Page 215 - Deep Learning

198 Adaptation

hence might come into play in other contexts, such as programming. in rule
notation, if the learner has already acquired the two rules
RG S → A
:,
1
1
and
RG S → A,
:,
2
2
it might be useful to create the new rule

′ RG:, ( 1 S 2 )S ⊂ → A,
where (S ∪ S ) is a symbol for whatever S and S have in common. This mech-
1
2
2
1
anism produces a new rule that is more general than either R and R and
2
1
hence might apply in situations not covered by either of those two original
rules. in such situations, action A might turn out to be correct, because that
situation shares some features with S and S . Strengthening, bottom-up rule
1
2
creation and various rule generalization mechanisms have been incorporated
into several computer models of skill acquisition.
Feedback can be negative as well as positive. When a person is acting in
an unfamiliar or changing task environment, his repertoire of rules will fit that
environment to some extent, but not completely. Consequently, some of his
actions will have their intended outcomes, but others will either be impossible
or irrelevant or they will produce unexpected outcomes. in everyday parlance,
we call deviations between expected and observed outcomes “errors,” and we
are used to regarding errors as undesirable. But in the context of skill acquisi-
tion, errors provide information. it may seem obvious how to learn from this
type of information: do not repeat the action just taken. But the problem of
learning from error is more complex than it first seems. it is discussed in depth
in Chapter 7.

Stage 3: Optimization
The optimization stage begins when the target skill has been mastered, that is,
when it can be executed reliably, and lasts as long as the person keeps practicing.
improvements can continue for a very long time. What improvements are pos-
sible, once a learner knows how to perform the task correctly? What additional
types of information become available after mastery? As practice progresses, the
stock of memories of past actions, answers and solutions grows. Mechanisms
that can make use of this type of information include shortcut detection and
optimization on the basis of statistical regularities in the task environment.