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?"

Comments

If you are really bad, the enemies kill themselves

[Anonymous Coward]

You are too good. The way it works is that the worse your aim is the more likely enemies will randomly jump infront of your line of fire. IF you actually have aim though, then the enemies actually attempt to take cover.

Or...

[Mike Hawk]

Maybe it just determined you suck.

Deep breath, its a joke.

Shandyometer!

[Flat Feet Pete]

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)

Re:Shandyometer!

[Bustbang]

Into the wonderful.

Re:Shandyometer!

[Anonymous Coward]

"Gods..."

I remember that game. I recall that I stopped playing because I literally could NOT find my way through level 2 or 3. I don't remember what happened, I was quite young at the time, but depending on what kind of assisting they implemented for dumb players like myself, I think it bugged because I stopped playing after a while.

Is there somewhere that has more info on this game? Somewhere where I can see what they did in it from a technical standpoint?

Re:Shandyometer!

[Flat Feet Pete]

looks like it may be coming out for GBA!!

http://www.bitmap-brothers.co.uk/our-games/futur e/ gods-gba/

I remember reading about the technique in a magazine, over 10 years ago.

Descent 3

[Apreche]

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?

Re:Descent 3

[jonadab]

> While Descent 2 was the pinnacle of the series Descent 3 had revolutionary AI.

Actually, the AI in Descent 2 was fairly impressive in certain ways, though you
wouldn't usually notice it most of the time. The thing that caused me to first
notice it was in a level that I was creating. I'd positioned three Diamond
Claws (the nastiest/scarriest of the melee bots) together at one corner, which
was just past a fly-through trigger that tripped a producer at the opposite
end of the hall (behind you). You could lure two of the DCs out into the
corridor and kill them, but the third would always hide stubbornly around the
corner until you came far enough to trip the producer. Then, when there were
things shooting at you from the other direction, he'd come out to play. I
didn't, as a level designer, even anticipate much less plan this. It was the
combination of the corner and the trigger that caused it, I think. He knew
if he waited he'd have reinforcements. The only suitable way to deal with
it was to come flying at top speed through the trigger and around the corner
hard enough to push the Diamond Claw just far enough that you could get past
the corner long enough to kill him. Pushing a Diamond Claw is *not* a normal
strategy in D2, since touching you is how they kill you, but you couldn't
leave your back turned to the Seekers (coming out of the producer) long
enough to kill him, and you *sure* don't want to turn your back on a
Diamond Claw who's that close to you. It was nasty. I had to put some extra
shields at that corner to compensate the player. (The other thing I could
have done is widen the corridor, but for design reasons I didn't want to.)

Re:No.

[Dreadlord]

it wasn't really adaptive, simply when you die too many times, the difficulty level is reduced, which includes the accuracy of the enemies, their damage, and your damage, many games have similar features, Warcraft III comes to mind, the only difference is that Max Payne does it automatically, no big deal.

Well

[dtfinch]

It had better give you a score based on how tough you were.

I made an adaptive pong game once. Almost anyone could beat it. If you sucked enough, the computer's paddle would essentially stop moving altogether. If you were very good, it would predict exactly where the ball would land and become unbeatable.

Re:Well

[IamNotWitchboy]

If there was a point where it became unbeatable, then it was not very good at adapting, was it?

The essential point in adaptive AI on games is to be difficult enough for anyone to be entertaining, without getting frustrating, or the opposite, that it's so trivial to beat it that it becomes uninteresting.

Re:Well

[dtfinch]

I doubt I ever got it to play unbeatably though, as I myself am not unbeatable. The game just went on forever anyway so there was no winner or loser in the absolute sense. It was like a practice mode for when you're not playing humans.

Re:Well

[jonadab]

> The essential point in adaptive AI on games is to be difficult enough for
> anyone to be entertaining, without getting frustrating

I think you want the player to get *slightly* frustrated *occasionally*. Not
badly, and not often, but if the player always wins without putting in some
extra effort, that's no fun either. When the player's tactics and skills
stagnate, you want to start beating him some of the time. (Not all of the
time. Not, even, most of the time, I think. But some of the time.)

One way to make game difficulty adaptive, without serious AI, is to make
expectations scale with accomplishments. This is easier in some types of
games than in others. Tetris is a great example of a game that can become
infinitely hard. Is the player stacking perfectly, covering no empty spaces?
Well, then, raise the probabilities on the hard pieces. Is the player in
deep doodoo, stacked past the middle of the board? Throw him some easy
pieces. (Most tetris games don't do enough with easy and hard pieces.)
And of course there's the ever-increasing speed. Not all games have such
easy ways of raising or lowering the expectations, without fundamentally
changing the consequences of the player's actions and therefore the strategy.
Games like freeciv are particularly in need of good AI.

Re:Well

[IamNotWitchboy]

I think you want the player to get *slightly* frustrated *occasionally*. Not badly, and not often, but if the player always wins without putting in some extra effort, that's no fun either. When the player's tactics and skills stagnate, you want to start beating him some of the time. (Not all of the time. Not, even, most of the time, I think. But some of the time.)

I think that was the essence of my point, just worded differently

Tetris is a great example of a game that can become infinitely hard. Is the player stacking perfectly, covering no empty spaces? Well, then, raise the probabilities on the hard pieces. Is the player in deep doodoo, stacked past the middle of the board? Throw him some easy pieces.

I think Tetris is a bad example for adaptative AI or similar term, becasue in the essential way of tetris, you are not playing against an AI. The pieces are random, and the only variable is the speed at which they fall. Since you technically _CAN'T_ win in Tetris, you only play against yourself by lasting longer, or making a higher score.

Perhaps one of the gazillion variations of tetris would make a better example, like one of these where you actually battle someone by dropping pieces in their board or obstacles or whatever.

Adaptive AI and it's drawbacks

[TheRoachMan]

After reading that little snippet on gamespot, I've got the feeling that the game will be 'letting you win'. It states that if it takes you 20 tries to do something, the game will lower it's standards for you. Why did finishing Splinter Cell make me feel good? Because it makes me feel I've accomplished something. I've mastered the game, no matter how difficult the timing was, no matter how pixel-perfect I had to aim to kill that guy, no matter how hard it was to master. Unless they (Ubisoft) implement this Adaptive AI perfectly and unnoticeable (and I hope they will), I'm going to feel as if no matter how bad I play, or how crummy my timing is, I'm still going to master the game. Adaptive AI could really take the challenge out of any game.

Re:Adaptive AI and it's drawbacks

[jabberjaw]

Then perhaps selective AI is better left as an option for now?

Re:Adaptive AI and it's drawbacks

[SuiteSisterMary]

The counterpoint to that is that the game has to be perfect, in terms of not punishing you by interface.

Splinter Cell, for example, like many games, had quite a few points where you had to die to figure out how you should have acted. Hell, one of the opening missions, where you need to sneak past some cops shaking down a drunk, there's too much sillyness. The cops can't hear you grunting and wheezing as you go past, hand to hand, but if you climb up too fast, you're hosed.

Prince of Persia: Sands of Time is another example of this.

Re:Adaptive AI and it's drawbacks

[Cuthalion]

BUT prince of persia has the ability to wind time back so if you have to die twice to figure something out, that still makese sense within the logic of the game.

Re:Adaptive AI and it's drawbacks

[SuperMo0]

That's not the same thing, though. Making something a "two-try" situation is not the same thing as adapting to the skill level of a player... all this "two-try" does is make you think "Well... THAT wasn't the right way." It doesn't change the difficulty of the CORRECT path in any way, shape, or form, and this is what adaptive AI is about (at least, from what I've seen.)

