Case Study: Genetic Algorithms

1
Case Study Genetic Algorithms

• GAs are an area of AI research
• used to solve search problems with potentially
  better performance than traditional search
  problem
• also possibly used as a form of learning
• In an AI search problem, you have a search space
• search space is the space of possible solutions,
  for instance
• in chess, it would be all possible board
  configurations
• in design it would be all of the possible
  components available to build the device along
  with all the possible ways to configure them
• often, AI problems require a brute-force search,
  but sometimes a heuristic is available to help
  guide the search
• in chess, we use a function based on the worth of
  each piece and the strategy of moving them into
  certain board locations to see if a move is worth
  pursuing or not
• in design, we might rate components based on
  their utility, cost, weight, and their
  configurations based on how easy it is to get to
  a component
• the guided search, known as heuristic search, is
  still intractable (for n moves, there are 2n
  board configurations)
• GAs are an attempt to get around that problem

2
Search Using Randomness

• Assume we are searching for the proper values to
  place into a list
• (a b c d e f) each of these represents
  something meaningful
• for instance a number of nails to use, b
  number of boards to use, c number of wall
  outlets, d number of windows, etc
• Assume we have a function that can take a vector
  (or list) and tell us approximately how good it
  is
• I can try to generate all possible combinations
  of vectors but that is intractable
• now consider many vectors are probably worthless
• if I can identify a worthless vector, I probably
  wouldnt want to use any variations of that
  vector
• similarly, if I find a good but not great vector,
  I might try variations of it
• What is a variation?
• Some minor but random change to the original
• (a b c d e f) becomes (a b d c e f) or (a b c d
  e f)
• where a is 1 greater or less than a

3
Inspired by Natural Selection

• GAs get their names because they are an attempt
  to mimic the evolutionary process of genetic
  material
• Our vector represents a chromosome
• We start with a population of chromosomes that we
  will breed
• The children will be made up of much the same
  genetic material as their parents (that is, the
  childrens chromosomes will be very similar to
  the parents chromosomes)
• but there will be some variations
• possible mutations a particular gene mutates,
  perhaps for the better, perhaps for the worse
• possible cross-overs one child might inherit
  the genes and another worse genes
• Now we use natural selection
• we pick only a subset of children, which are the
  fittest to survive, and use them to be parents of
  a new generation
• to determine the fitness of a child, we need a
  fitness function
• We iterate through many generations, hopefully
  moving our genes towards a solution guided by our
  fitness function

4
General Procedure for GAs
Starting with a base population of parents,
mutate them into children, rate the children with
the fitness function, select from the children
those to go on to the next generation Notice
that they take efforts to make copies, these
steps can be skipped if we dont mind working
with the original parents
5
Requirements

• We need to be able to express our problem as a
  vector
• the vector will be the set of attributes whose
  values we want to assign or learn
• consider that we want to make a better chocolate
  chip cookie, our vector might include the amount
  of flour, sugar, salt, water and chocolate chips
  that we will put into our mixture
• We need a fitness function to evaluate the
  children
• for a cookie, we might ask humans to eat the
  cookies and rate them, this of course means that
  our evolutionary process is greatly slowed down
• We need to know how to mutate our parents into
  new children
• cross-over, point mutation, inversion
• We need to know how many children to generate
• such decisions might dictate how quickly or
  slowly we reach a solution

6
Vectors

• Vectors represent features (attributes)
• they could be binary values
• (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)
• vectors might store integer values
• vectors might contain a mixture of values, in
  which case we have to be careful when doing
  things like cross-over
• 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)

7
Natural Selection Methods

• Evolutionary mechanisms are
• Inversion - moving around features in the vector
  such as reversing 3 features
• Point 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 using two parents and swapping
  portions of their two chromosomes
• The choice of which mechanism to use is made
  randomly, and the choice of how/where to apply it
  is made randomly

• Natural Selection mechanisms include
• 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)

8
Choices/Diversity

• How many vectors should make up the population of
  one generation?
• 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?

• 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