Defining Decision Theory - Behavioral Mathematics for Game AI - page 47

The effectively random nature of dealing cards makes it possible that we will lose
the hand despite making the suggested play. However, if one were to do the math,
the result still shows us that our best chances lie in a certain direction. Therefore, we
should follow it. Any kind dealer (and even the often overly helpful people at the
table) will tell you what your play should be—making it fairly apparent that the
proper Blackjack play is predictable.
Therein lies the problem. Using normative decision theory exclusively as a
source of game AI makes those decisions and actions very predictable. If, as a
player, you can track the same criteria the agent is using to make those decisions,
you can then match them with the decision the agent makes when those criteria are
in place. As these patterns repeat, you will be able to determine what those
“shoulds� are. Thus, you can predict with complete accuracy what the AI agent is
going to do at any given time. All you are doing is connecting the cause and effect
model that the agent is depending on completely for its thought process. You have
exposed what seems to be a complicated process for the Rube Goldberg machine
that it is: Everything happens for a reason—everything happens in order… because
it should .
The other problem with using normative decision theory is that it is entirely
mechanical. I don't necessarily mean mechanical in the sense of completely deter-
ministic as explained above. I mean it in the perceived sense that it doesn't seem
human. Certainly, much of that effect is due to the rigid cause and effect chain.
However, there is another factor in play: Not all people act the same way all the
time . If you were to face 10 enemies that are using the same model, and they all act
the same way every time they are faced with a constellation of criteria, the notion of
humanity (or any other - ity you want to apply here) is lost. We are exposing our
agents for the completely mathematical and logical bots that they are. Not only does
that not look “real,� but how long do you think it generally takes for the cause and
effect rules to make it onto the Web or into strategy guides? “If you want to beat
[this game], do [this]—the AI will do [that]… every single time.�
Using normative decision theory for game agents causes people playing against
them to subscribe to using it as well. We have reduced the exchange to the deter-
ministic equivalent of Tic-Tac-Toe. There is an answer for everything on both sides.
If you want to win, you must follow it. To harken back to Chapter 1 and Mr. Sid's
idea of interesting choices, it would seem that we have unwittingly removed both
the “interesting� and the “choice� from the experience. There must be a better way.