Slashdot Mirror
Adaptive AI in Games - Does it Really Work?
qasimodo asks: "I was recently reading a preview of Splinter Cell: Pandora Tomorrow, and then I came across this article at GameSpot saying Pandora Tomorrow will feature adaptive AI which 'will adjust itself to players' skill level'. I remember (and is also mentioned in the PT article) Max Payne also featured this, but I never noticed it. I guess that's the best way to know if it works, since it adapts to your gaming skills, but does it really work? Have you noticed it? Do you have proof of it?"
I remember reading an article about Gods by the Bitmap Brothers where they had helper triggers for people that had been stuck for a few minutes. They called it a shandyometer, after shandy drinkers obviously.
X-Com had a shandyometer, my old housemate used to send men who were very poor and irritating out, let them get slaughtered, then send in his main team and the game would have made it easier.
(For the non-Brits, Shandy's a mix of lager and lemonade (as in 7up/sprite), the old lore is that its drinkers are somehow unable to handle real beer)
Descent 3 always had adaptive AI. I remember when I first played it. I had played the prequels so I went into 3 with the same strategy I always had. It got me through quite a bit of the game, but it wasn't easy. Early in the game I was able to fly around really fast picking enemies off one at a time. As I played more they started to run away, regroup and attack in force. I countered that strategy by using bigger guns to destroy them. They then started to change formation to minimize the damage I could do to their whole group. While Descent 2 was the pinnacle of the series Descent 3 had revolutionary AI.
Also, I think that the sea battle AI in puzzle pirates could possibly be adaptive in some way. A couple updates ago they allowed brigands (computer controlled boats full of booty) to fire canonballs. Since then it has been widely agreed upon by players that they have increased in difficulty each and every day. There was an update last night, so we'll have to see what happened. I'm still a little unsure of this because if the AI was adaptive in some way, wouldn't they tell us?
The GeekNights podcast is going strong. Listen!
as much attention as it is. After all, I believe it was NBA Jam which introduced this concept, though they called it "CPU Assist". Essentially, a player who was losing would get more and more help from the computer as his deficit grew and grew, making his shots much more likely to go in and letting him knock opposing players over much more easily. On the flip side, a player with a big lead would find most of his shots hitting iron, and his players would lose the ball and get knocked over if opposing players even looked at them.
I also recall reading many years ago in an issue of Sega Visions (Sega's failed answer to Nintendo Power) that the Jurassic Park game for the Genesis would have "Dynamic Play Adjustment". The only example I can recall of this is that if the player was doing well, gaps to jump would get wider. I'm sure there were other examples, but that's the only one I remember.
So, in other words, this ain't new.
This is the NFL, which stands for "Not For Long" if you keep making those bulls*** calls.
Maybe its becoming more sophisticated in games like Pandora Tomorrow, but as a general game design concept, it's been around for a while.
The concept is mentioned in Rules of Play. Although I don't have it accessible right now, the example they used was in Wipeout XL (but could be any other Wipeout game for that matter). If your racer took a spill in the first lap and the AI of the computer racers didn't change, you'd have no chance at making it back up to the front of the race. However, because Wipeout XL used apadaptive AI, the conputer racers slowed down a little bit in this kind of situation making it possible, although more difficult, for the player to still win the race. That's not to suggest that this was the first instance of adaptive AI either. Conceptually, I'd wager that its been in games (probably in the sports genre) for quite some time.
The "adaptive AI" in Max Payne was simply a dynamic difficulty slider. I wouldn't be surprised if it was simply a matter of monitoring your kill rate, perhaps adjusted by your injury rate, and increasing enemy effectiveness (fire rate, accuracy, maybe damage multiplier if they really wanted to be devious) accordingly. I imagine the AI topped out fairly low, though, so the overall effect was negligible. I don't believe it was adaptive in the sense that it dynamically adjusted its battle tactics according to your battle tactics, as opposed to simple numerical effectiveness.
Say I tend to shootjump to the right when I head into battle. The AI couldn't care less. Now if it _did_ notice that enemies tended to die more often when I did so, and cause them to proactively fire where I would, statistically speaking, very likely end up, that'd be an AI to write home about.
The most remarkable AI in modern gaming that I've encountered of late is that of Virtua Fighter 4: Evolution (a bargain at $20 new, btw), in which battle profiles of players in the Japanese arcade circuits was distilled into what prove to be fairly different AIs. Dynamically speaking, the game tracks the areas that you tend to attack successfully (high, middle, or low), whether those attacks are strikes or throws, and whether you won or lost with those percentages...and, judging from its effectiveness at smushing me in the long run, adjusts its behavior accordingly. So while a given AI profile might tend to, say, try to counter middle throws often, that tendency might be further exaggerated as the bulk of my throws tend to come from that area.
