Compositional pattern producing networks
-
Upload
warunikaranaweera -
Category
Documents
-
view
3 -
download
0
description
Transcript of Compositional pattern producing networks
2/22/2015 1
Compositional Pattern Producing Networks:
A Novel Abstraction of Development
Kenneth Stanley
Genetic Programming and Evolvable Machines ‘07
What is a CPPN?
• A variation of artificial neural networks
What is a CPPN?
Adopts the process of biological development and evolution to evolve images, 3D shapes, robots, etc
Development and evolution
• Development: Single cell (genotype) to functioning organs (phenotype)
• Evolution: Single-celled creatures to complex beings
Biological development: Fruit Fly
Anterior: head end
Posterior: tail end
Step 1: Identifying the position Head and tail ends are identified by the morphogen gradient
Biological development: Fruit Fly
Anterior: head end
Posterior: tail end
Anterior
Posterior
Step 2: Generating a pattern Based on the position, the body is repeatedly segmented
Biological development: Fruit Fly
Anterior: head end
Posterior: tail end
Anterior
Posterior
Anterior
Posterior
Step 3: Expressed phenotype Each segment gets symmetric pairs of body parts (e.g.: wings and legs)
CPPN encodes development
Step 1: Identifying the position - x,y pixel coordinates
Step 2: Generating a pattern - CPPN
Step 3: Expressed phenotype - Attributes of the pixel
Generating Patterns: CPPN
• Uses functions to represent patterns
– Sine function (repetition)
– Gaussian function (symmetry)
Generating Patterns: CPPN
Symmetry in fruit fly: Wings, legs
Gaussian function can model symmetry
Patterns generate more patterns
Symmetry and repitition
Final CPPN
Evolving CPPNs
• CPPN-NEAT: Neuroevolution of augmenting topologies
• User chooses parents
• Mutation: Edit weights, add edges and nodes
• Cross over: Cross two or more CPPNs
Evolving CPPNs
Generation 1
Generation 15
Generation 33
Applications
• Picbreeder (evolving 2D images): Stanley et al., CHI 2008
• EndlessForms (evolving 3D shapes): Clune et al., SIGEVOlution, 2011
• Evolving soft robots: Clune et al., SIGEVOlution, 2014
Picbreeder
EndlessForms
Thank you!