Page 237 - Deep Learning

220 Adaptation

of attention, loss of information from working memory due to processing
overload or too much noise in the activations and strengths of knowledge
30
structures. We lack information about the prevalence of the different causes
of error in everyday life.

Constraint-Based Specialization

If the main cause of errors is that the initial rules for a task tend to be overly
general, then to adapt a strategy to a task is to gradually specialize those rules
by incorporating more and more information about the task environment
into them. As learning continues, the condition side of each rule becomes
more and more restricted and the rules will consequently become active in
fewer and fewer situations. As this process continues, a rule is eventually acti-
vated only in those situations in which its action is correct. According to this
Specialization Principle, the direction of change during practice is from general
but ineffective methods toward methods that are specific to a particular task.
To adapt is to incorporate information about the task environment into the
knowledge that controls action.
Given these concepts, the problem of learning from errors can be stated
with precision: If a rule R with goal G and situation S as its conditions and
action A as its right-hand side,

RG S → A,
:,
is activated in some situation S , and if the execution of A leads to a new
1
situation S that violates constraint C = <C , C >, then what is the appro-
s
r
2
priate specialization of R? That is, which conditions should be added to
the condition side of R so as to prevent the rule from violating C in future
situations?
For example, suppose that a novice car driver has discovered that driv-
ing behind a bus makes for slow progress on the highway and consequently
acquired the rule, If I am trying to get ahead, I am in the right-hand lane,
another vehicle x is also in the right-hand lane, x is ahead of me, and x is a bus,
then I switch to the left lane. Formally, this rule can be written as

R: Goal = (Make progress)
Situation = [(In me rightlane) & (In x rightlane) & (Ahead x me) &
(Isa bus x)]
⇒ SwitchLeft(me),