Page 94 - Deep Learning

The Production of Novelty 77

reduce their number to a smaller set of promising options. Sources of selectiv-
ity – which Newell and Simon called heuristics – can be derived from the goal
(how does the current situation deviate from the desired one?), task instructions,
background knowledge and other sources. in general, a heuristic is a piece of
knowledge that can be used to select among the options in a problem state. a
collection of heuristics forms a strategy for how to navigate the problem space.
Problem-solving steps are not only tentative but also anticipatory. Problem
solving alternates between the problem space in the head and the physical task
environment. to decide what to do next, the problem solver tries this or that
action tentatively in the mind’s eye (what would happen if I did this?). By antici-
pating outcomes, the person can think through and evaluate a course of action
before executing it. Such look-ahead has multiple advantages. Searching in the
head is faster and cheaper (and sometimes less dangerous) than searching in
the flesh.
The evaluation of an action outcome – real or anticipated – can be thought
of as mapping an anticipated problem state onto a value on some quantita-
tive dimension, hence the term evaluation function. For example, chess play-
ers evaluate chess positions in terms of what they call “material.” each chess
piece is ranked with respect to usefulness for winning the game, and a board
position can be evaluated with respect to which player has the stronger set of
pieces left. if one move leads to a stronger position than its alternatives, per-
haps it should be preferred. as the example illustrates, evaluation functions
tend to be task specific. They can be derived from goals, task instructions and
background knowledge. Like strategies, evaluation functions provide selectiv-
ity and hence help tame the combinatorial explosion; indeed, in Newell and
Simon’s works, the term “heuristic” refers both to tactics for choosing among
options and to evaluation functions.
The selectivity provided by heuristics – in both senses – is a matter of
degree. if a problem is truly unfamiliar, the problem solver will, by definition,
not possess any applicable heuristics, so the search will be random and unse-
lective, or nearly so (e.g., find the jar of tea bags in someone else’s kitchen). if
the problem is well known, the person already possesses an applicable strategy
and the search process collapses into a confident walk down the correct path,
as if produced by an infallible generator (e.g., multiply two 4-digit numbers).
For unfamiliar but solvable problems the process will fall somewhere between
these two extremes.
The heuristic search theory is a powerful theory of analytical thinking.
it explains the nature of the mental effort that goes into solving problems. it
explains why problem solving is possible even though success is not guaranteed.