PowerPoint Presentation
Learning by Correcting Mistakes
Metacognition
Learning by Correcting Mistakes
Meta-Reasoning
Ethics in Artificial Intelligence
Lesson Preview
Explanation-based learning revisited
Isolating mistakes
Explaining mistakes
Correcting mistakes
A Cup
A cup is an object that is stable and enables drinking.
An Object
This object is light and made of porcelain. It has a decoration, a concavity, and a handle. The bottom is flat.
Can we prove this object is a cup?
Which of these objects are cups? Check the box in the top left for each cup.
Which of these objects meet our definition of a cup? Check the box in the top right for each object that meets our definition.
ο ο
ο ο
ο ο
ο ο
ο ο
ο ο
A Cup
A cup is an object that is stable and enables drinking.
An Object
This object is light and made of porcelain. It has a decoration, a concavity, and a handle. The bottom is flat.
Can we prove this object is a cup?
Questions for Learning from Mistakes
1. How can the agent isolate the error in its former model?
2. How can the agent explain the problem that led to the error?
3. How can the agent repair the model to prevent the error from recurring?
handle-
fixed
handle-
moves
Algorithm for Isolating Mistakes
To find suspicious true-success relations:
Intersect all true successes (∩T)
Union all false successes (∪F)
Remove assertions in union from intersection (∩T – ∪F)
To find suspicious false-success relations:
Intersect all false successes (∩F)
Union all true successes (∪T)
Remove all assertions in union from intersection (∩F – ∪T)
Old Rule
If:
Object has bottom
Bottom is flat
Object has concavity
Object is lightweight
Object has a handle
Then:
Object is a cup
New Rule
If:
Object has bottom
Bottom is flat
Object has concavity
Object is lightweight
Object has a handle
Handle is fixed
Then:
Object is a cup
Brick
Brick
Brick
left-of
Current Concept
Brick
Brick
Brick
left-of
Not an Arch
touches
touches
supports
supports
Brick
Brick
Brick
left-of
New Concept
¬touches
¬touches
must-
support
must-
support
must-
support
must-
support
Questions for Learning from Mistakes
1. How can the agent isolate the error in its former model?
2. How can the agent explain the problem that led to the error?
3. How can the agent repair the model to prevent the error from recurring?
Object
Cup
is
Object
Bottom
has
Bottom
Flat
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Porcelain
made-of
Object
Decoration
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Object
Stable
is
Object
Cup
is
Object
Bottom
has
Bottom
Flat
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Porcelain
made-of
Object
Decoration
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Object
Stable
is
Object
Cup
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Handle
Fixed
is
Object
Cup
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Handle
Fixed
is
Is this a good way to fix this error?
ο Yes, because it shows only fixed-handle cups enable drinking.
ο No, because it will exclude some actual cups.
ο No, because some non-cups will still be included.
ο No, because it will cause incorrect decisions about other objects.
Object
Cup
is
Object
Concavity
has
Object
Liquids
carries
Object
Drinking
enables
Handle
Fixed
is
Object
is
Object
Handle
has
Object
Liftable
is
Light
Object
Cup
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Handle
Fixed
is
Object
Manipulable
is
Object
Cup
is
Object
Concavity
has
Object
Light
is
Object
Handle
has
Object
Liquids
carries
Object
Liftable
is
Object
Drinking
enables
Handle
Fixed
is
Object
Manipulable
is
Object
Orientable
is
Brick
Brick
Brick
left-of
Current Concept
Brick
Brick
Brick
left-of
Not an Arch
touches
touches
supports
supports
Brick
Brick
Brick
left-of
New Concept
¬touches
¬touches
must-
support
must-
support
must-
support
must-
support
Assignment
How would you use diagnosis to design an agent that could answer Raven’s progressive matrices?
To recap…
Explanation-based learning and incremental concept learning revisited
Isolating mistakes
Explaining mistakes
Correcting mistakes
/docProps/thumbnail.jpeg