Quite impressive not only for its dynamism but also for the wide and finely graded range of difficulty among the AIs. As you gain ranks in the Quest mode (from 1st kyu to 10th, 10th dan to 1st, and beyond), your opponents very slowly become more difficult such that you can actually observe effective tactics emerging and adjust your _own_ behavior accordingly. Quite a far cry from Street Fighter II, mm?
Considering the advances made in computing, I'm surprised that current gaming AI is still so sucky. It seems that our games have advanced only graphically and in size (I'm thinking world size -- but physical size has also grown) -- largely due to advances in memory and storage media.
As far as I can tell, AI has not advanced with current technology. I'm reminded of this while playing modern RTS games, where it seems that all computer opponents have similar stratagies, but never seem to ever actually "learn" anything (or even show a hint of adaptability). Of course, this is all purely antecdotal, but not without merit.
Any game developeres care to back me up? Or am I full of shit on this?
-Turkey
I can vouch for Max Payne's variable AI working. I played the game through once and thought "This is fantastic." So I decided to play the game through again. The first level was a breeze. I single shot to the headed every single enemy. Then, next level, all hell broke loose. The enemies were rolling, took at least four shots each, ducked bullets, hid behind things, shot back very accurately.. they couldn't kill me but I ran out of ammo quick. I never made it past the second level because the AI had jacked up to the highest level. If I recall correctly, when you die on a level it lowers the AI's abilities just in case it had them too high.
Aren't you guys messing the adaptive difficulty level with adaptive AI. Adaptive AI should learn, not simply adapting number or speed of enemies to the lameness of player.
What they need to do is adjust to the players styles. I like to camp and snipe in just about every game. I wish it would then cause the AI to start moving more rapidly, and check sniper spots more often. Do you always do something when you enter a room? Then the AI should brilliantly counter it, so I have to get a new strategy. Do I always camp in the same place? Then nade me.
Second of all I dont want it to turn pathetically easy, even though every game should have a difficulty setting called baby or wuss. (especially racing games like Gran Turismo) that would let you win. I labored so many damn hours perfecting my skills to no reward in Gran Turismo Three and I want my Formula One cars NOW!!! But what is the fun if I never die and never get hurt and don't get that rush when you complete a challenge. Because when I beat GT3 I will be so happy and thrilled and I will feel my $50 and racing wheel paid off. I will buy GT4, hence a happy customer and money for the corporations giving a financial incentive to the suits.
What it should do is offer a hint, to really bad players(flash bang a room with possible enemies and friendlies!!). Adapt to my style. If I want to snipe, then those guards better give up the grenades and take up some binoculars and a rifle. If a player is good at one method, make the objectives possible using that strategy, but encourage all the others too! Adaptive AI should enhance game play and make the game last. If I completed the it the first time sniping, and I want to be forced to try close quarters combat next time, without setting something I have not done any harder and make me have to vary my strategies to complete it. If Splinter Cell pulls off what I want, then UbiSoft will be very rich and I will be one very happy gamer.
Mod Wisely.
When playing Max Pain I found out about the adaptive AI the hard way.
I naturally adapt a one shot kill strategy in most games to be more efficient (less ammo wastage/less health wastage), unfortinatly in max pain this ended up with the first few level being very easy and then suddenly becoming impossible to complete.
Never the less max pane was religated to the dust bin as I was so fustrated and anoyed by this that I hardly felt like replaying half the game.
Just goes to show that Q&A testng is vital when implimenting a new tech as this should have most definilty been picked up.
GPLv2: I want my rights, I want my phone call! DRM: What use is a phone call, if you are unable to speak?
Not entirely true. Racing games don't HAVE to have adapative ai. The Grand Prix series never has. Grand Prix Legends, from what I gather, is kind of adaptive. It will set the ai times by your best lap to a degree.
NASCAR Racing never used to have it until this last iteration. As I said, it works pretty well. Starting a race from the back at Daytona, I managed to fight up to 17th place by the end (10% race distance) which is pretty good, and it was a huge fight all the way.
One thing a racing games DOES need it for is consistency. Many people find themselves great at one track, and not so hot at another. Without adaptive ai, you can crush the AI at one track, then lose badly at the next. It's more a question of realism than anything else. It's rare these days to see a driver lead a race at one track, then get lapped by everyone at the next.
See my comment here for an old, old case that I do consider learning.
I've had this sig for three days.