Re:Adaptive AI and it's drawbacks

[bugbread]

While that may be true for many people, the converse is probably true for many people as well. I never finished Splinter Cell, because I kept getting stuck in the exact same area, forever. After a few days of trying, I realized that while I was spending my time in the game, I wasn't actually enjoying myself, I was just getting frustrated and angry. It's bad enough when I get frustrated or angry at work, but using my free time voluntarily to get into a bad move struck me as a pretty bad idea, so I quit the game.

Maybe it's an issue of whether it's the solution to problems or the process of finding these solutions that appeals to different players.

Re:Adaptive AI and it's drawbacks

[Pyro226]

Why did finishing Splinter Cell make me feel good? Because it makes me feel I've accomplished something. I've mastered the game...

There are an increasing number of single player games, even console ones (Metal Gear Solid 2 for the PS2), that let you upload your record to the internet after you have completed a game. In the future, "Hardest AI" may share a spot on online ranking ladders next to fastest time and most head shots.

By making it a goal to do so well that the game gets harder, you are adding another level of replay value for the advanced gamers, while making the game beatable for the novice players. Some games alreay do this on a less advanced level; the offer special extra difficult settings once you've beaten the game.

Re:Adaptive AI and it's drawbacks

[toasted_calamari]

what would be *really* cool is they higher diffiuclties were made more difficult not only becuase of AI changes but plot and map changes, this would be more work, but it would dramatically increase what you term the "replay value" of the game.

Re:Adaptive AI and it's drawbacks

[SuiteSisterMary]

Yes, to a certain point. But I think they still took it too far.

It also doesn't help that a) they have to show you a walkthrough, basically, to get you through the levels, and b) very often, I found, a pole you were supposed to grab was difficult to make out against the wall, or a jump that looked too long was doable, while a jump that was doable looked too long.

Re:Adaptive AI and it's drawbacks

[SuiteSisterMary]

Yes, I know, and that's the problem.

Besides, most 'adaptive AI' is of the 'lower the health of the beastie that's whipping the PC's ass' or 'increase it's accuracy to Godlike levels' style.

What it SHOULD be is 'PC likes to crawl along ceiling pipes and rain down death from above, so watch the ceiling' or 'PC goes for head shots for quick kills, so bust out the helmets' or 'PC likes to shoot out the lights, so bring floodlights into a room we think he's going to, wait 45 seconds after he shoots out the lights, and turn them on.'

In-game characters reacting intelligently to what the PC is doing should be the order of the day. Supposedly Splinter Cell 2 does some of this, but I get the feeling lots of it will be pre-scripted; when x number of enemies is killed, do RunToArmorRoomAndGearUp(), for example.

Re:Adaptive AI and it's drawbacks

[muzthe42nd]

a jump that looked too long was doable, while a jump that was doable looked too long Yeah, that's the same thing twice there sir...

Re:Adaptive AI and it's drawbacks

[SuiteSisterMary]

Negative.

There were many times that a you had to jump a space that was wider than jumps you previously coulnd't make.

Then there jumps that looked impossible, instant death, that Princey would simply traipse over.

However

[Joe the Lesser]

What's the main reason players get disinterested in a game? Because they come across a level that they can't beat, and they get sick of the same ol' stuff over and over again.

If less people are buying these games, because they just aren't the master that you apparently are and would rather not get halfway through the game only to quit in frustration, it hurts the company so this move makes sense.

However, to satiate you, they should add an option to set the AI on the hardest possible skill level.

Re:However

[Acidic_Diarrhea]

I agree. That's an easy fix to the problem of people not feeling like they earned it - simple options screen selection could just turn off the adaptive gameplay. One thing I would like to note is that this isn't AI. Detecting how many times a player has failed and lowering enemy health or making enemies less sensitive to movement is not AI.

The field of artificial intelligence is nowhere near having a good handle on simulating thought but Splinter Cell isn't where the breakthrough has come. And I'm not trying to pick on you - AI gets bandied about and associated with way too many products and is actually becoming a watered down term. I mean, the electric blanket I got over the holidays had the slogan, "The blanket with a brain!" printed on the outside. Again, sensor detection and response is about as AI as a smoke detector detecting smoke and then beeping.

Re:However

[shaitand]

That would depend on how the game actually goes about determining you suck and adjusting it's gameplay.

Let me give you a clue, being AI does not exclude and will certainly include analyzing statistics and responding by adjusting variables. The minute analyzing comes into play and it's something artificial doing it, THAT IS AI. Whether it's very bright or not is irrelevant. I know human beings that aren't as intelligent as Splinter Cell.

I do however believe that thanks to people like you, NOTHING will ever be entirely acknowledged as "TRUE" AI, even if it were identical in function to the human brain and perfectly recreated the chemical reactions and responses we call emotions, morals, feelings, 6th sense, etc.

Re:However

[Anonymous Coward]

Thanks to people like me? Give me a clue?

Take a few classes on neural networks and you might be up to my level.

Re:However

[shaitand]

I've written a neural network, nothing especially fancy tis true. But nonetheless I'd say it easily covers the ground of "a couple classes".

Neural networks are merely an extremely flawed emulator of the actual function of neurons... or rather something of a combination of the above and the results of our poking a stick at the above and seeing what it does.

Truth be told they have little to do with AI, and AI has little to do with them. If you want to emulate biological intelligence, you look to a neural network. If you want to create artificial electronic intelligence, you don't.

From two comments (which certainly isn't much) I suspect you either believe you are some sort of superior authority on the subject (which you obviously are not if your level can be achieved in a couple classes, or any number of "classes"), or your are the type who will never be satisfied with any sort of AI that has a mechanism and function which we entirely fathom and understand. Neural networks are such a thing, although on a micro level we understand them, we don't fully know the how and why of the way they achieve their end results.

Biological neural emulation is great and a fascinating study. But's it's not even the most successful and advanced of AI progress out there. The most probable chance (of human level intelligence, since that's make believe standard set for some before they'll call it AI) we have has to be grown and matured like a child and is in the process, thus far it has been passing the turing tests for it's ages as it grows.

Re:However

[Thing 1]

What's the main reason players get disinterested in a game? Because they come across a level that they can't beat, and they get sick of the same ol' stuff over and over again.

I stopped playing Infocom's "Hitchhiker's Guide to the Galaxy" because I couldn't open the fucking door (about 1/2 hour into it, on the Vogon ship). Yes, if it was adaptive, I would have given that game more time.

And I worked at Infocom in high school!

I'm surprised this is getting...

[Tickenest]

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.

Re:I'm surprised this is getting...

[Jerf]

Ha! Kids these days.

Try at least as far back as Astrosmash, an 1981 Intellivision game.

It keys the difficulty to the number of extra lives you have. At the lowest level it's almost impossible to lose, and extra lives are handed out generously.

I think it's actually good in a way they backed off from this; once you start playing this game it's hard to stop, because you almost inevitably have to leave a game in progress, either by powering off or by deliberately dying enough times to lose, which is about as easy psychologically.

This is at least a candidate for "first", though I wouldn't be surprised if someone comes up with an Atari 2600 game that used it (before 1981).

I'm also somewhat surprised the arcade games didn't do more of this; this dynamic difficulty level is much more addictive then the monotonically increasing (and always huge) difficulty employed by modern games.

Re:I'm surprised this is getting...

[Tickenest]

Hah! You're right. But come on, give me credit for at least not thinking that the history of video games begins with the Nintendo 64.

Check it out.

Re:I'm surprised this is getting...

[Jerf]

But come on, give me credit for at least not thinking that the history of video games begins with the Nintendo 64.

Consider credit given ;-)

Re:I'm surprised this is getting...

[Mechanik]

I'm also somewhat surprised the arcade games didn't do more of this; this dynamic difficulty level is much more addictive then the monotonically increasing (and always huge) difficulty employed by modern games.

