the challenge of generating programs that allow computers to play games.
Two general strategies (by no
means mutually exclusive) have been used.
One approach is to program the computer to carry out a series of logical
In chess or checkers, for example, a program can apply a
predetermined sequence of algorithms
the best response to a given board position.
There are, however, some daunting obstacles to this way of doing things
in any but the simplest situations.
Since there are about 10
possible board positions in chess, and since the so-
called maze or tree of the game--that is, the playing-out of all possible games--is estimated to be about 10
sequences, this complexity (and the limited power of computers) prohibits even the best programs from a very deep
analysis of the value of the possible moves (which is, of course, what makes games of this sort so interesting to
“These limitations to logical analysis as a means of solving complex problems led people to think of other ways in
which machines (or brains) might play such games (e.g., von Neumann & Morgenstern, 1947; Shannon, 1950; Turing,
1950; Newell et al., 1959; Samuel, 1959).
An alternative would be to have the computer operate
accumulating and updating experience with the relative success or failure of moves initially made at random.
continually adjusting and refining the responses to a given challenge on the basis of the success or failure of a growing
body of experience, successful moves in a complex situation such as a given board position could be made without
any preconceptions about the rules of the game or how they might be used to explore the situation in logical terms.
this way of proceeding, at least in its extreme form, the machine would make no analysis of possible outcomes and
indeed would not have been provided with any logical rules by which to decide on an appropriate move.
response to any situation would be determined simply by the relative success or failure of all previous moves in the
same, similar, or related situations.
Given some way of profiting from experience, performance would gradually
improve as feedback form particular responses progressively altered the organization of the program (and, effectively,
the connectivity of the computer).
“This wholly empirical strategy would result in chance-level success at first, since little or no experience about useful
responses would have been available to guide behavior.
Over time, however, performance would improve and, given
enough experience, would ultimately become very good.
Even in this more sophisticated state, however, an analysis
of the altered structure of the program, depending on the complexity of the challenge, might given little or no insight
into the problem it had mastered or the rules of the game, even though the behavior of the computer would give the
appearance of being strategically informed in these ways.” (italics added)
--Dale Purves and R. Beau Lotto (2003),
Why we see what we do:
An empirical theory of vision