Game Development Reference
In-Depth Information
account all of the possible inputs. We will explore the exact mechanics of what went
wrong in a moment. By not taking key factors into account, it allowed itself to be
led into alternating between decisions. This is a phenomenon known as
strobing
,
flip-flopping
, or even just
flipping
.
While strobing can manifest in many ways, the common thread is the effect
that strobing has—
nothing
. In mild cases, strobing AI can look inefficient or mud-
dled. At its worst, strobing leads to the behavior seen in our example above where
the fleet alternated between sailing toward one city and the other, never attacking
either one.
I
GNORING
F
UTILITY
While that example implied a nefarious player intentionally exploiting the decision
algorithm of the fleet, AI agents can strobe between behaviors based on less direct
and less obvious changes in their decision factors. In fact, a worse problem is when
our game model generates strobing behaviors entirely on its own.
Agents that depend on each other for decision criteria can enter into a joint
concert of strobing. To illustrate with a simple example (with yet another nod to
Brother Occam), imagine two agents facing off in a Rock-Paper-Scissors contest.
There is a twist to the game, however: each agent can see the other player's inten-
tion
before
they make their play and elect to change what they are going to do.
At first, each player selects randomly. Then, upon examination of the pending
match, they can elect either to proceed or to change their selection. Because this is
a zero-sum game, the only two outcomes are a win-loss combination or a draw.
Unless there is a draw,
one
player is going to be on the losing side of the match.
(Even a draw is less than preferable.) Naturally, one would expect the soon-to-lose
player to be dissatisfied with this outlook and decide to change his play. Additionally,
because he can see the other player's choice, he changes his selection such that it will
put him in the position to win.
This act, of course, causes his
opponent
to be in the potential losing role. The
logical reaction is for the second player to switch
his
play as well. You can see where
this is headed, I'm sure. None of them will ever be satisfied with the pending out-
come to actually
commit
to playing. Between the two of them, they will spend all of
their time switching “weapons� in the battle and never fire a shot.
While this example might seem absurd, consider a scenario such as the Dude
assault from Chapter 14. As part of the decision process, our agent took into account
both the weapon that he was using and the one the
other
player was using. If the sce-
nario was drawn up so that there were distinct strength vs. weakness correlations
(as there are in Rock-Paper-Scissors), two agents with the same decision algorithm
may engage in much the same dance.