Although it definitely helps the profitability of an arcade game to be fun/addictive, one has to remember that the point of an arcade machine is to keep the player plugging those quarters in. If they made it too difficult to lose the game, you'd never put any more money in.

The trick is to somehow take the player through some optimally contrived series of peaks and valleys wherein you addict them just enough that they will put that one more quarter in to continue should they die/fail, and then crush them mercilessly so that they are in fact forced to do so if they want to keep playing.

And yes, I know it's been a damn long time since arcade games cost a quarter, but I'm a child of the 80's :-)


Mechanik

Re:I'm surprised this is getting...

[Anonymous Coward]

NBA Jam's computer assist was bullshit!! Why the fuck couldn't you blow out your CPU opponent?? It would do things like make Hakeem Olajuwon (if you were playing him) miss a layup from 2 ft. away, or make Spud Webb miss all his dunks. Bullshit!! Meanwhile, on the computer team, Luc Longley (or any big-ass center) will start draining 3-pointers left and right, even though they had a single or no bars of skill in that area...goddamn that used to piss me off. And the code to disable the CPU cheats NEVER worked!!! I really liked that game, but that's some of the biggest bullshit I've ever run across in all my years of game-playing.

Re:I'm surprised this is getting...

[Anonymous Coward]

Everyone knows it started with the Nintendo Entertainment System.

Re:I'm surprised this is getting...

[shaitand]

"And yes, I know it's been a damn long time since arcade games cost a quarter, but I'm a child of the 80's :-)"

That's true, nodays the coin slots are mostly just for looks. After all, why bother charging yourself a quarter? And it's not like there are Arcades anymore.

Adaptive AI has already been in sports games

[WordUpCousin]

All you do is turn on "Computer Assistance" and the losing team will get a boost. So in basketball, your shots will start going in if your losing by lots. In racing, your car will go faster the further behind you are.

So how will it work in first person shooters? Make your bullets hurt more? Make the opponents not shoot as much? I guess the best way to notice is to chase them around without shooting and see if they just run away.

Re:No.

[Goldberg's Pants]

Warcraft 3 had one? Are you sure? Post links please:)

A lot of racing games have had adaptive AI's. Gran Turismo certainly does. And Papyrus' NASCAR Racing 2003 has it as well, it's an explicit option you can turn on, and it seems to work pretty well.

Its nothing new

[5 Second Rule]

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.

Re:Its nothing new

[77Punker]

This was also used in F-Zero X for N64. The difference between driving the game like a grandma and an expert had on the AI was as much as 50 seconds on the AI's total race time. Also, I recall a vector game I used to play on MAME every now and then that had a setting on the dip switch for difficulty called "Adaptive". The game was from the early 80's, so it was probably one of the first attemps at such a thing in video games. Unforunately, I didn't notice any difference between the difficulties when playing.

Re:Its nothing new

[one4nine4two]

Not to mention Super Mario Kart and Mario Kart 64 (maybe Double Dash too, haven't played it much). Maybe it's just because I had gotten quite proficient at the games, but the adaptivity just pissed me off. I would lead an entire race, fending off multiple attacks from racers who I should've been going faster than, only to get passed at the last second by Toad when I'm flying top speed with Wario.

I hear Diddy Kong racing didn't have this adaptive AI stuff, so if you had a lead, the computer wouldn't cheat to catch up.

Re:Its nothing new

[_Sexy_Pants_]

I don't know if I'd call this adaptive. In fact, I tended to just call it "cheap". Adaptive would be more like if you tend to shoot shells behind you, and the computer keeps a shell handy at all time to counter that shell

Re:Its nothing new

[kisrael]

Yeah, Diddy Kong was great...very fair, and the same items always appeared in the same places-- so as long you were more than a split second behind someone, the item would regenerate. Plus you could stack items, up to 3, each level more powerful. Also, great karting, airplanes, and hovercraft modes, and better battlemodes than Mario Kart ever had.

On the other hand, races among good players tended to be decided in the first quarter of a lap or so, it was TOUGH to catch up. And the tiny pipsqueak characters (mouse and turtle) seemed to have a big advantage, at least among the people I was playing.

Re:No.

[Dreadlord]

Yep, sure, I couldn't find a link over Blizzard's site, but when you lose a game, a dialog box will appear asking if you want to reduce the difficulty level, check out the faqs at gamefaqs.com, I remember that a faq there mentioned this feature.

Max Payne

[ArmenTanzarian]

What that game really needed was a way to adapt to your attention span. When he would launch into one of those speeches about the rain on the pavement and how it made him feel, I wanted to shoot my Playstation.

Re:Max Payne

[Anonymous Coward]

Next time the urge strikes you, do us all a favor and shoot your computer.

Re:Max Payne

[Merusdraconis]

So is it or is it not true that guys don't want to know about feelings? I mean, here's Max Payne, badass extraordinare, taling about the goddamn rain, and here's you, a Slashdot user, getting pissed off at his sojourn into the realm of rain.

Damn, that made no sense.

problem with AI and difficulty

[*weasel]

the catch 22 is: make the game too easy, and players will complain. make the game too tough, and players will complain.

personally, i don't think a 'difficulty' slider should come into affect with AI. The AI should always -try- to behave the same way.

Whether you intend for them to be tacticians, civilians, or just mindless grunts. on 'Easy' or 'Difficult' a bad guy should still know he should take cover, call for backup, etc.

The 'difficulty' should come into play when deciding their accuracy, movement speed, 'scoring' (penalties for shooting hostages, raizing conquested territory, etc). It could also come into play in deciding the scarcity of resources. on Easy, there should be extra resources for the hero, and less for the enemy.

Adaptive -AI- is the wrong approach. Adaptive -difficult- is still a good idea though. but don't make enemies dumber; just make them slower, more inaccurate, fewer in number - don't give them as many grenades and leave more health packs around.

oh, and i also don't appreciate the 'difficulty' sliders that just scale the damage you receive up and down. that is an awfully 'cheap' hack imo.

Re:problem with AI and difficulty

[nick_davison]

Whether you intend for them to be tacticians, civilians, or just mindless grunts. on 'Easy' or 'Difficult' a bad guy should still know he should take cover, call for backup, etc.

I'd disagree:

A typical civilian is liable to run scared, shoot while running, empty their clip desperately, that kind of thing.

A regular soldier is liable to call for cover, move to a braced position or whatever for shooting, but still make some dumb mistakes.

A truly elite solider, the target is probably never going to know they and their squad were there until they are lying dead on the ground. They'll plan tactics in advance, have them well practiced and executed silently. They'll be better at scanning for targets, they'll be better at moving silently through the best areas of cover. When they do move, they'll be covered by an already well placed squadmate. Their communications will be as much via silent gestures as noisy radios.

Within that spectrum (and me not knowing much about elite forces), there's a huge range of variables for a given enemy. That security guard could be nothing more than a civilian in a uniform on easy difficulty yet an ex special forces soldier on high difficulty. They're still both "just a security guard" yet there's a huge range of differences in how their AI might act.

Ghost Recon has my all time favorite example of that. On the first mission, if you go around to the north of the valley, there's a slope that troops come down.

I'd played it once on easy. I managed to clear them all with just a single sniper and good reactions. As soon as they started getting shot, they just tried charging, firing wildly, hitting nothing.

On elite, the sniper died almost as soon as he gave away his position as half a dozen guys found cover, found him, lined their shots, then took him down.

Next I took a squad of three guys, including a light MG. I found some bushes near the bottom of the hill. I waited for them to come in to the open, then opened up. Quickly they dropped to the ground. Then they started moving in pairs back in to cover behind a boulder, the others providing supressing fire. As they kept the range between them and I, they were much harder to take down and I got maybe two of the six before they were safe. Now it became a case of do I have to clear them out or will they come after me? They answered that for me and came after me. Yet even then, they maintained great covering fire, from braced pairs who were taking advantage of the accuracy, moving from cover to cover.

