approval or disapproval feedback

10

• From Saunders, 2001

11
Novelty Complexity

• Axiom 6 Sometimes it is better not to try
  pursue novelty directly, but something that is
  correlated with it
• Prefer outputs on the subjective "edge of chaos"
  That almost, but not quite, elude understanding
  of that agent at that time
• Pleasure of an output is a hump-shaped function
  of the effort required to predict it
• Result Sing-song and white noise are boring, but
  catchy tunes are not

12
Novelty Dynamics

• Axiom 7 Let dynamics play a role in
  appreciation
• Process is temporally sensitive in several ways
• Pleasure associated with "getting it" depends on
  how much time it took to get there
• Even if earlier portions are unpredictable (gt
  not pleasurable), work as a whole can be if it is
  coherent
• Since the system learns, what it finds
  challenging, but possible, to predict (
  pleasurable) will change over time

13
Novelty Self-appreciation

• Axiom 8 Patterns in one's own states can be the
  objects of appreciation
• Will only be a path to novelty if agent has
  limited access to its own processes
• Can only change internal states indirectly, by
  changing world
• Uses model of its processes to predict its own
  behaviour, rather than using those very processes
  themselves

14
Novelty Embodiment

• Axiom 9 The best way to make outputs in the
  real world is to be embodied in the real world
• Avoids the output bottleneck
• Robot doesnt require intervention for it to
  generate and appreciate
• Allows for serendipity, in the space between
  expected and actual outcomes
• Imposes naturalness relation, making some
  transitions non-arbitrary (value)

15
Implementation issues

• Intended platform
• Two AIBO ERS-7s

• Problem
• Disembodied sound generation

• Solution
• Translate bodily movements into sound

16
Thank you!

• Thanks to
• Maggie Boden
• Rob Clowes
• Simon Colton
• Jon Rowe
• Rob Saunders
• Aaron Sloman
• Dustin Stokes
• Mitchell Whitelaw
• for helpful comments and discussions