Slashdot Mirror
Believable Stupidity In Game AI
Gamasutra is running a feature written by Mick West, co-founder of Neversoft, about creating game AI that is dumb enough to defeat, yet intelligent enough that its "mistakes" are similar to those a real player would make, thus preserving the illusion that the AI is not just throwing the game.
"The simplest way to introduce stupidity into AI is to reduce the amount of computation that it's allowed to perform. Chess AI generally performs billions of calculations when deciding what move to make. ... The problem with this approach is that it decreases the realism of the AI player. When you reduce the amount of computation, the AI will begin to make incredibly stupid mistakes — mistakes that are so stupid, no human would ever make them. The artificial nature of the game will then become apparent, which destroys the illusion of playing against a real opponent. ... By reducing the amount of computation, we create an AI opponent that is trying to win, but has been crippled in a way that leads to unrealistic gameplay."
Actually I don't have much issue with the quality of AI's used in games, but I don't like that so many of them cheat. A few games of gotten it right, but many AI's can track my movements through walls. In order to defeat the AI, I have to figure out how it works and so I'm constantly aware it has superhuman abilities and I find that very distracting from the realism.
Have a PRNG have the outcome of 313373 cause the bot to either:
- Fall on own grenade.
- Rocket-jump at 25 health.
- Hump the face of the nearest corpse.
Finally had enough. Come see us over at https://soylentnews.org/
So, basically, we have to determine how many "calculations per second" equivalent an average human can manage. Then we have to allow a range on either side of that since not everyone has the same capacity. Once we manage that, game AI would start being more realistic, huh?
Somehow I doubt it's that simplistic but still sort of interesting.
You know the thing about UDP jokes? I don't care if you get it or not.
This must be the problem facing the team creating Duke Nukem Forever.
They needed the AI to be dumb enough so you could hear the comments all the time during the game.
Oh, the humanity! Football games drive me nuts when the AI does stupid things no real person would ever do. Why the hell did my fullback just brush by the linebacker that's right in my RB's way?! Why can't I get my linebacker to stay in his lane on running plays?!
I'd love to see the difference in difficulty in Madden being the difference between playing a Jerry Glanville-coached team vs a Tom Landry-coached team. Instead, all increasing the difficulty does is make your opponents more talented but no smarter. Even on the highest difficulty, FBs don't understand their blocking assignments.
Hmmm. Possibly this is a sign that TFA was written by an AI agent. Asserting that there are mistakes that are too stupid for for any human to make is a mistake that is too stupid for any human to make.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
Year after year, I read these kind of articles that report how game AI increases in leaps and bounds...and I still don't see it. Bad guys in GTA still seem to rush towards grenades, Halo/Gears of War enemies are either completely impulsive or avoidant. I'm not knocking the programmers...I think game AI must be very difficult to achieve, and even harder to detect for the layman (such as myself).
Does anyone have an example of really good AI in action games (or any non-RPG, non-RTS games)?
(-1, Raw and Uncut is the only way to read)
Fritz has a sparring mode which does a little better than that. It will find a route that sets up a position where the player can force a win of a piece or a pawn. Essentially it sets up a tactical middle game puzzle live in the middle of a game.
The player doesn't know when it will do this or even if it will happen at all. But it is most likely to happen when the player puts the computer under pressure. This is great because it teaches the player to press the computer and coordinate pieces while also constantly keeping an eye out for the wins.
The whole chessmaster series features near-perfect play alternating with just flat out dropping pieces.
Even with Fritz though, "easy" mode is still well above beginner.
t
If it is so difficult to raise the computers to the level of human intelligence, it is probably impossible to reach the level of human stupidity.
Unfortunately, the AI generally only gets a tiny slice of the CPU time. If it were to use as much as the graphics, etc, then sure: the extra processing required to make the AI believably stupid would be easy. The innovations are there, it's just they are too slow or processor intensive to be implemented in most games.
The author has obviously never played chess with me.
My first rule of game AI is that the computer should have access to the same information and controls as a human player. I hate games where the computer knows about your units and buildings that it hasn't scouted.
