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

“Karl Sims” and the Digital Evolution (Co-)Evolution of Morphologies and Behavior • Aesthetic (Graphics) • Complex behavior • Natural Morphology CPSC 607 – In Class Presentation Russel Ahmed Apu Building Blocks of Evolution Brief Outline � Evolution & Co-evolution � Evolution of Morphology (Plants) � Evolution of Morphology and behavior Genotype Expression Phenotype Selection � Evolution based on competition Reproduction & combination Morphology: Evolving 3D Plants Evolution of Plants Morphology � Parameter set & parameter space � Crossover (Random Percentage of parents) � Mutating & Mating parameter sets � Random interpolation of parents � Parameters: � Fractal limits � Branching factor � Scaling � Stochastic contrubution � 21 Genetic parameters

Genotype and phenotype Morphologies and behavior � L – System � Basic Morphology (Adjacent Boxes & muscles) � Grammar based � Neural Network based brain � Recursive � Automatic generation of morphology and neural � Mutating properties system to control muscles (GA) � Different joint type � Accomplish tasks (Fitness function): � Different phenotype part properties Swimming � � Embedded neural Walking � structure Jumping � � Sensors & Effectors Following � Neural Structure Neural Structure Example � 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 Evolving Creatures Swimming � Physics Engine � No gravity � Viscous resistance � Behavior Selection � Straight line � Measure fitness (10 sec) swimming rewarded � Reproduce the fittest � Continuous � Suspend sim. Time for unfit phenotypes movement rewarded � Maximum distance from COG

Walking Jumping � Gravity � Maximum height above the ground � Static friction for lowest part � Fitness equals distance traveled � Average height � Speed is during the rewarded simulation � Falling is prevented Following Hurdles of artificial evolution � Creatures having � Flaws of physics engine � Selection of proper strategy is hard light sensors � Large number of phenotype for given genotype � Heading � Fluctuations on small changes � Light at different � Demo Videos (by Rob Leclerc): location Rob1 � Rob2 � Speed � Rob3 � Rob4 � Co-Evolution based on competition Challenges of Co-Evolution � A game of occupying a � Fitness highly dependent on behavior cube � Higher complexity � Interaction between � More interesting evolving creatures � Competition � Dependent on environmental factors � Evolving strategies (I.e. � Dependent on opponent morphology and blocking opponents way) behavior to gain control of the � Intra and inter species competition resource

The contest rules (for selection) Evolution strategy � Inappropriate creatures are removed � Sim for 8 second before contest � Winner gains the most control over the cube � Number of offspring proportional to � Points for closing& fitness surrounding � Survivors kept � Choose competition � Mutating graph strategy for faster result � 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)