Page 239 - Deep Learning

222 Adaptation

The added condition elements are shown in bold face. The rule condition now
says that it should not be true that there is a vehicle behind me in the left lane.
More colloquially, the revised rule says, switch to the left lane only if there is
no car coming up from behind in that lane. This rule recommends switching
lanes only when the constraint is irrelevant. The procedure by which the new
restriction on the rule is identified is general and well-specified: Negate the
relevance condition of the violated constraint and add it to the situation part
of the rule.
The second response to the constraint violation avoids future violations
by creating a second version of the original rule that applies only when the
constraint is guaranteed to be satisfied. This is done by adding both the rel-
evance and satisfaction conditions (without negating them) to the rule. In the
driving example, the relevance condition is once again that there is a vehicle
coming up from behind in the left lane and the satisfaction condition is that
this vehicle is moving more slowly than the driver’s vehicle. If those conditions
are added, the second revised rule becomes
Rʹʹ: Goal = (Make progress)
Situation = {(In me rightlane) & (In x rightlane) & (Ahead x me) & (Isa bus
x) & (In x leftlane) & (Behind me x) & not-(Faster x me)}
⇒ SwitchLeft(me),
where the added conditions are once again shown in bold face. This second
new rule says, approximately, if there is a vehicle in the left lane, switch to that
lane only if that vehicle is moving more slowly than your own.
When these two revisions are completed, the initial disposition to switch
into the left lane as soon as there is a bus ahead has given rise to two special-
ized versions of itself. The first new rule says that one can switch into the left
lane if that lane is empty; the second says that one can switch into the left lane
if any vehicle in that lane is slower than oneself. Between them, these two new
rules exhaust the possibilities. The rule revisions identify the weakest condi-
tions under which the original rule will violate the constraint, and then con-
strain the rule so as not to be active under those conditions. The result of these
revisions is that the wisdom encoded in the learner’s declarative knowledge is
incorporated into his practical knowledge for how to drive. In A. Newell’s ter-
minology, structure has been moved from the evaluator (the constraint base)
to the generator (the rule set). 32
The learning algorithm works because constraint violations are informa-
tive. Mismatches or deviations between expectations and observations do not
merely signal that an error as been committed; they also provide information