Most Impressive Game AI?

togelius asks: "I have the feeling that when developers make the effort to put really sophisticated AI into a game, gamers frequently just don't notice (see e.g. Forza). Conversely, games that are lauded for their fantastic AI are sometimes based on very simple algorithms (e.g. Halo 1). For someone who wants to apply AI to games, it is very interesting to know what AI is really appreciated. What is the most impressive game AI you have come across? Have you ever encountered a situation where it really felt like the computer-controlled opponents were really thinking?"

Comments

Come off as cheap

[Romancer]

A lot of AI that is used in games today can come off as cheap since the computer can think and compare much faster than a human player. Imagine fighting an opponent that can react 10X faster than you.

Another way to look at it is if you think that the AI is learning patterns and adjusting for tactics.
That's been played out in many genres, the most recent to come to mind is the Stargate SG1 episode where a character must face a situation that adapts to his efforts and becomes impossible to beat since the game can react faster than he can and has a perfect memory.

It's a ballance that game AI must match, playability and difficulty.

Re:Come off as cheap

[Romancer]

Exactly.

The most human AI would see patterns and adapt. The learning also plays a role, since just randomly playing out pre-programmed moves till one works leads to repetitive gameplay. The AI must have a very low level of options to piece together so it can make larger combinations that turn into tactics. The smaller each action is and the more actions it has to work with will let it find the best action. But that still doesn't mean that it has learned anything if it starts over with each situation. It has to have some sort of loose pattern recognition to see similar situations and apply the most likely solution.

Add all these things to an ability to totally screw up and you'd have a good AI. :)

Re:Come off as cheap

[2short]

Targeting by calculating from internal data and perfect knowledge of the physics model, rather than analyzing screen output, is not what I would typically call "AI". Not all games are First Person Shooters, and FPS are not particularly good tests of AI since targeting is such a big part of them.

In simple, pure-strategy board games like Chess, the best AIs are only on par with the best humans. In more complex mostly-strategy games, like the various Real Time Strategy games, the AIs are hopelessly outmatched, and the difficulty of player-vs-computer scenarios is typically adjusted by how huge a head start the computer is given, or how much it is allowed to cheat in various ways.

supreme commander

[Anonymous Coward]

supreme commander's ai was the first rts ai that i was not able to beat using the "survive the rush, build large army, overrun" recipe.

Fact or fable?

[lawpoop]

When I was contemplating learning video game programming, I was reading a guide that told you first to program a pong clone, and then a pac-man clone. Why pac-man? It teaches you AI. The ghost behavior is actually fairly complex. One ghost wanders randomly, another tries to get on the opposite side of the board from wherever pac-man is. The other two form a hunting pair: one tries to cut off your escape while the other goes for the kill.

I never thought that the ghosts would be so complex!

Galactic Civilizations 2

[Zarhan]

Well, one of the greatest experiences (And still is), AI wise, is Stardocks XXXX-type space strategy game, Galactic Civilizations 2. I especially like, when on easier levels, you do something, and the AI race sends a message "It seems that you are making a massive buildup for war. However, with this difficulty level, I pretend not no notice it until you actually make your strike." or something to that effect.

Simplfy the game and the AI gets better

[kinglink]

The Best AIs I've found are in some games like Yu-gi-oh the card game I'm currently working on nightmare troubadour and most of the opponents I've played always make a "great" move. On the other hand there's a couple opponents who are dumb as bricks, yet these enemies are suppose to be dumb as bricks (they are first time players in the story) And it's amazing how poorly they play (the play "well" for stupid AI, but they make bonehead moves that a new player can easily capitialize on. The player feels like each player has a different style, not just a different deck, and that makes for a much better game. (this is coming from a 25 year old guy in the game business).

The reason it's great is that there's simple rules to the game that the AI can know. There's been one point in the game where the AI got confused mainly because I blocked her in with a couple traps, but overall the Ai's abilities in the game are outstanding.

The important think to know about AI in games is it's not "AI". It's scripts or code that simulates scripts. There's no neural nets or anything else because we can't get the power for a neural net in an active game. In chess we can but then chess no longer is fun unless we tone down the "intellegence".

Some other great AIs are Gears of War (On insane they do great flanking maneuvers and such) Ghost recon (they really seem to know how to take cover and make it a challenge for the player to take them out. however the friendly AI leaves.... alot to be desired), Oblivion (watching random people walk around is pretty impressive, it helped build up that game.) and others, but there's none that make me think I'm fighting a real person.

There is a push to create truer "AIs" in games, Gran turismo created a way to train Drivers, Forza 2 is improving on it's drivtar system, Virtua fighter 4 had a way to teach an AI fighter, which was cool and indepth. But these are all "Scripts" taken from player experiences, not exactly AI. There's other games working on "true AI" but even then it's still toned down because we don't have the tools to make the driver "think" yet. It's just rail following and teaching the computer how to follow rails or when to break away from them.

I wouldn't say the molyeniux's games had great AI but they have good AI that at least learns a bit. Yet they feel like it's all you telling the game what to do, and it trying to figure out what you want it to do (and it fails) where as the Sims has interesting AI, but never feels real (mainly because the game never feels real).

So overall if you want to see good AI, look at simple games, expecting full 3d world simulations to have great ai is still a long way off but it's slowly coming. However this push for "graphics graphics graphics" won't help AI in the long run, but hopefully in a couple generations we stop worrying about graphics and work on AI and physics which seem to be more beneficial to the player then higher polycounts.

