Artificial Intelligence: Advanced Search and Machine Learning

1
Artificial Intelligence Advanced Search and
Machine Learning

• Lecture 14

• Learning - Intro
• Learning and Search
• Genetic Algorithms

2
Learning Intro

• Another fundamental aspect of human intelligence
  is the ability to learn.
• A system that can learn can adapt to change, and
  develop more sophisticated rules/reasoning with
  experience.
• May also be easier to program - start off simple,
  but give program ability to learn and improve.

3
Learning - Intro

• Many types of learning, e.g.,
• Learning from example (inductive learning).
• Give someone pictures of 10 different tigers..
• Learning by being told
• Tell someone A tiger has a long tail, stripes,
  short ears,
• Learning by analogy
• A tiger is like a big cat with orange stripes
  and big teeth
• Learning by discovery.
• AI focuses on learning from examples (inductive
  learning)

4
Inductive Learning

• Goal is to find some rule or function that lets
  you draw useful conclusions, e.g.,
• Rule letting you determine student performace
  from entry qualifications.
• Rule letting you classify images (into
  widget/wodget) given some features.
• Rule letting you determine whether a word in a
  sentence is a noun or a verb.
• Provide learning system with data/cases where
  result is known..
• Student results, manually classified images..
• Rule should work for new cases.

5
Classification

• Many learning problems can be expressed as
  classification problems.
• Given various input features (e.g., size of
  teeth, stripyness..) goal is to decide which of a
  number of classes (or output categories) the
  case falls into (e.g., tiger vs rabbit).
• Consider
• character recognition
• disease diagnosis
• student performance prediction.
• All classification tasks.

6
Classification

• One way to formalise it as looking for a function
  that maps features to categories
• f(feature1, feature2, .) --gt Category
• More informally, if we have features such as
• teethbig, covering stripy
• we want a rule allowing us to conclude that
  this example is a tiger.
• We start with cases where we know the right
  classification.

7
Learning and Search

• As a search problem, we are looking to search
  through all possible rules/functions, and see
  which best accounts for our example data.
• Ie, given data on 10 tigers and non tigers,
  search through all possible rules and find the
  one that (most) correctly classifies all the
  examples.
• Want to find the simplest such rule that covers
  the example data.

8
Example

• Data
• Assume rule is just a conjunction of required
  features (e.g., stripy AND longTail -gt tiger)
• Which is the simplest such that predicts tigers
  given this data?

Positive example
Negative example
9
Searching the rule space

• How do we search through all possible rules to
  find best one?
• In this (much simplified) example we could just
  go through all the possible rules
  systematically..
• But in general the search space is very large
  and this is impractical.
• So inductive learning is partly about clever ways
  of managing search for possible rules.

10
Genetic Algorithms

• One novel search approach is the use of genetic
  algorithms.
• Normal search procedures have some way of
  creating successor node from existing node in
  the search tree. The new node is sometimes called
  a child node, and the dominating one the
  parent.
• Genetic algorithms extend this to allow two
  parents! The child has features of both parents.

11
Genetic Algorithms (GAs)

• To further the biological analogy, nodes in
  genetic agorithms are (sometimes) called genomes
  or chromosomes.
• A genome represents a possible solution to the
  problem. We are looking for the genome with the
  best performance..
• Algorithm to find this involves maintaining a
  population of genomes, killing off bad ones,
  creating offspring, and iterating until best
  solution found. (cf survival of fittest).

12
Genomes

• Genomes (representing possible solutions to a
  problem) are often represented as string, often a
  bit string (ie, sequence of 0s and 1s).
• Consider the following problem
• You are trying to decide which books to buy for
  your course. There are 8 possible books. Some
  books cover more than one module. Some books cost
  more than others
• We could represent a possible solution as a bit
  string where each bit corresponds to one book.
• 10010111 (buy first book, dont buy second ..)

13
Fitness function

• Genetic Algorithms need a way of evaluating how
  good a solution is.
• We need some function assigning a score to a
  potential solution.
• For books problem we could invent some scoring
  system based on cost of books, and how well the
  books covered all modules.

14
Genetic Algorithms

• Given some representation of a candidate
  solution, and some fitness function, the
  algorithm is..
• Start with some set S of candidate solutions.
• While S doesnt contain an acceptable soln
• Score each solution in S using fitness fn.
• Select pairs of solutions from S, favouring pairs
  with high scores.
• Produce offspring from these pairs.
• Replace weakest candidate in S with these
  offspring.

15
Offspring..

• Given two parent solutions, how do we produce
  offspring?
• Cross-over and mutate genomes..
• Cross-over take first half of first genome,
  second half of second one (and vice versa) to
  produce two offspring.e.g.,
• parents
  children
• 11101000 11100011
• 10010011 10011000
• Mutation randomly change one element in the
  string. E.g., 11100011 --gt 11000011

16
Example

• For book problem
• Start with a population of possible solutions
• say 10001000, 11101111, 10101010, 11001100
• Score, find best (say, 2nd and 3rd)
• Cross-over to give 11101010, 10101111
• Replace weakest with these, to give population
• 11101111, 10101010, 11101010, 10101111
• Continue until good solution found.

17
Genetic Algorithms and Inductive Learning

• We can apply this to machine learning
• We represent a possible classification rule as a
  string, where each position in string corresponds
  to one feature
• e.g., YN? big teeth, not stripy, dont care
  about the tail.
• Use a fitness function that reflects how good the
  proposed rule is to cover the example data.
• YN? Only gets one example right (the stripy
  animal which isnt a tiger), score1.

18
GAs and Learning

• We have a population of candidate solutions..
  YYN, ??Y, YN? Etc.
• Score, cross-over, mutate etc, until (hopefully)
  we end up with YY? (ie, big teeth and stripy,
  dont care about the tail).
• We hope that this solution will correctly predict
  other examples!

19
GAs and Learning

• GAs very effective for certain kinds of problem
  (e.g., timetabling)
• Used widely for all sorts of problems.
• Still somewhat unfocused as a learning technique.
• Next lecture will look at other techniques.

20
Summary

• Inductive Learning is about finding general rules
  from examples.
• Often used for classification tasks - what kind
  of thing is it, given some features?
• Learning involves search - search through all
  possible rules.
• One approach is to use Genetic Algorithms
• advanced search procedure that uses two parent
  solutions and produces offspring using
  cross-over, and maintains pool of possible good
  solutions.