|
[Neversoft co-founder West presents a thought-provoking look at improving the believability of AI opponents in games by upping their use of "intelligent mistakes", in a piece originally written for Game Developer magazine.]
Twenty years ago, I was working on my first commercial game: Steve
Davis World Snooker, one of the first snooker/pool games to have an AI
opponent. The AI I created was very simple. The computer just picked
the highest value ball that could be potted, and then potted it.
Since
it knew the precise positions of all the balls, it was very easy for it
to pot the ball every time. This was fine for the highest level of
difficulty, but for easy mode I simply gave the AI a random angular
deviation to the shot.
Toward the end of the project, we got some feedback from the client
that the AI was "too good." I was puzzled by this and assumed the
person wanted the expert mode to be slightly less accurate. So I
changed that. But then I heard complaints about the decreased accuracy,
and again that the AI was still too good.
Eventually the clients paid a visit to our offices and tried to
demonstrate in person what they meant. It gradually came out that they
thought the problem was actually with the "easy" mode.
They liked that
the computer missed a lot of shots, but they thought that the
positional play was too good. The computer always seemed to be leaving
the white ball in a convenient position after its shot, either playing
for safety or lining up another ball. They wanted that changed.
The problem was, there was no positional play! The eventual position
of the white ball was actually completely random. The AI only
calculated where the cue ball should hit the object ball in order to
make that object ball go into a pocket.
It then blindly shot the cue
ball toward that point with a speed proportional to the distance needed
to travel, scaled by the angle, plus some fudge factor. Where the white
ball went afterward was never calculated, and it quite often ended up
in a pocket.
So why was it a problem? Why did they think the AI was "too good" when it was actually random?
 Humans have a tendency to anthropomorphize AI opponents. We think
the computer is going through a thought process just like a human would
do in a similar situation.
When we see the ball end up in an
advantageous position, we think the computer must have intended that to
happen.
The effect is magnified here by the computer's ability to pot a ball
from any position, so for the computer, all positions are equally
advantageous.
Hence, it can pot ball after ball, without having to
worry about positional play. Because sinking a ball on every single
shot would be impossible for a human, the player assumes that the
computer is using positional play.
Design or Code?
Is this a design problem or a code problem? To a certain extent it
depends on the type of game, and to what extent the AI-controlled
opponents are intended to directly represent a human in the same
situation as the player.
In a head-to-head game such as pool, chess, or poker, the AI
decisions are very much determined at a pure code level. In a
one-versus-many game, such as an FPS, there is some expectation that
your opponents are generally weaker than you are.
After all, you are
generally placed in a situation of being one person against countless
hordes of bad guys. Other game genres, particularly racing games, pit
you against a field of equal opponents. Here the expectation of
realistic AI is somewhere between that of chess and the FPS examples.
The more the computer AI has to mimic the idiosyncrasies of a human
player, the more the task falls to the programmer. The vast majority of
the AI work in a chess game is handled by programmers. Game designers
would focus more on the presentation.
In an FPS, the underlying code is
generally vastly simpler than chess AI. There is path finding, some
state transitions, some goals, and some basic behaviors.
The majority
of the behavioral content is supplied via the game designers, generally
with some form of scripting. The designers will also be responsible for
coding in actions, goals, and responses that emulate the idiosyncrasies
of human behavior.
|
Printer-Friendly Version
When we were working on Rumble Racing, the goal was to make the AI opponents feel like people. We had a matrix of possibility for introduced errors, where at each pathing junction the AI had a choice to take a fast line or slow line. This matched up how people drove the tracks, where small errors in braking and taking the corners made the cars seem to go slower because of poor driving, not simply going slower.
Hey Chris - Rumble Racing was (and is, as I still play it today) one of those great games where one really got to see the AI do some things you'd expect a live opponent to pull off, so it was always a blast to race the same tracks multiple times and almost never see the same result.
g.
This gives moments of hesitation and panic which look extremely realistic (at least in something as simple as driving or fleeing). I've tried to think of a good way to cause the AI to flipflop between decisions only when they are in a panic, however that seems to require breaking good decision making skills (requiring a set amount of time before they can change action).
It's nice to see an article on this subject. A good deal of thought has to go into creating lifelike artificial behavior, and I'm glad to see people in the industry upping their game.
eg: I try to rush into the range and view of my opponents from whence they previously had no knowledge of my presence. Current A.I. would allow all notified opponents to 'snap to' my character and begin firing/ducking for cover with their speed and accuracy determined by difficulty. What I want to happen is for the opponents "psychology" to determine their actions, even though they may be 'trained soldiers.' I just initiated a "surprise attack" and expect them to act surprised depending on a combination of how alert the opponent may be and how efficiently I got within range.
