Learning

1
Learning

• A big criticism of AI has been that machines are
  unable to adapt to new situations or unprogrammed
  circumstances
• Learning - ability to adapt to new situation or
  ability to adapt internal representations when
  faced with something new
• Skill refinement ?? Knowledge acquisition

2
Rote Learning

• Simplest form of learning and also the weakest
  and least useful form
• Occurs any time a new piece of information is
  stored
• Caching is a form of Rote Learning
• Need a means of organizing new info
• Some form of Generalization would be helpful

3
Advice Taking

• Using advice to alter processes
• In Hearts Try not to take points
• In Chess Try to control center of board
• Need a way to operationalize the advice
• Translate advice into actual code
• Unfold advice into more detail as needed
• A useful form of learning if general knowledge is
  being provided by a teacher/trainer

4
Example of Unfolding Advice

• See Lisp code on page 451
• Start with me avoid taking points
• Do not take points during the trick
• While each player is playing a card, do not be
  the winner if there are points
• While each player is playing a card, when you
  play your card, if there are already points
  played, play a card lower than at least one
  played so that you are not the winner of the
  trick but if the trick does not have points, you
  can be the winner by playing a higher card

5
Learning in Problem Solving

• One of the most common forms of learning in AI is
  to improve the systems performance as it solve
  problems by fixing any problems that arise
• Parameter Adjustment
• as new cases are seen, adjust the parameters to
  be able to handle new case
• Macro-Operators
• if a new sequence of operations leads to a
  solution, compile them into a macro-operator
• Chunking
• similar but contains a wrapper around it that
  designates when the chunk is useful

6
Induction

• Learning how to better categorize items
• Often performed during classification
• Idea - we categorize based on features
• We need to learn what features should be found
  and what features should be found for an item to
  be categorized into a given class
• Induction altering a set of features for a given
  category, positive features that should be in all
  instances of the class, and negative features
  that should be in no instances of the class
• Winstons Learning Program
• Version Spaces
• Decision Trees

7
Winstons Learning Program

• Start from a basic structural description of an
  object
• Examine known instances of the object, generalize
  to include them
• Examine near misses, restrict description to
  exclude them
• Generalize add links, generalizing nodes
• Restriction alter links

8
Mitchells Version Spaces

• Maintains a set of descriptions
• Given a representational language of the features
  and values in the space, and a set of positive
  and negative examples, form a set of descriptions
  making up the concept space
• Use a process of candidate elimination to narrow
  the set Everything into the concept space

9
Candidate Elimination

• Initialize G to null (everything)
• initialize S to first positive example
• If positive, remove from G descriptions that do
  not cover the example. Update S to contain the
  most specific set of descriptions that cover the
  example
• If negative, remove from S any descriptions that
  cover the example. Update G to contain the most
  general set of descriptions that do not cover the
  example
• Repeat until S G converge to the same singleton
  set

10
Decision Trees

• Attributes of the objects are distributed as a
  tree
• Descend the tree, going down positive or negative
  branches based on whether the attribute is
  present for the current example
• If a leaf node is reached and it is not the
  example, add a new node and branch

11
Explanation-Based Learning

• Learning from a single example accompanied by an
  explanation
• EBL requires
• Training Example
• A Goal Concept
• An Operationality Criterion
• A Domain Theory
• The system then generalizes its representation

12
Example of EBL Chess Fork

• Training Example a representation describing a
  situation where a knight can take two pieces of
  high value
• Goal Concept that this is a good situation to
  or strategy if you are in the position to perform
  a fork, but a bad situation to fall into
• Operationality Criteria concepts that are
  usable for this situation are that the knight be
  moved to a position where it cannot be captured,
  where the knight has the ability to capture two
  pieces of high value
• Domain Theory the rules of the knights
  movement and how it captures pieces, the rules to
  assign values to each piece, and a rule that
  states that the pieces threatened by the knight
  are of high enough value

13
Discovery

• AM Deriving new mathematical theorems
• Used a variety of techniques including Generate
  and Test and heuristic search with 250 heuristic
  rules
• Example generate a new concept if some element
  of B are in A but not all elements of B (i.e. why
  is B not A?)
• One concept was of divisors for numbers
• AM used this to learn about prime numbers and
  squares

14
Other Discovery Systems

• BACON - data-driven discovery
• Given data, analyze it for concepts within the
  domain
• Was able to derive the ideal gas law from data
  relating the variable values in the equation
  (pV/nT8.32)
• AUTOCLASS - clustering of classes
• Learned new classes of stars from infrared
  spectral data

15
Analogy

• Transformational Analogy
• Transforming a previous solution into a new
  solution
• e.g. mathematical problems -- converting a proof
  for line segment lengths to angles
• Derivational Analogy
• Using the how a solution worked to solve a
  different problem
• e.g. converting a sort program from Pascal to Lisp

16
Neural Net and Genetic Algs.

• Instead of writing a program to manipulate
  internal representations upon new examples, have
  the computer alter its representation
• Both methods use a variety of mathematical
  algorithms to perform this manipulation
• We will look at Neural Nets in Ch 18.