In both cases, they were just about as accurate as before (they just used more accurate firing positions) and could take just as much damage (Ghost Recon is great for accurately handling how much mess a single bullet causes). It was entirely down to their use of tactics that they went from being a group of idiots to mop up to scarily hard adversaries.

Granted, those were skill settings, not an adaptive AI system. Still, that's what differing AI levels should be like. I can't stand games that differ difficulty by making shots do more or less damage, by simply upping numbers of enemies, by suddenly making enemies perfect shots while they still move in exactly the same way they always did.

Re:problem with AI and difficulty

[shaitand]

Personally I'd agree and disagree with both of you.

For the most part I agree with you, my direct parent poster. I don't believe that more or less skilled players should recieve different amounts of mana from heaven and increased health, or decreased enemy health etc.

I believe that there should be two basis on which to analyze the player (high level basis that is) skill and intelligence. The two are certainly NOT the same thing, although a player may be skilled and intelligent both. Basically the AI need only mimic the player in these respects.

Skill would be things like REACTION speed, enemy or beastie A should not move slower because a player is not skilled, enemy or beastie A should react slower and/or be less accurate when firing. Fast reactions and accuracy can be natural or trained but BOTH are matters of skill.

Intelligence is something else altogether, having more to do with tatics. To think of clever and workable tatics one must be intelligent and/or knowledgable. So the AI's tatical manuvers should be adjusted in this case.

A skilled and intelligent player will of course encounter AI that is very clever and employs it's clever tatics quickly, is skilled with it's arms and rarely misses a difficult shot.

An untalented and ignorant/stupid player will encounter AI that knows few moves and reacts slower, that AI is not as good a shot.

Of course the difficulty of the shot must be considered as well, even the most poorly skilled AI will never miss at point blank range, although it may hesitate when it should pull the trigger.

There should also be a lower boundy, the game only gets so easy. Since there is never mana from heaven, and health and speed never magically increase there is no need to worry about upper boundry, upper boundy is limited by the programmers ability to code intelligent tatics. An impossible shot is an impossible shot of course, but lightning fast reflexes do not make for an impossible to beat game.

Re:problem with AI and difficulty

[jonadab]

Adaptive AI is needed for some games, where "intelligence" is a key feature
of playing the game. (I put "intelligence" in quotation marks because I'm
using the traditional definition that includes quite a lot of things that
computers can do, such as examine by brute force all the possibilities for
the next N moves in a chess game, or test various board positions against
every single word in a large dictionary in a Scrabble game. If you think of
intelligence in terms of abstract reasoning and qualitative learning, then
we'd be doing well just to make a game that can do those things at all, much
less adaptively.)

Re:problem with AI and difficulty

[NEOtaku17]

I completely agree with this man mod him up! If I die I don't want my enemies to just stand there and let me shoot him, a game like that would be boring and dumb. If I'm really that bad I would want the game to help me aim better and kill the opponent more easily, not make them more unrealistic.

Re:problem with AI and difficulty

[*weasel]

i don't think i conveyed myself clearly as i think the original responder did not get my intent.

yes, i think that -reaction- speed should be a factor adjusted by difficulty, not movement speed.

and yes - i never meant that a civilian should be an expert tactician if you turn a game to 'difficult'. i was suggesting that a civilian should behave like a civilian no matter what the difficulty. just because you select 'hard' doesn't mean that joe-blow on the street is suddenly a kung fu master. similarly a military opponent shouldn't just stand still and fire 1 shot a minute at you on 'easy'.

of course, it isn't all misunderstanding: i still do not think enemies should change their tactics based on a difficulty slider. perhaps adjust their 'resolve' to determine at what point they flee/surrender - but i don't think sudden increases in tactical operation are appropriate.

quite frankly that's a waste of developer effort, to make the enemy more intelligent, and then have 80% of the game players never encounter that intelligence. it is not an efficient use of AI resources, especially considering that every game to date has had some basic AI that could use much more attention. Perhaps if all else were perfect, i wouldn't mind. but when enemies have code to allow them to use cover and suppressing fire on 'hard' - but they all have terrible pathfinding on any difficulty -- it makes it a bit hard to understand.

increased number of enemies doesn't always make sense on a difficulty slider, but sometimes it can. the difference for me on that point is: does combat still essentially play the same, or are they just throwing waves of mindless guys at me to test ammo management?

increasing the -number- of sentries in a game like Deus Ex makes sense. but increasing them to the point that sneaking and hacking become essentially impossible doesn't.

i only suggested scaling up enemy accuracy as one of many factors. it is not a silver bullet, but it's much preferrable to simply giving enemies bigger guns or making their shots do more damage.

and mana from heaven is not at all the kind of difficulty adjustment i was proposing. i was thinking more like adjusting the number of friendly units you have at your disposal or adjusting the prevalence of resources in a strategy or sim game.

I've seen this before

[Film11]

I've seen so-called "adapative AI" before. I can't remember what the games which utilised it were called but what I think they did, is take your score and increase or decrease the AI's skill accordingly. Although in SC:PT I think this might be actual in game, instead of calculating between levels. Either way, I don't think its that hard. For programmers, not me.

Guess it comes down to the type of game

[hrieke]

A RTS game and FPS game are two very different beasts, and how the AI acts in those types of games can either kill the game play or make the gmaer keep coming back.

One thought on this subject is that the game should know what the objectives are and allow the AI engine figure out the best way to achive the goal (as in RTS), which could generate interesting and unpredictable game play. Back when I was playing C&C, I quickly figured out that the AI could not deal with walls very well, which was a very simple solution to do a land grab by building walls everywhere.

For FPS, AI should control each monster / enemy and allow them to work as a team. A few games do this, but I don't think we'll see real inteligence until 64bit CPUs are more common and there are enough spare CPU cycles to compute their actions.

Another thought is to have the AI improve the more you play the game and allow the game the ability to learn from you as much as you learn from the game. But somehow I don't think we'd allow the game to spend the night analizing a few hours of Quake III playing just so the next time it can kick our ass...

Re:Guess it comes down to the type of game

[shaitand]

The C&C AI is easy to beat, build almost nothing, just mine ore as fast as you can and build a shitload of tanks. Make sure your ore is coming fast enough that you can build tanks nonstop, whatever you are doing, it's secondary to keeping the tanks rolling out.

When you have a sufficient legion of tanks, send them out and crush the enemy. If you don't crush them that's ok, so long as you never stopped building tanks. When you have another legion of tanks send them out. Of course choose your targets strategically, you want to cripple the enemy after all. And always makes sure each of your harvesters is guarded by a couple tanks.

When playing the computer building up is a waste of time. The computer of course will try to build up, and thus it will waste ore and ultimately never be able to win. If ore permits it's best to be able to build two tanks at a time of course. Three is wasting ore that could be used to buy tanks!

Re:Guess it comes down to the type of game

[Anonymous Coward]

That's way too risky an approach. The wall game is the only sensible strategy. Sandbag all the way out to the enemy base, sandbags all the way round the enemy base, wipe out any enemies that remain outside. You now have the entire map to yourself. Harvest everything, build your mighty army, and the average enemy base will last around ten seconds.

Cheap? And?

Re:Guess it comes down to the type of game

[shaitand]

bah, it's risky against another player. I haven't ever lost even ONE scenerio against the computer this way.

Walls are nice and all, but they cost too much money and take two seconds to blow up. Hell the 3rd rank or so of my tanks hasn't even fallen into place before they've already eaten through your wall. Of course you have to manually control your tanks and concentrate the full force fire on individual targets so that most everything dies or is blown up instantly. Dividing your fire is great, but you'll find reducing enemy numbers rapidly is a better technique then spreading fire among multiple enemies. Each enemy taken down isn't shooting you anymore.