It may sound like what I'm asking for is scripted A.I. set-pieces. In one way yes, but this has much larger scope. To me, realistic 'mistake making' A.I. mean an opponent that "feels." Will retreat because he fears, will fight because he's confident. I'm an artist so I don't know how to implement it, but it would be good to see the A.I. literally go "whoops"
Personally, I like gmaes that use the old D&D system of % to hit; etc; as I always feel like it's an even playing field between the player and the computer. Many games do a great job with AI (like Fallout 3 most recently), but I think some FPS and action games where the AI seems to be on a straight path to the protagonist can benefit from some rpg-style stats under the hood.
John Smith: Try explaining to your producer that you need to develop two different physics systems - one for the game to use, and one for the A.I. ;) Seriously though, your example assumes that with the right interfaces, an A.I. "brain" would act similarly to a human brain, which simply isn't true - even if your A.I. was unable to be accurate, it doesn't mean that it would feel like you're playing against anything analogous to a human opponent.
Games are experiences in which you need to create a suspension of belief, and the A.I. is a tool in that - there will never be a fair fight between an A.I. and a human, but with your A.I. making the right mistakes, you can trick your player into thinking that there is. And at the end of the day, that's all that matters! ;)
C&C 3 is case in point for this.
You mention the prospect of setting up a situation that the player can exploit. That's definitely a fine solution to the handicapping problem! My interest is in the preconditions for setting up such a situation.
As you show with the poker examples, there's usually a reason that a player makes an error, or a type of error which is more typical. (Folding to a big raise, against the odds. That one is a very good and clear example! Folding when "unclear" about the odds... I think this means that you can teach the machine which hands are complicated for a human to calculate the odds for.)
Let me give another example to sharpen the angle I'm getting at.
Let's say we were writing a chess program aimed at beginners (defining 'beginner' as not absolute novice, but rated under 1400). We'd want the machine to make beginner-like mistakes. The first step in "dumbing-down" the AI would be to figure out what kind of mistakes beginners make, and under what conditions. We'd be doing pretty well if we programmed the machine to threaten checkmate when it can be beaten by mate in three. We'd be doing even better if we figured out which kinds of mistakes are more obvious for players (and avoid those). So, the machine wouldn't fall for a sufficiently "obvious" mate in three.
In other words, we'd want a function which evaluates the "cleverness" of a move; the machine, imitating a beginner, is less likely to notice clever moves. Then, when we're setting up a scenario for the player to exploit, we have a believable reason for making the "error."
The AI is not playing the game to win.
The AI is a granny playing with a spoilt brat who throws a tantrum when it loses.
Personally I prefer AI with exploitable weak points, but most gamers are of the spoilt brat variety.
UT used to give the bots fewer shards in the Flak Cannon at lower AI levels, and make the rockets slower... pretty crazy.
We talked about this on the games-and-industry podcast Visiting The Village this week (www.visitingthevillage.com).
For example, take the game Monopoly. there would be specific strategies or "Tendencies" that were predetermined by the programmer to have top priority over other decisions regardless of how useful it was.
For instance, Maud will always buy level 4 properties (The green and blue properties) if able, and place houses on them as soon as possible. This would include a bias of being willing to offer VERY large sums of money in order to acquire there properties or giving up any lesser (according to the bias) property. while at the same time its shunning Level one properties entirely; this would include certain reckless actions, such as trading properties that were not level four even if it owned two out of three of a lesser property of the same color.
This sort of AI gave the illusion that this AI personality had a strategy. you would notice how "badly" he wanted boardwalk and you would hold out for a better offer. You got the feeling you were outsmarting him. and if you picked up on the tendencies, you felt much smarter, and rewarded for victory over them.
I think it would be a nice little tweak to shooter games if personalities were used. Players would know that there were very tactful enemies, as well as overyly aggressive (Reckless) ones. In this instance players can have waves of stupid enemies, mixed with a few, "Hard to kill" Bruce willis types. I think a dynamic AI such as this will enrich most shooters.
Schaeffer recommends that you tune your player so that it is risk-seeking. This doesn't necessarily make it inherently weaker, but if it gives its opponent an interesting game they wil be happier even if they lose.