Learning by Explaining Experience

• Explanation principle: if you want to understand a concept, try to explain it to someone else.

• By doing exercises

• e.g. MACBETH û works on precis of Shakespeare's play

  • matches an exercise description to a precedent description, then transfer explanation from the precedent to the exercise.

  • construct an antecedent-consequent rule to perform explanation-based learning.

  • knowing a lot helps you to learn more.

• Example: (exercise and precedent)
  Greed (exercise)
  This is an exercise about a weak noble and a greedy woman. The noble is married to the woman. Explain why the noble is likely to want to be king.
  Macbeth (precedent)
  This is a story about Macbeth, Lady Macbeth, Duncan, and Macduff. Macbeth is an evil noble, and Lady Macbeth is a greedy, ambitious woman. Duncan is a king, and Macduff is a noble. Lady Macbeth persuades Macbeth to want to be king because she is greedy. She is able to influence him because he is married to her and because he is weak. Macbeth murders Duncan with a knife. Macbeth murders Duncan because Macbeth wants to be king, because Macbeth is evil, and because Duncan is king. Then Lady Macbeth kills herself. Finally Macduff gets angry and kills Macbeth because Macbeth murdered Duncan and because Macduff is loyal to Duncan.

• Representation of the precedent:

  
  

  • to further describe the link "murder", treat murder as a node 0 reification (reified link) [treating something abstract and difficult to talk about as though it were concrete and easy to talk about]

• Explanation transfer:

• based on assumption: if two situations are similar in some respects, then they are likely to be similar in other respects as well.

• "Cause" links in a precedent situation are likely to identify links that can serve as a basis for justification or prediction.

• Example:

  • Precedent: in semantic net form:

    

  • Exercise

Explanation Template: consisting of the "cause" link in the precedent and all the links that the "cause" links tie together.

• determine how the people in the exercise correspond to the people in the precedent.
  [noble ½ Macbeth; his wife ½ Lady Macbeth]

• transfer the explanation templates from Macbeth to Greedy.

  

• When a single precedent cannot supply the total explanation template needed, MACBETH attempts to chain several explanation templates.
  e.g. to establish that the noble is weak, using a second precedent, Linda and Dick story.

• Example:
  Domination (exercise)
  This is an exercise about a noble and a domineering greedy woman. The noble is married to the woman. Show that the noble is likely to want to be king.
  Linda and Dick (precedent)
  This is a story about Linda and Dick. Linda is a woman and Dick is a man. Dick is married to Linda. Dick is weak because he is tired. He is tired because Linda is domineering.

Generating Rule-like Principles

• the explanation templates essentially supply an "And tree" that span the gap between the relation or action to be explained and that already known.

  root node: relation or action to be explained

  leaf node: relation or action already in place

  in between nodes: links joined by cause relations in the explanation templates.

• Can use the And-tree to generate new description.

  • the new description is called a recollection (because it involves bits of many explanation templates)

To summarize, to learn using MACBETH,

• find the And tree implied by the explanation templates used to work the exercise.

• Using the climb-tree induction heuristic, generalize the agents and objects that appear in the And tree.

• Build a new description using the links that appear in the And tree and the generalized agents and objects.

To generate rule-based description:

• label the root node of recollection with "then",

• label the leaf node with "if".

  
  

• The rules obtained:
  The Want-to-Be-King rule
  If ?noble is weak
  ?noble is married to ?woman
  ?woman is greedy
  then ?noble is likely to want to be king

Winston's Principle of Parallel Evolution

• The longer two situations have been evolving in the same way, the more likely they are to continue to evolve in the same way.

• Given a set of useful looking precedents, use the one with an explanation templates that matches the largest number of "cause" links in the exercise.

• For example (domination and fatique)
  Domination and Fatique
  This is an exercise about a noble and a greedy woman. The noble is married to the woman. The noble is tired because the woman is domineering. Show that the noble is likely to want to be king.

Learning About From and Function

• learns about objects look like from functional descriptions and descriptions of particular examples.

  • require a physical example, otherwise, there would be no way to know which precedents are relevant.

  • require precedents: otherwise, there would be no way to know which aspects of the physical example are relevant.

• e.g Cup examples.
  A Cup Description
  This is a description of an object. The object is a cup because it is stable and because it enables drinking.
  A Particular Object
  This is an exercise about a light object that is made of procelain. The object has a decoration, a concavity, and a handle. The object's bottom is flat. Show that the object is a cup.

• use of precedents
  A Brick
  This is a description of a brick. The brick is stable because the brick's bottom is flat. The brick is heavy.
  A Glass
  This is a description of a glass. The glass enables drinking because the glass carries liquids and because the glass is liftable. The glass is pretty.
  A Briefcase
  This is a description of a briefcase. The briefcase is liftable because it has a handle and because it is light. The briefcase is useful because it is a portable container for papers.
  A Bowl
  This is a description of a bowl. The bowl carries liquids because it has a concavity. The bowl contains cherry soup.

The final rule:
If ?object has a flat bottom
?object has a concavity
?object is light
?object has a handle
Then ?object is a cup

Matching

• Matching between objects in exercise and objects in precedents is required for this explanation-based learning.

• matching should be driven by the purpose of the match: so that it is immune to irrelevant properties.

The matching problem can be reduced to backward chaining:

• treat the precedent as though it were a source of antecedent-consequent rules.

• use the exercise as though it were a database of assertions.

Example (the Macbeth example in figure 17.3)
Desired matching:

Macbeth	Greed	Role
Macbeth	The noble	Victim
Lady Macbeth	The Woman	Villain

Rules:
R1 If ?x1 is loyal to ?x2
?x3 murders ?x2
Then ?x1 kills ?x3

R2 If ?x3 is evil
?x3 wants to be king
?x2 is king
Then ?x3 murders ?x2

R3 If ?x4 persuades ?x3 to want to be king
Then ?x3 wants to be king

R4 If ?x4 is greedy
?x4 is able to influence ?x3
Then ?x4 persuades ?x3 to want to be king
R5 IF ?x3 is weak
?x3 is married to ?x4
Then ?x4 is able to influence ?x3

The process of converting individual names into match variables is called variablization.

Individual	Variables
Macduff	?x1
Duncan	?x2
Macbeth	?x3
Lady Macbeth	?x4

consider the use of Greed as a source of assertion.
The woman is greedy
The noble is weak
The noble is married to the woman

• with these 3 assertions, and the about 5 rules, ask whether Macbeth supports the assertion:
  the noble wants to be king.

• R3 creates the following binding:

Variable	Binding
?x3	the noble

• From R4:

Variable	Binding
?x4	the woman
?x3	the noble

• Finally,

Variable	Binding in Greed	Origin in Macbeth
?x3	the noble	Macbeth
?x4	the woman	Lady Macbeth