I've tried the wall technique, because the computer AI does put FAR too much weight to walls and trying to go around rather than through them. C&C has weak walls so this is stupid. I've found you sometimes lose with it, because it diverts attention and energy away from building tanks. And valuable ore, later in the game you will probably make up for it, but by then the computer has it's own army. Hell by the time I have a legion of twenty tanks coming in the computer maybe has about five.

Possibly not Possible, Plausible or Wanted

[daki]

AI has it's limits as we all know. It after all is still just a computer and computers, allthough being able to do things faster than humans, cannot act like a human. It still requires the user. Now adaptive AI is a thought, and seems to work well from the games I've played (Splinter Cell, pretty much anything by Tom Clancy), allthough I think there is a limit on not only what the computers are capable of, but also what we want them to do. The whole premise of video games is to get an adreneline rush by doing something we normally would not do, but a completely realistic game is not at all wanted. Would you go to the movies to see a James Bond movie where James gets shot in the first fight scene because he cannot avoid the hundreds of shots coming at him while he is running around? I don't think so. We need our heros to have a bit of a superhero quality and our villans to have inherant flaws. Having a real video game would suck, because as opposed to being able to be shot multiple times in some games, one shot and you would be down pretty much. As for the AI, we need them to be stupid, and not be expert marksman. We want challenges, but if these situations were happening in real life, there is absolutely no way the person would survive. Anyway, thats fust my thoughts on the matter. =)

Re:Possibly not Possible, Plausible or Wanted

[shaitand]

I'm with you on the rest but please tell me haven't bought into this mumbo jumbo that humans are anything more than biological computers?

Re:Possibly not Possible, Plausible or Wanted

[Anonymous Coward]

Funny - AFAIC the mumbo-jumbo is the theory that humans are nothing but biological computers. When I see a computer that approaches the capabilities of a human, or even a theoretical proof that it's possible, I'll reconsider that one.

Re:Possibly not Possible, Plausible or Wanted

[shaitand]

There already are plenty of theoretical proof's. But I'll make a new one for you. Humans are merely biological computers, therefore it's theoretically possible to equal their intelligence with an electronic computer.

There you go, theoretical proofs. Hell it's more solid than MOST of the prevailing theoretical facts out there that we consider foundations of modern science.

When YOU will see a computer that approaches the capabilities (I assume you mean intellectual capabilities, it might be awhile before you see the robotic technology and AI technology combined in a package that is equal to or superior to a human) of a human is one thing.

But several have seen it, and psychologists have spoken to a computer AI which is being grown like a human being and so far hasn't had any trouble fooling those child psychologists into believing it is a real child of it's age at every stage of the game. Because it's being grown however, it will be another decade before we find out whether it will continue to successfully learn until fully grown.

Re:Possibly not Possible, Plausible or Wanted

[Anonymous Coward]

"Would you go to the movies to see a James Bond movie where James gets shot in the first fight scene because he cannot avoid the hundreds of shots coming at him while he is running around?"

Yes I would! It is very hard for me to suspend disbelief and enjoy the movie when it is so obvious that some of those bullets should be hitting him.

"We need our heros to have a bit of a superhero quality (by not getting hit by the bullets)."

How can we view the hero as being a superhero when he's not being hit either because of sheer dumb luck, or else because the bad guys are incredibly bad shots?
Unless of course we posit that he is a superhero because he has a superpower called "incredible luck."

If I remember correctly,

[OgdEnigmaX]

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?

Re:If I remember correctly,

[Samhaine]

Nitpick... it's 10th Kyu to 1st and 1st Dan to 10th That is all.

Re:If I remember correctly,

[h0mer]

When did you notice the difficulty start going up? I was 88-2 and got sick of having to win 500 matches to enter the tournament for each arcade. I found the only difference between 1st kyu and 1st dan to be that against the dan, you have to block.

Re:If I remember correctly,

[way2slo]

"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."

That would be interesting to play against, but I don't think it would be very realistic. Each AI person that you enounter in the game would have no idea what your tendancies were at first. They should not know how you fought in the last room, unless they saw you. Now if you were in a room full of guy and you did the same move to kill off the first handfull, then it would be realistic if the AI controlling the individual guys would adapt and counter your move. Otherwise, if they "just know" how you have been fighting the whole game, how would they explain it "in game"? Does one of them make a phone call while you are killing his friends? Are they the Borg?

For example in Prince of Pursia: The Sands of Time, I always tend to use the vaulting head-slash move, and the computer bad-guys just stand there and let me do it. (unless they are a specific type, then they always counter that move) Instead, they should have a small chance of blocking any move at first, but the more I use a move in a row in the same room, the bad-guys in that room should have a better chance of countering that move because they expect it that much more.

Re:If I remember correctly,

[OgdEnigmaX]

I'd have the information propagate according to some set of the following criteria:

1) Whether the agent in question has seen the behavior's effectiveness (necessitating a crowded room, a narrow escape or, yes, a telephone. I imagine that one might bump into ridiculously elite commandos that are in constant radio contact, for example)

2) "Ingenuity factor" - whether the agent is apt to pick up on such things and develop effective responses (how quick on the uptake are they? Some folks might never understand how you do what you do, regardless of how many times they see you hit enemies in the legs to stun them before finishing them off and how many warnings they receive from their compatriots about it. Furthermore, some folks might never figure out an effective counter to your tactics, or, if they do, might not have the time or resources to implement it. Not everyone has access to kevlar long-johns, and even those that do might not be able to get them in time to protect themselves from the action hero about to break down the door)

3) Whether the agent is talkative (percentage score from zero to one, adjusted by faction interrelationships...how much does group A care about the fate of group B?)

4) Whether the potential receivers are receptive, and how adaptable they are (ditto, again adjusted by faction interrelationships, maybe by rank as well...a good leader is far more likely to get his subordinates to do something than a grunt is to do the opposite)

5) Whether or not there's a language barrier, if you happen to be that sadistic

Of course I don't think the full set of rules would be appropriate for most games. The Max Payne series, for example, is a little too arcade-oriented for AI this involved. I think the propagative ideas would only work in a game with massive scope, like some tactical mix between Deus Ex 2 and Vice City...which is _hellish_ to think about design-wise :). I've never seen a game that huge with combat that sophisticated.

The other major problem remains how one goes about designing AIs that can adapt. How does one break down player behaviors into discrete parts that can be assigned weight and identified as significant, and thus as candidates for adaptive behavior? I imagine some variant of the VF4: Evo body region tracking could be implemented, assembling a profile of what goes into a kill and a loss (escape, I guess, in this case): method of attack (weapon) and area damaged. "Mmm," says a reasonably smart AI. "75% of the kills I've seen him are resulting from headshots, which account for 90% of the damage he's incurring in those cases. People he fails to hit in the head, however, tend to escape. I'll be sure to keep myself better covered in the future," potentially reducing accuracy ("No way I'm peeking my head out!") but potentially increasing survivability in the long run, and in any case forcing the player to adjust his tactics.

I know this is probably a recipe for neural net hell, and the fact that firefights are usually really chaotic and over very quickly makes the idea of a lucidly observant enemy kind of silly (though it would be a lot of fun, I think, to code a text-based proof of concept), but I do think this sort of social AI, or whatever you want to call it has a place in at least *some* genre. So FPS behaviors are hard to break down. It might do well in an RPG, for example, in which spell slots are few (that is to say, not at all like the Final Fantasy games, in which nearly every spell tends to be at your disposal all the time), and as such the choice of defensive spells and counterspells to prepare before battle is a fine art. After a few months of adventuring, perhaps news spreads among the Orc tribes that a human wizard that wields fire has been cutting a swath through the settlements near human lands. Their mages, however few, might begin to research anti-fire spells just in case, and deploy duly equipped mages and magic items near the border zone in order to conta

This has been around a while...

