Using Artificial Life to evolve Artificial Intelligence

1
Using Artificial Life to evolve Artificial
Intelligence

• Virgil Griffith
• California Institute of Technology
• http//virgil.gr
• virgil_at_caltech.edu

Google Tech Talk - 2007
2
What is Artificial Life
Life
3
Evolution an abbrev intro

• Evolution is an algorithm
• Given only
• Variable population
• Selection
• Reproduction with occasional errors
• Regardless of substrate you get evolution!

4
Forming body plans with evolution

• Node specifies part type joint and range of
  movement
• Edges specify the joints between parts
• Population
• Graphs of nodes and edges
• Selection
• Ability to perform some task (walking jumping
  etc.)
• Mutation
• Node types change/new nodes grafted on

5
Blocky Creatures Movie
6
Using Artificial Lifeto evolveArtificial
Intelligence
7
How to model Intelligence

• Marionettes (ancient Greeks)
• Hydraulics (Descartes)
• Pulleys and gears (Industrial Revolution)
• Telephone switchboard (1930s)
• Boolean logic (1940s)
• Digital computer (1960s)
• Neural networks (1980s - )

8
Nervous Systems

• Evolution found and stuck with nervous systems
  across all levels of complexity
• Provide all behaviorsincluding anything that
  might be considered intelligencein all organisms
  more complex than plants
• Some behaviors are innate so the wiring diagram
  (the connections) must matter
• But some behaviors are learned so
  learningphenotypic plasticitymust also matter

9
(No Transcript)
10
Not to be confused with
11
What Polyworld is

• Making artificial intelligence the way Nature
  made natural intelligence
• The evolution of nervous systems in an ecology
• Working our way up the intelligence spectrum
• Research tool for evolutionary biology
  behavioral ecology cognitive science

12
What Polyworld is not

• Fully open ended
• Accurate model of microbiology
• Accurate model of any particular ecology
• though could be done
• Accurate model of any animals brain
• though could be done

13
Polyworld Overview

• Organisms have
• evolving genes and mate sexually
• a body and metabolism
• neural network brains
• initial neural wiring is genetic
• At birth all neural weights are random
• Hebbian learning refines synapse weights
  throughout lifetime
• 1-dimensional vision (like Flatland)
• No fitness function
• Fitness is determined by natural selection alone
• Critter Colors
• Red current aggression
• Blue current horniness

14
Movie - Sample
15
Body Genes

• Size
• Strength
• Max speed
• Max lifespan
• Fraction of energy given to offspring
• Greenness
• Point-mutation rate
• Number of crossover points

16
Brain Genes

• Vision
• of neurons for seeing red
• of neurons for seeing green
• of neurons for seeing blue
• of internal neural groups
• For each neural group
• of excitatory neurons
• of inhibitory neurons
• Initial bias of neurons
• Bias learning rate
• For each pair of neural groups
• Connection density for excitatory neurons
• Connection density for inhibitory neurons
• Learning rate for excitatory neurons
• Learning rate for inhibitory neurons

17
Polyworldian brain map
18
Polyworld Brain Map (actual)
19
All about Energy (Health)

• Get Energy by
• eating food pellets
• eating other Polyworldians
• Lose Energy by
• mating moving existing
• having large size or strength
• but get benefits in max-energy and fighting
• brain activity
• for computational reasons and parsimonious brain
  size

20
Behavior sample Eating
21
Behavior sample Killing Eating
22
Behavior sample Mating
23
Behavior sample Lighting
24
New Species Joggers
25
New Species Indolent Cannibals
26
Emergent Behavior Visual Response
27
Emergent Behavior Fleeing Attack
28
Foraging Grazing Swarming
29
Observations from Polyworld

• Evolution generates a wide range brain wirings
• Selection for use of vision
• Evolution of emergent behaviors

30
Ideal Free Distribution in agents with evolved
neural architectures
Early
Middle
Late
31
Predator-Prey Cycles
32
(No Transcript)
33
But is it Alive Ask Farmer Belin

• Life is a pattern in space-time rather than a
  specific material object
• Self-reproduction
• Information storage of a self-representation
• A metabolism
• Functional interactions with the environment
• The ability to evolve

Farmer Belin (1992)
34
But is it Intelligent

• No obvious way to measure intelligence
• (aka We dont know)
• even biologists have a hard time on this
• But were in a simulation that means we can use
  techniques not available to biology!
• Information theory
• Complexity theory

35
Neural Functional Complexity
36
Is there an evolutionary arrow of complexity

• Yes Darwin Lamarck Huxley Valentine
• No Lewontin Levins Gould

37
Evolution drives complexity
38
Genetic complexity over time
39
Neural Complexity Room to grow
40
Future Directions

• More
• measures of complexity
• complex environment
• food types
• agent senses (touch smell)
• Behavioral Ecology
• Optimal foraging (profit vs. predation risk)
• Evolutionary Biology
• Speciation (population isolation)
• Altruism (genetic similarity)
• Classical conditioning animal intelligence
  experiments

41
Source Code

• Source code is available!
• Runs on Mac/Linux (via Qt)
• http//www.sf.net/projects/polyworld/

42
But is this a good idea
43
Special Thanks

• Larry Yaeger
• Chris Adami

44
Plasticity in Neural Function
Function maps
The redirect
Mriganka Sur et al Science 1988 Nature 2001
45
Plasticity in Wiring
Patterns of long-range connections in V1 normal
A1 and rewired A1
Mriganka Sur et al. Nature 2001
46
Hebbian Learning Structure from Randomness
John Pearson Gerald Edelman
47
Real and Artificial Brain Maps
Distribution of orientation-selective cells in
visual cortex
48
Neuroscience Recap

• Intelligence is based in brains
• Useful brain functions are created by a
• suitable initial neural wiring
• general purpose learning mechanism
• Artificial neural networks capture key features
  of biological neural networks
• Thus we could make useful artificial neural
  systems with
• An evolving population of wiring diagrams
• Hebbian learning

49
Thanks to

• Larry Yaeger
• Chris Adami

50
What can Evolution do

• Optimization
• Traffic Lights
• Air Foil Shape
• Fuzzy Problems
• Sonar response from sunken ships versus live
  submarines
• Good for management tasks such as timetables and
  resource scheduling
• Even good for evolving learning algorithms and
  simulated organisms and behaviors

51
Neural Group Mutual Information
52
Evolution drives max complexity