L3-
Learning by Correcting Mistakes
isolate suspicious relations that may cause failures
explain why those facts cause problems
repairs knowledge
Isolate Suspicious Relations
use a "near-miss group" to identify the suspicious relation û relations in the true-success set and false success set.
a simple matter of using set operations on the relations that appears in the true successes and false success
To isolate suspicious relations using FIXIT,
To isolate the true-success suspicious relations,
intersect all true successes. Call the result ÃT.
Union all false successes. Call the result ÈF.
Remove
all assertions in the union from the intersection. These are the
true-success suspicious relations, written mathematically as
ÃT-ÈF.
To isolate the false-success suspicious relations,
intersect all false-successes. Call the result ÃF.
Union all true successes. Call the result ÈT
Remove all assertions in the union from the intersection. These are the false-success suspicious relations, written mathematically as ÃF-ÈT.
In general, there will be more than one suspicious relation.
the more true successes and false successes you have, the fewer suspicious relations there are likely to be.
For example, the fixed handle is a suspicious relation, because both +ve example have fixed handles while both near misses don't.
Intelligent Knowledge Repair
Example: Cup example used in the previous chapter.
Original Model:
The new model should be:
To make repair, do a breadth-first re-examination of all the relations in the model's And tree, looking for a relation with an explanation that needs to be replaced.
For each relation, look for precedents that tie the re-examined relation to at least one of the true-success suspicious relations. If found, replace the subtree beneath the re-examined relation using those precedents thus explaining the re-examined relation in a new way.
The new explanation should be as short as possible.
The longer the chain of precedent-supplied cause links, the less reliable the conclusion.
Initially, the re-examination effort is limited to:
Head Set ¥ the original precedents that are used to learn the existing model.
Tail Set ¥ those precedents in which one of the true success suspicious relations cause something.
When re-examine a relation:
look for a way to explain that relation using all but one of the precedents in the combined head/tail set.
The exception is the precedent that was used to explain the re-examined relation previously.
If the re-examined relation is not connected to the suspicious relation, then try to find more precedent.
Augment the head/tail set with a smaller precedent set (i.e. Tail set in the example) by looking for precedents that extend the cause link chains that lead through the existing head/tail set precedents.
The Repaired Cup-Identification Rule
IF The
object has a bottom
The bottom is flat
The object has a
concavity
The object is light-weighted
The object has a
handle
The handle is fixed
THEN The object is a cup.
To summarize, to deal with true-success suspicious relations,
Until the false successes are accounted for,
For each relation in the model tree, starting with the root relation and working down, breadth first,
Form a head set consisting of all precedents that contributed tot he malfunctioning model tree.
Form a tail set consisting of all precedents in which one or more true-success suspicious relations cause another relation.
Try to find a new explanation for a relation in the model tree using the head set, minus the precedent that explained the relation before, plus the tail set.
If an explanation is found, replace that portion of the model that lies beneath the newly re-explained relation
Augment the smaller of the head set and the tail set with other precedents that extend the casual chains found in the set's existing precedents.
Censor
Use censor to handle false-success suspicious relation.
e.g. Even one knows the pail contain "handle is hinged" relation, the algorithm may still try to establish "handle is fixed" relation.
Create Hinged-handle censor:
if handle is hinged then handle is not fixed.
To stop it from trying to explain a relation that can be shown to be false by a single cause link.
Create censor when program cannot account for the false successes using the true-success suspicious relation.
To create censor, perform a breadth-first re-examination of all the relations in the repaired model tree, looking for precedents that tie the negation of each re-examined relation to a false success relation.
Initially, the precedent set is limited to:
initial head set ¥ precedents in which the negation of the re-examined relation is caused by something.
Initial tail set ¥ precedents in which one of the false-success suspicious relations causes something.
To find an explanation for the negation of the re-examined relation that include at least one of the false-success suspicious relations. If one can be found, can create censor from that collection of precedents.
Eventually, the suitcase example will be need:
A
Suitcase
This is a
description of a suitcase. The suitcase is liftable because it has a
handle and because it is light. The handle is not fixed because it
is hinged. The suitcase is useful because it is a portable container
for clothes.
To summarize, to deal with false-success suspicious relations,
Until the false successes are accounted for,
For each relation in the model tree, starting with the root relation and working down, breadth first,
Form a head set consisting of all precedents in which the negation of the relation is caused by another relation.
Form a tail set consisting of all precedents in which one or more false-success suspicious relations cause another relation.
Try to find an explanation for the negation of the relation in the model tree using the head set plus the tail set.
If an explanation is found, create a new censor.
Augment the smaller of the head set and the tail set with other precedents that extend the causal chains found in the set's existing precedents.
If there are no suspicious relations in a near-miss group,
There may be just a few explanations, each of which is common to a subset of the true successes or the false successes. The problem is to partition situations into groups, inside each of which there is a consistent explanation for failure.
Try to use a relation that occurs in some, but not all, true successes as the suspicious relation. See whether that relation can be used to explain the failures. If it can be, ask whether that true-success relation was left out of the other true-success descriptions by oversight.
May indicate that the situations must be described at a finer grain, adding more detail, so that an explicit explanation emerges.