[Asprin]

IIRC, Doom and Quake had adaptive AI, too.

No, I guess it doesn't really do anything different from a random number generator.

Re:This has been around a while...

[waaka!]

Didn't Doom and Quake have user-selectable difficulty levels? Not sure how that would be adaptive, really.

Re:This has been around a while...

[addaon]

Dragon Warrior IV is the first game I can think of that had real adaptive AI, but it was AI on your side... the bad critters never changed but, in Chapter 5, your team members would keep track of "what worked" (seems to just be how close an action was to a kill), and increase the probability of choosing that action. It was a good implementation, too. They started with 5 (I think) user-selectable probability distributions (defensive, offensive, etc), and when a given one was selected, training would only happen within it... so you essentially have five modes within which learning takes place, separately, and the user could (and occasionally really should) disable those modes which learned weird habits.

Re:No.

[Anonymous Coward]

According to and old chat on the TuxKart-devel mailing list, racing games have to have an adaptive AI or be very frustrating- in a real, all-out race, if you fall behind once, you're hopeless unless the other drivers all make mistakes the equal of yours.

Usually quite simple. If you're behind them, they slow down slightly; if you're ahead, they speed up.

New SC Pandora Tomorrow Trailer was put out today

[AIX-Hood]

Kind of on topic, that they happened to release a new trailer for Splinter Cell Pandora Tomorrow today also. Torrent: http://www.filerush.com/torrents/splinter_cell-pan dora_tomorrow-trailer2.zip.torrent

State of AI in games

[j-turkey]

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?

Re:State of AI in games

[Anonymous Coward]

There's two reasons for this.

First, AI isn't given much processing time in games, especially real-time games. Most of the processing time is going to graphics, physics, etc.

Second, before AI can really get "good" in any satisfying sense, there's a LOT more theoretical work that has to be done. And game developers, God bless 'em, are practical creatures - they need to get games out the door, so they're not going to have time to futz around with inventing new algorithms.

Re:State of AI in games

[ajd1474]

I think people are confusing difficulty with intelligence. Use a game like Dead or Alive or Tekken as an example. Increasing the difficulty adaptively would involve the AI hitting faster or using more advanced moves the better you got. That is not adaptive AI, that is adaptive difficulty.

Adaptive AI in this case would be more like that if you only use a certain set of attacks the AI learns what you are likely to do and defends appropriately. Just like a real person would. These would be the end of using set-patterns of attack against a stupid AI. Its not just to make things harder, but to make them more realistic. No more "left jab, right kick, knee, knee, knee, rinse and repeat"

Or in a game like Age of Empires, if your strategy is to always use cavalry to rush the opposition, the AI would learn and create Pikemen to counter your repeated attacks. Having AI which adapts to your moves forces the player to continually revise their strategy... not neccesarily make it more difficult, but forcing the player to ALSO become adaptive instead of using a tried and true method to beat the AI.

Re:State of AI in games

[DeadScreenSky]

You are essentially correct, though some RTS games are starting to get some very nice AI. I recently got into the RTS Rise of Nations, and the AI is nicely clever in it. Not super "same as an expert player" smart, but much better than what I find in games like C&C and Warcraft. The fact that only the highest levels of AI cheat, and the game tells you how and warns you, make it even better. Very fun game, too, if overwhelming at first.

Though even the new C&C Generals has some nice AI for probing defences - generally if they are killed too much going head-on towards the center of your defences, they will try and attack the far edges of your defence. Still not really smart, per se, as it relies too much on suicide assaults, but it makes for a more exciting game. (And on a similar note, I recently discovered there are mods for Generals and Yuri's Revenge that do nothing but improve the AI. Haven't tried it for Generals yet, but the ones for Yuri were truly impressive, breathing completely new life into my various LAN battles.)

I think part of the problem with RTS games is just that making a truly good AI would be ridiculously hard, not even considering CPU power. If you look at a really strategic game like Go (and to a much lesser extend, Chess), it is just really hard to make a good AI. A good RTS should at least approach remotely a similar level of strategic depth, but compounding the problem is that the games just aren't around very long. We haven't had twenty years to perfect the AI for a game like Civilization III, for example. So it cheats like crazy to compensate. :(

I imagine things are getting better overall, but we still have a long way to go!

Re:State of AI in games

[CuteAlien]

AI has advanced in several fields within games, but there are some reasons why learning AI is usually not used.

One of the hardest problem is doing leveldesign for an unpredictable AI. Hell, it's hard enough to do leveldesign for those unpredictable players already :). When you give up control over the AI behaviour (and you have to let loose control somewhat when you allow your bots to learn), it's very hard to tell in advance if the bots will learn behaviour which will advance the fun for the gamer. Maybe the AI will learn to use moves which make it harder to kill, but just look stupid (duck-jumping instead of running to be faster p.e.).

It's just preferable to make your AI as good as possible from the beginning and once it's getting hard enough to finetune theire level for a good game (making it more stupid is always easy!).

AI has nevertheless advanced within the last years. Don't let yourself confuse by the bad AI which are still released much too often (those will never vanish). But maybe the advances are not so obvious like in 3D graphics. Good designers hide the weaknesses of the AI since years, so except the obvious bugs you could often not see what was missing in older games. AI in 3D was the first big step. AI in 3D where bots can roam whole levels was another. Then came the Levels with stairs and other more complicated geometry (taking care of roofs is hard!). By now bots start taking cover behind simple objects and more complicated cover-finding algorithms are in the work. Physics makes it harder once more (even thought i don't know if any games do really handle that well yet - they certainly will in the future). Also group behaviour has increased very much within the last years. Pathfinding nearly ain't a topic anymore, some bots don't even need waypoints anymore but use polygon meshes instead.

I'd say AI is getting better about at the same level as the gaming worlds are getting more complex. Maybe it would already be possible to make bots which seem more intelligent in a game with reduced world-size and simpler geometry. I guess it'll be possible to do a quite good AI for a sims-like game (how about some intelligent pets? I'd wish this company would release the script language). Some years ago creatures did just that - nice graphics, but technically they were years behind, but advanced AI (even learning!). Unfortunatly complex worlds seem to be more in demand currently.

Re:State of AI in games

[j-turkey]

AI has advanced in several fields within games, but there are some reasons why learning AI is usually not used.

I spent a little bit of time in college working with AI and evolutionary computation and I can say with very little authority that AI has, in general, come a long way in the last 10 or 15 years. Also, my post was probably more of a semi-ontopic rant than anything else (with very little to do with adaptive AI). Still, I don't consider AI with adaptive difficulty all that far from a strategicly adaptive AI -- the difference is a few different variables, and a little more time spent thinking about the AI strategy.

However, to address your response, it is clear that the state of AI has actually progressed -- and I see the point that you raised about the level of AI rising with the complexity of gameplay (particularly with the addition of the Z axis to many modern games).

What I'd really like to see (to finish off my rant) is a more strategicly adaptive AI, where (for example, in RTS games) a computer opponent will start with a basic strategy, or even a pseudo-random choice from a variety of basic strategies. The basic strategy will have a number of variables in it -- like a defensive/offensive scaling variable, as well as a few others. When things don't go so well, the computer can randomly change tactics -- with a limited amount of total tactics, but just enough to keep it interesting. Strategy variables are altered until the sytstem finds that the fitness function is being met (winning the game, kill/loss ratio, etc) -- that strategy is used until the game is over, or the fitness function is no longer being met.

