Page 203 - Deep Learning

186 Adaptation

out without further breakdown. if the relevant subgoal is thereby satisfied,
control reverts to the next goal upward, where the process repeats itself. in
the terminology of computer science, this is called top-down, left-to-right
execution. This pattern is familiar to writers who use outlines.
At each decision point, the situation at hand is evaluated by testing all
relevant rules – that is, all rules that share the currently active subgoal – for
the satisfaction of their conditions. if there are no relevant rules that match
the current situation, then the person’s practical knowledge is incomplete; he
has not yet mastered the task. if there is a single rule that matches, the action
of that rule is executed. if there are multiple rules that match, they are entered
into a temporary memory structure called the conflict set. A single rule is cho-
sen from that set by a process called conflict resolution, and the action of the
25
selected rule is performed. There are various conflict resolution schemes. A
specific scheme will be described in Chapter 7.
in short, a set of rules is executed by four main processes: First, a match-
ing process compares all relevant rules against the person’s perception of the
current situation to determine the available options. Second, a conflict resolu-
tion process selects a single rule from the set of matching rules. Third, the
action of the chosen rule is performed. Finally, a control process moves the
locus of control from node to node in the goal tree. The mental machinery
26
that performs these processes is called the cognitive architecture. The dis-
tinction between a task-independent architecture and a pool of task-specific
rules neatly slices cognition into its invariant and changeable aspects. The
former is possibly innate, the latter certainly acquired.

PROCESSES FOR SKILL ACQUISITION
if practical knowledge consists of a collection of goal-situation-action associa-
tions – production rules – then to master a novel task or to adapt to changing
task demands is to extend or revise such a collection. A process that takes one
or more rules as input and generates an improved rule as output is a learning
mechanism. Like other change mechanisms, a skill acquisition mechanism can
be partitioned into a triggering condition and the change process proper. A
situation in which the triggering condition for some learning mechanism is
satisfied constitutes a learning opportunity. The execution of the correspond-
ing change process is a learning event.
Learning events are brief compared to the time it takes to master a skill, so
the Raisins in the Dough principle applies (see Chapter 4). The majority of the
cognitive processes that occur during practice participate in the execution of