Page 238 - Deep Learning

Error Correction: The Specialization Theory 221

where the SwitchLeft action has the effect of changing (In me rightlane) to
(In me leftlane), while leaving everything else unchanged. This driving heuris-
tic is overly general, because reckless moves between lanes are dangerous.
This state of affairs violates the constraint that one should not switch into
the left lane when there is a faster car coming from behind. This constraint –
call it C – can be formalized as

C = C ={ [(In me leftlane ) & (In y leftlane) & (Behind me y ]
)
r
* ** C = not-(Faster y me)},
S
in which the double asterisk once again symbolizes the connective “ought to.”
C says that if there is a car coming up behind me, it had better not be faster
than me (or else I am in trouble). Activating rule R when a faster car is coming
up from behind in the left lane creates a constraint violation, which becomes
obvious to the novice when the other driver honks his annoyance at the for-
mer’s amateurish driving. When the novice switches lanes, the relevance
criterion is suddenly satisfied by the situation at hand while the satisfaction
criterion is not. What might the novice learn from this mistaken action? How
should the novice driver revise his rule for driving?
When we encounter an unexpected and undesired outcome, we react in
two ways. We infer that we should do A only when the constraint is irrelevant
(I should turn left only when the left lane is empty) or when it is satisfied (I can
turn left when any vehicle in the left lane is slower than I am). The first response
constrains the rule to situations in which C is irrelevant. The second response
constrains the rule to execute only in situations in which C is guaranteed to
be satisfied.
Consider the first of these two responses. If there is no vehicle in the left
lane, the constraint is irrelevant and hence cannot be violated. In the driving
example, the criterion for relevance is that there is a vehicle coming up in the
left lane. The first specialization of the rule is achieved by negating that rele-
vance condition and adding it to the situation part of the rule. (The precise
algorithm for computing the conditions to be added is available in technical
publications.) The revised rule is
31
R′: Goal = (Make progress)
Situation = {(In me rightlane) & (In x rightlane) & (Ahead x me) &
(Isa bus x) & not-[(In y leftlane) & (Behind y me)]}
⇒ SwitchLeft(me).