By now, it should be pretty clear where my knowlegde of game theory and AI ends -- but I'll still stand by my basic point. With a decent AI engine, I should be able to play a computer opponent which is different every time, and manages to be challenging without the AI having to cheat.

Re:State of AI in games

[CuteAlien]

Maybe another problem of AI in games are the release dates. AI is always finished last (the reason is simple - it's hard to finish the ai when the game ain't finished yet). Adaptive AI needs a system as basis which can already play the game - so the reason why it's usually not done might also simple be the reason that no team wants to wait some more weaks for the AI to increase some more, when it's already able to play :)

i'd love to see an adaptive ai in a rpg

[metalmario]

if they don't make it, i will myself. ;) already wrote one in my master's thesis...

Max Pain in Max Payne

[deblorvayn]

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.

Re:No.

[hubertt]

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.

Pffffffffft!

[msimm]

I want Weenie Level. It keeps me honest.

Its Not Adaptive AI

[Zevets]

It will be Adaptive Difficulty.

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.

Adaptive AI has porblems

[node159]

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.

AI?

[DarkGamer20X6]

So, basically, we're talking about a "pacing" issue, right? It's actually a relatively simple, and often used, technique in videogame AI.

Create a fitness value for the player based on whatever criteria you feel best indicates the skill level of the player (Time to complete a stage, Amount of damage taken, Accuracy rating, etc.). Measure the player's actual value against some expected value. Then, adjust certain game parameters (Enemy firing rate/accuracy, Availability of health packs, etc.) to compensate for the player's skill level.

There are certainly details to consider, such as how far back we consider the player's fitness criteria (the last few levels or the entire game?), the frequency/magnitude of "adjustment", or what expected value should be used for comparison. However, these details are usually adjusted through repeated trial-and-error testing.

The new Splinter Cell doesn't seem to hold much more to it's AI than that, but of course, I couldn't say; I didn't develop the game.

More sophisticated adaptive AI in videogames (incidentally, the subject of my thesis) is certainly feasible and has been done to a slight degree already, but as the focus in videogame development remains dependent on a) making the game "look" good and b) getting the game to market, AI takes a back seat and programmers must rely on quick and easy solutions, like pacing.

The real question is: When videogames begin using more sophisticated techniques and stronger AI, will many people believe it? This is the Turing Test, only you know ahead of time that you're interacting with a computer.

Re:AI?

[Anonymous Coward]

Is your thesis available somewhere? It's the subject of my Ph.D. thesis too, and I would be interested in seeing your work, since there is very little scientific work in this area.

That's the problem with AI.

[Ayanami Rei]

Once it's obvious how to do something that at first blush seems like an "intelligent" thing to have in an automated system, it suddenly becomes "not-AI".

Face recognition? Oh that's just statistical analysis, that's not AI.

And so on, and so on.

It's AI, either because it's a smart behavior that you didn't expect it to exhibit (you being the player), or because you didn't have to show it explicitly what to do in each situation (you being the programmer).

I think that's about as good as a useful defintion we're going to come by for AI.

Re:No.

[Goldberg's Pants]

That's more of an adaptive difficulty level though. Still, pretty cool. I didn't know it did that. (I only ever play skirmish, that's probably why.)

Re:If you are really bad, the enemies kill themsel

[Anonymous Coward]

If I have aim, there is no hiding; they're already dead.

Re:No.

[Goldberg's Pants]

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.

Re:No.

[Goldberg's Pants]

Not going to happen. At least not for a while. Too advanced I think. You can see rudimentary examples of this in some beat-em-up's. I can't remember which one is was now. Early to mid 90's anyway. Prior to that, a lot of them, you could find the weak spot and defeat the AI by repeating the same move over and over. (Generally a sweeping kick in my experience.) Then it came along... May have been Mortal Kombat, not sure. But the game learned. After doing the same move 2 or 3 times in a row, the game got wise and countered you.

I can't think of any other game off hand that has an adaptive ai that would count as learning, at least not one beyond "he's done that move three times, bet he does it again".

Feel free to prove me wrong as I would love to play a game that learnt and adapted.

D'oh!

[OgdEnigmaX]

Whoops, got the polarity wrong. It's been a little while since I've picked it up :)

Re:No.

[hubertt]

I can't think of any other game off hand that has an adaptive ai that would count as learning, at least not one beyond "he's done that move three times, bet he does it again".

You are right, it's not going to happen anytime soon. Adaptive algorithms are known - they are used among others in OCR but they are not widely used in games. As far as I remember, some of the chess programs (Chessmaster? I'm not sure) used played games as a library of moves for analysis. But this probably counts for "he's done that move again".

Most of games' enemies are just reacting to the behaviour of the player, they are not analysing the behaviour. But do we really need it? Would you like to play the game that is getting impossible to win if you are too good? I prefer the "learning curve" - when new enemy appears (in FPS of RTS games) I try to find the weak spots and behaviour patterns of the new guy and then use it against him. After that, I like it to be predictable, in fact!

Killer Instinct

[sofakingl]

I may be wrong about this (it's been a while since I've played it), but I remember hearing once that Killer Instinct had something like this in it. I do remember that certain combos didn't work so well after using them enough times against the computer player.

Re:No.

[shaitand]

"that is getting impossible to win if you are too good?"

The idea is that the difficulty should scale. Thus an uber player should be equally challenged as a first time player who has barely figured out the controls.

For you who is becoming a better player the game shouldn't get easier as you become more skilled, however it shouldn't be harder for you to win either, rather it should remain equally challenging. No this doesn't serve if your the type who likes to show his friends what a bad arse you are because you can wail through a game.

Re:No.

[cgenman]

Well, adaptive AI in gaming has always meant that the AI adapts to the player's skill level. What you want is a Learning AI, which in gaming generally means that an AI adjusts a slider to better counter a move that a player will make.

This isn't rocket science people. There have been, for example, nearly perfect Ai's that have played within the rules of a game and can still kick a player's tail. All games are developed with the idea in mind that every move has a counter, and every counter has a counter. Now, the AI development team knows the best instantaneous moves to counter other moves, and as they are the development team know most of the higher-level strategies that will be tried. If a development company wants an AI to beat you senseless, it can do so without changing the rules of the game. Is that fun? No. So you weaken it, and change difficulty levels around. Now you have a system that isn't learning, but is playing with the player. And playing is fun.

Re:No.

[addaon]

See my comment here for an old, old case that I do consider learning.

Adaptive AI / difficulty

[mobby_6kl]

If by adaptive AI in this case means that the performance of your opponents depends on how well you are doing, than NFS:Underground is an excellent example of this. Do a few 360s in one lap, pretend you are playing Carmageddon, or just crash somewhere. After this you would be 13sec behind and and next lap you are just two seconds behind the leader(no major mistakes). On the other hand drie like mad, full speed, good turns and no serious mistakes, but the opponent just flies past two turns before finish. This is really annoying, and looks like not only for me, but for lots of other people too.
Anyway, a clear example of adaptive difficulty in games would be Unreal Tournament. It can be turned on/off, and has an easy to test effect.

Re:Adaptive AI / difficulty

[Pulse_Instance]

When I first started playing UT, I found the Adaptive AI setting, I turned Bot AI to lowest, then played for 30 minutes, I took a look at the AI again and it was GodLike.

Make adaptive AI an optional setting?

[ScorpiusFan]

Adaptive AI seems like a good feature for some people. Why not make it an optional setting? Those who prefer the raw AI as it is can play it at that setting, while those who don't want to risk being held back by an overly frustrating opponent have a fallback to ensure gameplay is still fun for them?

not really...

[Anonymous Coward]

it's just another fancy name for getting more cheap.

as you kill more baddies without getting hit or killed, their damage/sensors increases until they kill you.

