Biomorphic Computing

1
Biomorphic Computing

• Professor Bill Tomlinson
• Tuesday 200-450pm
• Winter 2004
• CS 189

2
Week 2 Genetic Algorithms

• News
• Last Weeks assignment
• Natural Evolution
• Synthetic Evolution ( Sims reading)
• Assignment 2
• Lab Time

3
Go over Game of Life assignment

• Problems?

4
Go over Sims reading

• fixed-length bit strings vs. lisp expressions
• parameter spaces
• mutating and mating symbolic expressions

5
Natural Evolution

• Descent with modification
• The central process by which the abundance of
  life forms have come to exist.

6
Linnaean Classification

• Carolus Linnaeus, 1707-1787
• Initially based on visible, behavioral and other
  obvious characteristics.
• Kingdom, Phylum, Class, Order, Family, Genus,
  Species (pick your favorite mnemonic device)

7
Defining the species

• How do you tell whether two things are from the
  same species?

8
Defining the species

• Reproductive isolation (produce viable offspring)
• Morphological similarity
• Genetic similarity
• Ecological similarity

9
Questioning the species concept

• DNA studies are calling some of these
  classifications into question.
• Organisms that reproduce asexually - self-cloning
  bacteria
• Ring species
• Species over time
• Based on human decisions about classification
• A conceptually useful mechanism to work with the
  uneven genotypic/phenotypic distribution

10
Charles Darwin

• 1809-1882
• Voyage of the Beagle to the Galapagos
• Finches
• Origin of Species - 1859
• Expression of Emotion in Man and Animals - 1872

11
Darwin vs. Lamarck

• How does a giraffe get its long neck?
• Lamarck (1744-1829)

12
Mendelian Genetics

• Gregor Mendel
• 1822-1884
• Dominant vs. recessive
• Genotype vs. phenotype
• (expression)

13
Punctuated Equilibria

• Niles Eldredge
• Stephen Jay Gould (1941-2002)
• long periods of stability with abrupt periods of
  change and appearances of new species
• splitting off of daughter species (cladogenesis),
  rather than transformation of species as a whole
  (anagenesis)
• Founder effects

14
Mitosis vs. Meiosis

• Mitosis one cell
• becomes two cells
• with the same DNA
• Meiosis one cell
• becomes four cells
• with one strand each

15
Crossing Over

• Meiosis -gt production of germ cells
• Parts of two
• chromosomes
• get swapped.
• Also called recombination

16
Mutation

• Occasional misfirings of the replication process.
• Almost always harmful.
• On occasion,
• very successful.

17
Differential Survival

• Once there is variability (through sexual
  reproduction, crossover and mutation) in a
  population, the environment culls some while
  others survive.
• Natural selection.

18
Biological Diversity

• Simple mechanism
• of natural selection
• enables the propagation
• of vastly different
• organisms.

19
Convergent Evolution

• Same selection pressures cause similar
  morphology/behavior/etc. in different species.
• Other examples?

20
From Biology to Computation

• From natural selection to evolutionary computation

21
Simulating evolution

• Descent with modification
• Fitness landscape determines differential survival

22
Different Branches of Evolutionary Computation

• Genetic Algorithms
• Genetic Programming
• Evolutionary Strategies
• Evolutionary Programming
• etc.

23
Evolved Virtual Creatures

• Karl Sims

24
What problems are evolutionary techniques good
for?

• Discuss

25
What problems are evolutionary techniques good
for?

• Problems in which the solution can be
  characterized in terms of variables (or human
  feedback)
• Overcoming fitness landscapes with many local
  optima (i.e., where gradient descent wont work).

26
Shortcomings

• Computationally expensive

27
Applications Artificial Life

• Tierra
• Synthetic
• organisms
• compete
• for time
• and
• computer
• memory.

28
Applications Biocomputing

• Protein Folding
• RNA Folding
• Sequence alignment

29
Applications Evolvable Hardware

• Golem Project _at_ Brandeis (movie)

30
Applications Coordinating Groups of Robots

• Evolving teams for reconnaissance, terrain
  mapping, etc. - UCF/GMU

31
Applications Game Theory

• Prisoners Dilemma - Axelrod

32
Applications Art

• Karl Sims

33
Genetic Algorithms

• Project ideas?

34
Break
35
Assignment 2

• Implement a genetic algorithm to evolve a
  creature that can go as fast as possible to a
  point where you click.

36
Representation of a Creature

• Ability to sense and take action
• Conditional bits - perception
• Action bits - motor

37
Conditional Bits

• How are conditional bits hooked up to the sensory
  world?
• Dot product and cross product offer one way to
  slice perceptual world into chunks.

38
Sensor radius

• How far around them can they see?
• How do they detect the target?

39
Action Bits

• How do action bits allow the creature to move
  around or otherwise affect its world?

40
How a rule works.

• Rule 010010
• First four bits are perception
• e.g., front-left, back-left, back-right,
  front-right
• Last two bits are action
• go-straight-or-turn, if-turn-which-direction
• So this bitstring might mean If I sense
  something in back left quadrant, turn left.

41
Number of rules

• How many do you need?
• How many is too many?

42
The Bitstring Genome

• all the bits of all the rules
• equivalent of DNA
• Total length bits/rule rules.

43
Populations

• Genetically Diverse
• Provide the raw materials for survival of the
  fittest.

44
Initialize population

• Random
• Seeded

45
Evolving the creatures

• Initialize population
• Evaluate population
• Generate new population
• Loop between evaluating population and generating
  new one.

46
Population size

• Large enough for genetic diversity
• Small enough for computational speed

47
Evaluate Population

• Load population into simulated world (reset)
• Fitness criteria?

48
number of creatures loaded in

• can just be one, or can be several (take average
  of fitness)

49
Generate new population

• Find best (mom) and second best (dad) (or some
  other combination of top performers)
• Crossing over
• Mutation

50
Crossover rate

• How much recombination?

51
Mutation rate

• How often to change a random bit?

52
Number of time steps

• Too short and its hard to learn
• Too long and success is trivial

53
Number of generations

• Genetic change occurs in small stages over many
  generations.

54
Assignment 2

• Distribute Assignment sheet
• Distribute source code

55
The Files

• GAMain
• Creature
• Evolver
• TestWorld
• GridWorld
• Vec2

56
What needs to be written?

• Evolver. generateNewPopulation()
• Creature.sense()
• Creature.chooseAction()
• Creature.takeAction()
• TestWorld. testFitness()

57
Lab Time

• All please log into a computer
• If you dont have an account, please get one