Karl Sims and the Digital Evolution (Co-)Evolution of Morphologies - - PDF document

(Co-)Evolution of Morphologies and Behavior

CPSC 607 – In Class Presentation Russel Ahmed Apu

“Karl Sims” and the Digital Evolution

• Aesthetic (Graphics)
• Complex behavior
• Natural Morphology

Brief Outline

Evolution & Co-evolution Evolution of Morphology (Plants) Evolution of Morphology and behavior Evolution based on competition

Building Blocks of Evolution

Genotype Phenotype Expression Reproduction & combination Selection

Morphology: Evolving 3D Plants

Parameter set & parameter space Mutating & Mating parameter sets Parameters:

Fractal limits Branching factor Scaling Stochastic contrubution

21 Genetic parameters

Evolution of Plants Morphology

Crossover (Random Percentage of parents) Random interpolation of parents

Morphologies and behavior

Basic Morphology (Adjacent Boxes & muscles) Neural Network based brain Automatic generation of morphology and neural

system to control muscles (GA)

Accomplish tasks (Fitness function):

• Swimming
• Walking
• Jumping
• Following

Genotype and phenotype

L – System Grammar based Recursive Mutating properties Different joint type Different phenotype

part properties

Embedded neural

structure

Sensors & Effectors

Neural Structure

Unconventional Neural structure Specialized neurons Capable of functions (not just threshold) Assumed to create interesting evolutionary

behavior

Two brain steps for each time step Effector strength proportional to cross

section area of joints

Neural Structure Example Evolving Creatures

Physics Engine Behavior Selection Measure fitness (10 sec) Reproduce the fittest Suspend sim. Time for unfit phenotypes

Swimming

No gravity Viscous resistance Straight line

swimming rewarded

Continuous

movement rewarded

Maximum distance

from COG

Walking

Gravity Static friction Fitness equals

distance traveled

Speed is

rewarded

Falling is

prevented

Jumping

Maximum height

above the ground for lowest part

Average height

during the simulation

Following

Creatures having

light sensors

Heading Light at different

location

Speed

Hurdles of artificial evolution

Flaws of physics engine Selection of proper strategy is hard Large number of phenotype for given genotype Fluctuations on small changes Demo Videos (by Rob Leclerc):

• Rob1
• Rob2
• Rob3
• Rob4

Co-Evolution based on competition

A game of occupying a

cube

Interaction between

evolving creatures

Competition Evolving strategies (I.e.

blocking opponents way) to gain control of the resource

Challenges of Co-Evolution

Fitness highly dependent on behavior Higher complexity More interesting Dependent on environmental factors Dependent on opponent morphology and

behavior

Intra and inter species competition

The contest rules (for selection)

Sim for 8 second Winner gains the most

control over the cube

Points for closing&

surrounding

Choose competition

strategy for faster result

Evolution strategy

Inappropriate creatures are removed

before contest

Number of offspring proportional to

fitness

Survivors kept Mutating graph Mating

Evolved creatures Results

Diverse collection of interesting strategies Winners alternated between species

(Strategies and counter strategies)

Adaptive behavior Counters the opponent Most successful strategy: covering with

arms

Finally, enjoy the demo…

Concluding Demo (Karl Sims)