Here's how you keep people motivated

[TerryAtWork]

I was taught on a course the way to keep people motivated is set it up so they get %75 success rate.

Seems the AI should always move towards that.

the problem ive found atleast in FPS

[cyrax777]

is that it would just make the computer cheat instead of regrouping and such in the case of Unreal Tournement atleast it would just spawn in a couple of uber bots that could move faster then you could and could hit you if they could see you. Not to mention teleporting around the map at times.

Adaptive AI

[ThePyro]

I'll be the first to praise Descent's great AI, but I honestly don't think it's adaptive at all. Doing a search yielded no pages that indicated an adaptive AI. I even found an interview with one of the developers, and although AI was discussed briefly, no mention was made of adaptability.

The Descent robots were definitely smart - they could find you ANYWHERE in a level, could call for reinforcements, and some knew how to sneak up behind you when you weren't looking. But they didn't adapt to your playing style or learn from their mistakes. There were five difficulty levels: Trainee, Rookie, Hotshot, Ace, Insane. From what I can remember, the bots were just a little bit better at dodging on the higher difficulty levels (and did more damage, too).

Its really annoying

[bluGill]

Grand turismo had this, and it was really annoying. I'd have a great run, finish some race in 3:20, but take last place. Next attempt I'd take first despite taking 3:50 to finish the same race. Happened all the time, you had to force yourself to drive bad because when you did a good job you couldn't win.

What I hated most was taking a corner at the fast speed the car could handle, and seeing a car that handles worse pass me on the corner and not spin out afterwards. In other words it wasn't adaptive AI along, but a cheating adaptive AI that I hated.

Re:Its really annoying

[SurgeryByNumbers]

This is called "rubber-band AI" because it prevents you from getting so far ahead of the competition that you get lonely out there. That's the theory, at least. What it amounts to is that the enemy cheats to make up for poor performance, and that's a frustrating approach to AI development.

Older versions of Mario Kart had the same problem. Nothing like having a tiny shroom pass your heavyweight (high topspeed) kart like it was nothing. In Double Dash, either they fixed it or replaced it with real AI, because it hasn't happened to me yet. Definately a change for the better.

Gimme multiplayer!

[schwatoo]

To hell with adapative AI. Gimme multiplayer Spinter Cell! When I got Spinter Cell on the xbox last year I was already hooked on Ghost Recon on xbox live. I was upset when the only live feature of Spinter Cell was content download. From the preview it looks like Spinter Cell 2 is going to have a very very sweet multiplayer mode!

Couple Examples

[Deliveranc3]

There was a boss in soul caliber where my friend said if you let him take 1/3 of your life you could beat him easily. There was also a dm game Q1 I think where you gave the cpu a 19-0 advantage then you could just tear it apart. Both examples are kind of like easter eggs for noobs. Also you can actually tell on the adaptive AI in Max Payne 2 on the Dead Man Walking Levels.

Re:No.

[Pofy]

Isn't that controlled by difficulty level you chose? After all, no matter how good or bad I am at a game, at times I want to play an easy game, at times a hard game. If the game simply adappt to me, and sets itself based on how I play, that is not possible. In my opinion, most games should have much more settings that you should be able to change individually, not just a easy/medium/hard option.

Re:No.

[Anonymous Coward]

Would you like to play the game that is getting impossible to win if you are too good?

Yes, Mazer, I would. It still won't beat me.

Re:No.

[bugbread]

Or if you're the kind of person who enjoys the feeling of satisfaction that occurs as your increases in skill make previously difficult things easy.

I suppose there are some people who like that feeling of "banging your head against the wall", where no matter how much better you get, everything is just as hard. Personally, I like the feeling of getting better at something and being rewarded for my skills, not penalized.

Re:No.

[bugbread]

Ok, apologies, that came off as more snarky than I had intended. I just meant to say that there are other reasons to enjoy getting better at a game besides showing off.

NFL Blitz anyone?

[Hwaguy]

This title has been known for its AI to adapt a "handicap" whenever it is getting stomped. Say, for example, that you are three touchdowns up against some poor computer player. The computer player will suddenly make miracle interceptions and unbelivable tackles that cause you to fumble and lose the ball. They will also dodge any and all of your tackles!

This is not really adaptive AI, but rather, a tweaking of the stat system (instead of an fumble every 40 tackles, its a fumble every 5), but it seems that no truely adaptive AI exists yet, at least in any game that would be worth having an AI adapt to.

Re:NFL Blitz anyone?

[WebGangsta]

I can confirm this for the arcade version of NFL Blitz '99. Another item: if you're a registered user and you're trying to beat all 30 teams, as you get closer to winning all the games you'll suddenly find yourself without being able to enter powerups (that is, the machine will automatically go into 'tournament mode' without actually being told to go into that mode). Blitz doesn't actually learn how to play against you (as a true AI would, for example, notice that the quarterback always fades back and to the left before passing), it just cranks up the speed of the computerized players and their "random" factors.

exactly my point

[*weasel]

things like increasing the assistance from 'auto-aim' are ideal for making adjustments in difficulty.

and particularly, aside from whether it's the 'right' way to adjust difficulty, wasting time coding and testing varying levels of enemy behaviors that most people will never see is just plain waste.

Counter-Strike

[nicolas.e]

This game has a really good AI, and speech synthesis is well done.

Back to the 80's

[Anonymous Coward]

Zanac on the SNES had superb adapative "AI".

Great game

Its called catchup code

[Radius9]

That isn't adaptive AI, as the players in the game play the same. It's called catchup code. Its used often in racing games as well. Basically, in NBA Jam, if you are behind, your shooting percentage goes up, regardless of the player. Same with the AI. This has the drawback of the easiest way to win at NBA Jam in single player is to just make sure you are never in the lead for most of the game. In the 4th quarter, if you've let the AI maintain a lead for most of the game, you can start doing stuff like making full court 3 point shots, as well as always being able to succesfully steal the ball. Quite irritating in my view. Racing games tend to do this quite a bit too, Need for Speed Underground being a good recent example. No matter how far ahead you get, the other cars tend to always be right behind you, and vice versa, if you crash, you can almost always catch up. This just turns the race into something where you can crash as much as you want, as long as its not on the last lap or two. Getting ahead does nothing for you, as the AI is always right behind you. Its an old trick from the coin op days, where the games tend to last 2 minutes or so, so it makes it more exciting. But I find it quite irritating on home games.

Lemonade?

[Anonymous Coward]

Since when is 7up/Sprite considered lemonade?

it only adds more bugs

[sven_eee]

some of the games i've played with "adaptive AI" work of hit points/damage so if you only have a little bit of health the AI takes it easy on you, in some game if you only have 1% health the AI ignores you, so you can walk strate pass the bots without them seeing/targeting you.but that is more adaptive SKILL not adaptive AI.
Adaptive AI lets the AI adapt to your style of play, so if you camp it anticamps, real adaptive AI trys to out think/out plan your moves. so if you like to sneak up on the AI it will start to expect that and then act agenst that (claymores :P).
I've try making good ai for game but the biggest problem is performance, alot of "thinking" ai code takes a lot from the CPU and slows the game down where most gamers want good GFX and heads to pop in good gfx. what is why games like UT / Q3 still use waypoints and paths, it cuts back on what the AI needs to think about leaving more CPU free for GFX and other things

Street Fighter 2

[jonathan_the_ninja]

A very far cry indeed. Honestly, if you beat the first 3 characters you fight in direct succession, the next person you fight will be landing straight specials! It seems to me that Street Fighter 2's so-called "AI" is simply adding more special moves to the enemy's barrage of attacks, in addition to cheat combos, and making your own moves harder to land, and, perhaps even increasing the speed of the enemy to an extent. Don't get me wrong, I love Street Fighter 2...