The big advantage that computers have is that they can micromanage every unit with 100% efficiency. One way to reduce skill could be to limit the amount of attention the computer can spend, maybe in the form of "actions per minute". For a game like poker that could be a limit on how precisely the computer player calculates odds. A more experienced human player has a better feel for the game, so a more skillful computer player could dig deeper into the nooks and crannies of probability.
A way that computers often act too stupid is not accounting for how their interactions with one player will influence other players who aren't directly involved. For example, in a three-way game the computer player might throw everything against the strongest player, weakening them both and letting the third player win. Humans have millions of years of instincts for dealing with such situations. So the game AI might need to precompute some game theory and adapt to opponent reactions over a series of many games. Then it could be dumbed down by reducing its use of that experience and acting more like a newbie human player.
If you want an AI to make human like mistakes, you have to have at least a roughly human cognitive model. The simplest way to do this, it seems to me, is to give the AI competing goals. Rather than just have the AI "try to win", and then cripple its ability to do that effectively, you could give it multiple goals to strive toward, and then give it some degree of randomness in which goal it chooses to pursue. Victory vs. pain-avoidance, attack vs. finding time to recover, etc.
Asserting that there are mistakes that are too stupid for for any human to make is a mistake that is too stupid for any human to make.
HA! I see through your ruse!
Have you driven a fnord... lately?
You must wait a little bit before using this resource; please try again later.
And vice versa.
One thing that is stupid: trying the same things that failed in the past.
This is a component of smart behaviors as well. If you don't know what to do, try something and watch very carefully how it fails. If you are in a desperate situation and you are going to definitely fail through inaction, then try an action which failed in the past and hope you get lucky. Or vice versa, when the action that usually works looks like its going to fail, try doing nothing.
Animals freezing in the face of danger is an example of this. It's not much of a defense, but you might get lucky. Maybe the prey a couple animals down the line will get itchy and draw the predator's attention.
On the other hand, smart behavior can be a component of stupidity. If the red button gets you a treat 80% of the time and the green one gets you a treat 20% of the time, the intelligently stupid thing is to push the red button 100% of the time. The stupidly intelligent thing is to try to work out the pattern of red/green rewards.
If you want realistic model of stupidity, provide the NPCs with a range of decision making strategies, all of which work to some degree, but the better of which take more effort (computation). The NPC can choose between the strategies with a random function weighted towards the better strategies, but as time and "stress" come into play the function can shift towards the easy but less effective strategies.
That's a pretty good model of human performance "choking".
Evolution did not produce human stupidity because it was useless, after all.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
If this were a karate match, this would be the equivalent of the master having to fight someone he's never met before, but that person has studied the master's every move. Then, after the master wins the fight, he has to fight someone else who has studied his every move and acts differently. Not a fair fight.
F.E.A.R. is the only game to date that I've played and honestly felt like I was playing against human opponents. Honestly, they were as smart as or even smarter than playing against people online.
They did this through heavy use of waypoints and scripted events. If that's what you have to do to make it realistic, then by all means, do it.
Do we have programmers that are smart enough to program stupidity algorithms to be smart enough to be as stupid as humans?
A computer that plays on equal terms.
AIs often win in some games not by the virtue of being smarter, but by having an unfair advantage. Examples:
My suggestion: An AI should be coded as a bot, within the constraints given to the player. If the player can only see a part of the battlefield (like in Starcraft) then the AI should have the same limit and need adjust its own viewport to gain awareness of an area. It should also be limited by the fog of war, and lack the ability to see out of the back of its head. To put in another way, a fair Starcraft AI would be one implemented with a camera pointed at the screen, controlling only the keyboard and mouse inputs.
The idea is that I want to be beaten because the AI is indeed smarter, not because it's got a superior access to the battlefield I can never gain.
Chess probably comes closest to the sort of thing I want -- the AI and human are fighting on very equal terms. I don't see the calculating millions of positions a second as a problem, that's simply an implementation detail.
Well, I don't know Kasparov and I'm not a history expert, but I'm willing to bet that because he beat Deep Blue in 1996, he just figured he could do it again with no problem. He probably underestimated how much Deep Blue had improved, but that's just my speculation.
