Defining Decision Theory - Behavioral Mathematics for Game AI - page 51

In a similar vein, the practical joke of stopping in a crowded place to stare and
point up at something that isn't there is a great example of how you can influence
the public in this way. The more people start mimicking you, the easier it is to get
other people to look as well. Their thought process is just the same as the polling
data. They don't have the facts that would be required for a prescriptive motive
telling them they should look up. Instead, they have placed all their trust in the
descriptive sense of “Everyone else is doing it… I'm sure they wouldn't be doing it
without reason.�
T HE B EST OF B OTH W ORLDS
Certainly, there are merits in both of these areas, but while normative decision the-
ory (the bot) has obvious examples of how it should be used to make decisions in
games, how does descriptive decision theory come in? After all, we don't tap into a
database of survey information à la Family Feud to help our bots make decisions.
About the closest thing we have to that is capturing player data ahead of time and
attempting to make our decision models from that. If game AI were that simple,
these sorts of topics wouldn't be necessary. However, there is something to be
gleaned from the use of descriptive decision theory.
In game AI, we are tasked with something of a hybrid. On the one hand, we are
trying to determine, on any given game frame, what our agents should do. We are try-
ing to create a decision for the moment—the arena of normative decision theory.
That may or may not be a simple undertaking, depending on the decision or situation.
On the other hand, whether or not we can come to a decision about what
should be done, we are trying to emulate and/or re-create behaviors that look like
what people (or animals, or space aliens…) tend to do. Of course, these decisions
may or may not be what they should do . Hopefully, those tendencies that we analyze
with descriptive game theory (and attempt to re-create) will also be near enough to
the “most logical choice� as per normative game theory so that they don't look
ridiculous—slightly misguided or a bit erroneous perhaps, but not outright silly.
The result of all of this is that, to construct decisions that are meaningful and
realistic, we can't tie ourselves to the omniscient and purely rational tenets of nor-
mative decision theory. Looking again at the list of requirements for that to be in place:
Has all of the relevant information available
Is able to perceive the information with the accuracy needed
Is able to perfectly perform all the calculations necessary to apply those facts
Is perfectly rational