Page 120 - Deep Learning

Creative Insight: The Redistribution Theory 103

posing subgoals are usually deliberate and capacity-demanding processes, and
they tend to occur sequentially rather than in parallel. The set of options con-
sidered at each step along the way is called the conflict set and the process of
choosing an option is called conflict resolution. If the chosen step turns out to
have a negative outcome, the best response might be to back up to an earlier
state of affairs and do something different from what was done before. This is
often easier to do in the mind’s eye than in the material world.
Knowledge elements that decide choices among rival actions and operations
36
are called heuristics. Like inhibitory links in perception and memory, heuris-
tics constrain the set of options. There are two types of heuristics. Strategies
focus attention on the most promising actions available in a particular situation
or problem state. Some strategic heuristics apply generally because they abstract
from the task content. For example, if one action appears to be more promis-
ing than its rivals, it makes sense to explore it first. As the example illustrates,
general methods are not very helpful. Most effective heuristics are specific to a
task or a type of situation. For example, a locked door might trigger a search
for a key, while a swollen mountain creek might trigger a search for a bridge.
Evaluation functions, a second type of heuristic, support comparisons between
outcomes. For example, a chess player can calculate the relative strengths of his
own and his opponent’s positions after a move, and thereby evaluate the use-
fulness of the move. The application of evaluation functions presupposes the
application of strategic heuristics, because some cognitive operation or action
has to be executed before there is some outcome to evaluate. Strategies and
evaluation functions are retrieved from memory through the same process of
spreading activation as other types of knowledge elements.

The Power and the Limit of Experience

Analytical thinking proceeds by projecting past experience onto the situation
at hand. The initial interpretation of a problem situation is determined by the
biases laid down in the visual and verbal systems by prior experience. The
retrieval of affordances, inference rules and potential subgoals is determined by
the relative strengths of memory links, another repository of past experience.
The choices of which steps to take, which paths through the solution space
to explore is determined by heuristics – strategies and evaluation functions –
acquired in prior experience. Each of the main processes – perception, retrieval
and search – helps contain the combinatorial explosion inherent in solution
spaces by constraining the set of options to the most promising ones. In tight
contexts, this is effective. The more knowledge that can be brought to bear, the