I don't know Kasparov's motivations for playing Deep Blue at all, honestly. Beating Deep Blue wouldn't have won him any fame ("Oh look, he beat a computer. Computers suck at chess anyway."), but losing would look bad for him. He had nothing to gain and plenty to lose. I know after he lost the match, he demanded a rematch, but IBM refused and put Deep Blue out of commission. IBM had nothing to gain by beating Kasparov again, and their stocks had already started going up when the news got out that they had beaten the grandmaster (to their great pleasure, history did indeed forget the shady details I mentioned before, just as they hoped).
At the time that Deep Blue was being used, there was more of a focus on brute force search than tricky AI game play. Deep blue searched an average of 130 million nodes a second using a iterative-deepening alpaha-beta search, sometimes able to look 40 moves ahead. IBM declined a rematch after that game, but thanks to improvements in AI, a standard desktop PC running improved search algorithms is now a suitable match for even a grandmaster.
Deep blue also contained an 'opening book' of 4000 positions and 700,000 grandmaster games indexed.
Source: Artificial Intelligence A Modern Approach: Russell, Norwig
i'm not involved into any of these activities, but i know Kasparov as a great mind and chess player because big blue won against him.
and thats probably what the history will remind us.
Eventually, a human would have been beaten anyway.
Kasparov had immortality to gain. In fact I'm fairly certain most people today remember him only as The First Person To Lose Against A Computer In Chess and have no idea who deep blue is.
What! I was losing against my Sinclair Spectrum +2 as long ago as 1987.
Why is this debasement? Playing against a machine even in unfair settings is still a challenge. The goal isn't to pick only fights you can win but to test yourself and do something new and experience something new.
As for it being "unfair" it wasn't touted as fair it was a computer v grandmaster experiment not a ranking test.
The only shady part in my opinion is the refusal for a rematch. If the man is willing to play on uneven ground and give his time (even if paid) to your stunt then you should have the courtesy to let him challenge/redeem himself.
I read about in college. Kasparov did ask for the transcripts. However the Big Blue team told him no. Basically saying it would give Kasparov an unfair advantage because he could study Big Blue's moves and deconstruct the algorithm it used.
Sorry, but as you say I must almost completely disagree with your sentiment on scripted events in games. Scripted events are not good example of AI because there is no decision making, which makes them predictable and exploitable. So the only way to scale difficulty in a scripted scenario is to just pump up the AI's accuracy, damage, and health, which are not very interesting. Scripted events work fine for fairly linear narratives and setting up big cinematic set-pieces, but they don't make for interesting gameplay (mind you I did not say they don't make for "fun" gameplay).
The events you describe in AvP are not necessarily examples of scripted scenarios, but sound more like a scripted set of responses to stimuli the player provides (would the AI react differently if you didn't rip off a soldier's head?). The trick for interesting yet realistic decision making is to allow the AI a range of responses that it can make, and then every time you present the same scenario to the computer you may get a different response. So when the AI is actually making decisions, rather than just following a pre-scripted path, this allows the player to make actual decisions in response, rather than just following their own memorized trial-and-error derived path to success.
The key to a good scaling AI doesn't have to be that tricky, it can just be a matter of what range of choices you allow any given AI to make and what sorts of "mistakes" you throw in that pool of choices.
When I face "easy" opponents in an FPS, I want them to use simple tactics (not just be unable to hit the broadside of a barn) like charging forwards blindly, or getting scared easily and retreating or even panicking, and being easily suppressed by heavy fire. When I face more advanced opponents, I want the range of their choices to move up the tactical scale to include flanking maneuvers, suppression fire, use of cover, and tactical retreats. A good mistake for an advanced AI would be to assume you're in the wrong position if you duck out of view and to attack that wrong position vigorously (as opposed to the omniscience a lot of AI's seem to have). They don't have to be any more accurate or need any more bullets to kill than an easy bot, but at least they could present more of a real challenge without artificially increasing their stats. Granted this is harder to do and would require actual programming rather than just increasing a few numbers, but that's the price for good AI in your game.