Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence

Figure 12.5
Pruning the tree.

---

used to challenge chess grandmasters are able to evaluate around two
hundred million states per second.
Let’s compare what we’ve learned about automated chess play to
what we intuitively know about human chess play.
For starters, humans – at least those who play very well – also mem-
orise large numbers of opening strategies and variations thereof. For
seconds, humans are also very good at determining which moves are
unlikely to be made by an opponent playing well. Those humans who
are very experienced at chess also have a very good sense of how good
a particular game state is for them.
It seems then that there is a good fit between what we know of
automated chess play and what we can intuit concerning the play of
accomplished humans. There are, however, important distinctions
between the two, the most obvious of these being the sheer quantity
of search and evaluation carried out by chess computers.
It is clear that chess grandmasters do not explicitly apply a heuris-
tic function to some two hundred million states per second in order
to evaluate the goodness of the current game state and determine a
strategy. We need to be very careful here, however, as we might be
tempted to countenance a very poor argument against computation-
alism in light of this.
The very poor argument runs as follows. Computers need to search
two hundred million states per second in order to match the best
human players. These human players, however, clearly do not. So
human players are not enacting formal systems in playing chess.
Therefore computationalism is false, since there is at least one cogni-
tive task whose function cannot be accounted for in terms of the
operations of formal systems.
There are a number of reasons why this argument is bad. The most
telling criticism of it is that there is a distinction between human per-
formance and our best attempts at recreating it.
It may well be the case – and will be if computationalism is true –
that human players do actually employ these methods. It is just that
humans have very good guiding principles that allow them to engage
in significantly less search to determine winning strategies. The fact
that these human players are unable to make explicit the principles
they are guided by does not speak against the possibility that their
cognitive functions are to be accounted for in terms of the opera-
tions of formal systems. After all, much of our mental life is opaque
to introspection. It merely speaks against the best results of our
empirical investigation in seeking to determine the operations of
these systems.

129

---

Another criticism of the argument against computationalism is
that it equivocates between the cognitive functions we explicitly and
consciously engage in and cognitive function per se. To say that a
grandmaster does not explicitly consider two hundred million states
per second is not equivalent to saying that their cognitive processes of
evaluation are not implicitly tantamount to consideration of such a
number of states. Relying on a claim of the latter when only the
former is demonstrable is an equivocation which begs the question
against the computationalist.
While we clearly need to be careful about the conclusions we draw
from perceived distinctions between computer chess play and human
chess play, there are yet further di
fferences that are worth mention-
ing.
For one thing, the minimax procedure assumes that one’s opponent
will always make the move which is optimal for them. Very often,
however, it will be the case that a human player will make a sub-optimal
move. It may be the case that they fail to recognise the optimal move,
or it may be the case that they deliberately play a sub-optimal move to
try to force their opponent into following a particularly strategy.
A computer may well, of course, also make a sub-optimal move,
simply through failure to correctly evaluate. Deliberately making a
sub-optimal move to achieve an e
ffect in one’s opponent’s strategy is,
however, another thing entirely.
In order to actively seek to influence the strategy of one’s opponent
by making an unusual move, one requires an understanding of how
one’s opponent thinks about the game. In other words, one requires
some appreciation of how one’s opponent is likely to move given
certain circumstances. The best way to discern this is to play as much
as possible against the opponent in question.
During computer versus human tournament play, it is generally
allowed for the programmers to modify the computer program
between games in order to counter strategies deployed by the human
player. It is important to note that where the computer requires
human intervention to be able to respond to the idiosyncrasies of a
given player’s strategy, rapid responsive adaptiveness to peculiarity
seems to be something that humans naturally excel at. This leads us
to the final point of distinction between human and computer game
play we’ll consider here.
Consider the fact that the principles which guide human evalua-
tion of chess states are su
fficiently superior to those employed by a
computer as to require significantly less explicit evaluation of descen-
dant states in order to play comparably well. Consider also the rapid
130
  

---

responsiveness to the idiosyncrasies of a given player’s strategy which
is evidenced by human players but not computers.
It appears that there is something subserving both these abilities
which humans are very good at but which hasn’t been mentioned at all
in this discussion of machine game play. Humans excel at extracting