Game Development Reference
The second part of our definition requires that the AI decisions be intelligent
ones. Making the AI decisions appear intelligent is a recurring challenge. The
hardest part is not always making the AI look smart; often, it is preventing the AI
from looking dumb. At an AI roundtable many years ago, I dubbed this
''avoiding artificial stupidity.'' Even purely random decision making may fit this
criterion. Pundits describe insanity as ''doing the exact same thing and expecting
different results.'' In this light, random decision making gives the appearance of
an AI that is learning from past mistakes or one that is changing how it plays to
keep the player from countering the AI's past successes. If players cannot detect
the purely random nature, they often interpret it as highly intelligent. The final
judge of ''intelligence'' is the player.
It may seem unfair that AI will be judged by how stupid it is in its low points instead of how smart
it is in its high points, but other aspects of games are judged in a similar way. For example, Id
Software spent well over a year perfecting the graphics engine for one of the very first full 3D
games, Quake. The hard part was not hitting 30 frames per second at the top end, but keeping the
frame rate above 10 frames per second at all times [Abrash96].
The third part of our definition requires the game AI to react to changing con-
ditions. It is generally accepted that a core part of what makes computer games fun
is a high degree of interactivity. Quality interactivity requires that the players'
decisions matter—that they change the state of the game. Players are free to make
whatever choices they find agreeable and to change the game state in such ways as
they can. This is the changing world in which game AI must exist. This changing
world places great pressure on the ability of the AI to appear intelligent. If the AI
cannot act differently in different situations, it will be only slightly more interesting
than watching a stick fall to the ground when you let go of it.
Those three parts give us a working view of AI, but this topic is about game AI. As
an entertainment product, everything must further the player's enjoyment. Our
AI must make the game more fun. A harder challenge appeals to only one of the
four basic ways people have fun [Lazzaro04]. Even worse, a harder challenge
lacks fixed definition—one player's ''harder'' is another player's ''frustrating.'' A
game AI programmer must never lose sight of fun as the primary goal. Making
the AI smarter and more sophisticated often makes the game more fun, but not
always. An online game that programmed the monster AI to target healers and
mages preferentially proved to be far more effective at defeating the players but
far less fun for the players.