ALife

1
ALife

• Artificial Life - just as AI is the pursuit of
  using computation to explore intelligence, ALIFE
  is the pursuit of using computation to explore
  biological life and simulate it electronically
• ALIFE-AI Claim The dumbest smart thing you can
  do is stay alive
• Alife is the lower bound for AI

2
Life Comes Before Intelligence

• This idea (or slogan) is being promoted by those
  who feel that Alife must precede AI
• There are many missing pieces to the AI picture,
  perhaps Alife can fill them in
• Missing common sense
• Missing learning abilities
• Missing intentionality
• Missing emotions
• Missing spirit and soul

3
Is life a new missing piece?

• Is Alife just some new piece of software?
• Is Alife something that emerges from an
  organization of physical moving parts?
• If the latter, then perhaps for AI to succeed,
  computers need more elements found in life, such
  as motion, perception and elements found in
  organic chemistry
• Alife investigates processes involved in
  self-replication, evolution, symbiosis as modeled
  in software

4
Forms of Alife

• Predominantly Evolutionary techniques
• Also Neural Networks to some extent
• Other forms of Learning
• Perception
• Vehicle Control
• Adaptive Systems
• Modeling the environment

5
Evolutionary Computing

• Using natural selection as a basis for techniques
  applied to learning in problem solving
• Three common approaches
• Genetic Algorithms
• Genetic Programming
• Evolutionary Programming

6
Genetic Algorithms

• Learning through manipulation of a feature space
• The state is a vector representing which features
  are present and absent
• Supervised learning requiring a method of
  determining how good a given feature vector is
• Binary vector - an associated feature is present
  or absent (using a 1 or a 0)
• Multi-valued vector - each feature is represented
  by an integer, real or other non-binary value

7
Example of a binary vector

• A vector may be represented as
• (a, b, c, d, e) where afever?, bsoreness?,
  cachyeyes? drunny nose?, ecoughing?
• We might wish to learn a concept like flu
• start with (0, 0, 0, 0, 0) (no symptoms) and
  learn that this is not the flu
• eventually we may learn that flu can be
  represented by the vector (1, 1, 1, ?, ?) (?
  means dont care what the value is)
• Allergies might be represented as (?, 0, ?, 1, 1)

8
Multi-valued Vector

• Features may include values pertaining to height,
  weight, color, etc
• Consider an animal vector (a, b, c, d, e, f)
  where aheight, bweight, c of legs, dcolor,
  eshape and ftail?)
• Elephant(4, 2000 lbs, 4, grey, round, y)
• Mouse(3, 2 lbs, 4, grey, round, y)
• Dog(1, 20 lbs, 4, varies, long, y)

9
Learning in GAs

• The idea behind learning a vector space is to use
  prior vectors and generate a new generation of
  vectors using natural selection techniques
• This is a search problem (searching for the ideal
  or optimal vector)
• The natural selection techniques will (hopefully)
  improve the performance of the search
• This form of learning can be used to learn
  recognition knowl., control knowl., planning and
  design knowl., diagnostic knowl.

10
Fitness Function

• This form of learning is supervised (requires
  feedback)
• Feedback provided via a fitness function
• Given a vector V, apply the function f(V)
• Use this value to determine this vectors worth
  towards the next generation
• A vector that is highly rated may be used in
  forming the next generation of vectors
• A vector that is lowly rated will probably not be
  used (unless randomly selected)

11
Natural Selection Methods

• Think of the vector as a chromosome
• Standard methods are
• inversion - moving around features in the vector
  such as reversing 3 features
• mutation - changing a features value to another
  value (in binary vectors, simply complementing
  the bit, in multi-valued vectors, requires random
  selection of a new value)
• crossover (requires two chromosomes, randomly
  swap some portion of the two vectors)

12
Selection Methods

• How do you determine which vectors to
  alter/mutate?
• Fitness Ranking - use a fitness function to
  select the best available vector (or vectors) and
  use it (them)
• Rank Method - use the fitness function but do not
  select the best, use probabilities instead
• Random Selection - in addition to the top
  vector(s), some approaches randomly select some
  number of vectors from the remaining, lesser
  ranked ones
• Diversity - determine which vectors are the most
  diverse from the top ranked one(s) and select it
  (them)

13
General Procedure for GAs
Unrated Fitness Rated
Population Function Population Mutation S
election Operations Function Population
of Reproduction Population of Parent
Copies Function Parents
14
The Evolution Process

• Create an initial population of one vector
• Mutate one or more genes producing a new
  offspring
• Mate one or more vectors
• Add mutated and offspring vectors to current
  population
• Create a new generation by keeping the best along
  with randomly selected individuals

15
Choices

• How many vectors for the population?
• If too low, vectors will be the same or similar
  to previous generations, if too high, computation
  time may be too long
• What is the mutation rate?
• If too low, changes will occur infrequently, if
  too high, there is no guided search
• Is mating allowed?
• Are duplicate vectors allowed?

16
Another Ranking Method

• Instead of selecting the best vectors based on
  the fitness function, use the fitness function to
  rank the vectors and then
• Sort the vectors on ranking
• Assign a high probability to the highest ranked
• The probability of the second is (1-p)f where p
  is the probability of the first and f is a
  predetermined fitness for the second item
• Compute the remaining probabilities similarly
• Select a candidate by randomly producing a
  probability

17
Diversity

• Based on the idea that diversity helps promote
  survival, it might also be reasonable to select
  vectors which are most diverse from other
  selections
• Select the first vector(s) for the new generation
  using the rank or fitness method
• Select the remaining vector(s) by finding one(s)
  most diverse with those already chosen

18
Genetic Programming

• Instead of fixed-length vectors as forms of
  representation, use dynamic representations (e.g.
  trees)
• Trees are formed out of syntactic parses of
  programs
• Evolutionary methods are then used to adapt a
  program into other forms

19
Adapting Programs

• Crossover swaps one part of a program with
  another
• Since the parse tree represents a syntactic unit
  as a subtree, swapping one subtree with another
  will leave both subtrees syntactically correct
• Therefore, crossover will not introduce errors
• Other adaptions (mutation, inversion) must
  maintain the syntactic correctness if used
• A fitness function is then used to determine if
  the new program behaves better or worse

20
Evolutionary Process

• Start with an initial population of first-order
  calculus terms (Horn clauses) (or use some
  provided by the user)
• Apply crossover techniques
• Use positive and negative examples and a weighted
  average as a fitness function to determine how
  well the given clauses define the concept
• Retain the best for the next generation and
  discard the remainder

21
Evolutionary Programming

• Unlike Genetic Programming, Evolution Programming
  takes a single program and mutates it using
  randomness
• On average, the child is like the parent with
  some distinctiveness
• The difference is determined by gaussian
  probabilities using a normal bell curve
• Different types of mutations are used based on
  the given syntactic structure of the program