Re:Simplfy the game and the AI gets better

[tmortn]

Eh, the chess programs are mostly just pattern matching against libraries of stored games and brute force projection of all possible moves from a given point. They rank the options and the top option always wins. I suppose when considering that the computer is storing your games (ie growing its library) then its responses could change over time. Not because it learned, but because the same 'intelligence' is applied to a different data set. Given the same data set the ranking system will always react the same way in the same circumstances. A very rudimentary form of 'learning'. Really it isn't learning any more than say the Pythagorean theorem which also gives different results based on different inputs despite the fact its method does not change.

I can't say I buy your neural net deal either. At least not as a distinction between 'real' and 'non-real'. Will certainly agree in your examples there is a distinct split between computationally expensive and computationally cheap ways of determining courses of action for a computer program. But no "real" AI exists yet, that is no Turing test capable AI... neural nets or no neural nets. I think the point the other response was trying to point out is that regardless of the method used to simulate intelligence (static script vs adaptive code) the important factor in determining if something is actually artificially intelligent is generally agreed to be the Turing test. That is in interaction with humans it is impossible for a human to distinguish if the responses of the machine are from a human or not.

Frankly, 'real' AI from something like neural net code is not something game companies desire at all (computationally expensive or not). The problem is such code is by its nature un-predictable because any such system of learning/mutation has to be based on the unpredictable input of the player. A game which is unpredictable is bad from an investment standpoint because you don't know what it will do or how it will respond. Thus deterministic scripts that can be relied upon to act in a consistent way beneficial to the game are generally far more desirable. Not to say I don't want to see it come to be... just that the likely hood of seeing serious work along these lines is pretty slim right now barring some kind of breakthrough.

Warcraft III (Insane!)

[nartz]

The best AI I have seen is probably in WC3. However, I feel that in many games, it isn't the AI that is good, but rather that the computer players sort of cheat by having knowledge of everything in their environment; for example, they know (from the beginning) where the bases of other players are, instead of having to search like a human player. This gives them a huge edge - think of it as a human player playing against another with a map-hack, very unfair.

Wesnoth

[Glowing Fish]

I would have to say that the AI for Wesnoth, an open-source Turn Based Strategy, is one of the better AIs I have encountered (for that genre of games).
Although it isn't that the AI is that well done, it is that the rule set is simple enough that an AI can follow it.
I've played Civilization, Heroes of Might and Magic and Masters of Orion, the trinity of TBS games. Although they were often very good, the AI could only win in all of them due to "cheating" of a sort. The reason was that the various different factors to be considered were behind the planning ability of an AI. For example, in Heroes of Might and Magic II, there were seven different resources that a player could collect. Often, towards the end of the game, even while it was badly losing, the AI would be running around trying to grab resources, and would lose because of it. In Civilization II, because there was so many different units and improvements to be built, the AI would produce useless units, or spend all their time building improvements to cities that were about to be captured. The algorithm for keeping track of so many factors is impossible to make in an AI. AIs can't understand what is relevant and what is not.

So, in Wesnoth, there is only one resource to be considered, gold. Damage is also a straightforward mathematical calculation. So with the simpler rule set, the AI can play in a relevant way. Not that the rule set is simple in the sense of easy, it has a few factors, but those few factors can be combined in intricate ways.

So Wesnoth has one of the better AIs in my view, although of course it can still be tricked and worked around, but then any AI can be.

Cheating

[dunezone]

Ive noticed that AI is not designed to beat your next move but is designed to cheat you without the player noticing. Command and Conquer and Gears of War are two games that have two well hidden cheats. Command and Conquer is twelve years old almost, the enemy AI was programmed to always have full resources as long as one harvester made it back. Therefor what would take you five harvesters would only require them one. Most players would of never noticed this unless their strategy was to cut off enemy resources instead of an full out assault. Gears of War was praised for having AI that used the environment to their advantage which helped cover a little cheat they had. The AI had a weird tendency to know exactly where you were as long as your cross hair covered them or came close to them. For example if you were to pop your head out and just happen to have your cross hair on an enemy turret that was always firing at a covered friend, it would immediately start firing at you, this would also goes for the regular grunts/guards. This is very noticeable on "Insane", since that mode requires you to use cover 90% of the time and better tactics then rush in and shoot everything that moves. AI is not designed to outsmart/out think/or consider your next move, in my opinion most AI is designed to defeat you by using small cheats in the programming that give it an unfair advantage and hopefully designed so that you wont be able to notice it.

Game where computer seems like it is thinking

[Brian_Ellenberger]

Galactic Civilization I and II (see: http://www.galciv2.com/) is one of the few games I have ever played where it seemed like the computer was thinking. If you have never played GalCiv, and you like strategy, I highly recommend picking them up. I consider them to be superior even to the Civilization series. Brad Wardell prides himself on the AI, and it definitely shows. The computer is very difficult to beat and does not cheat. It actually responds in a logical manner, which makes GalCiv go from just being a number-crunching exercise to an actual strategy game. For example, when making some "aggressive" moves towards an enemy (moving some attack ships to an "ally" to wipe them out) I've actually had the game pop up a message from my ally (before ever entering his space) saying something to the effect of "I used to play video games when I was a kid, and when I did I used to build my forces up and send them to sneak attack an opponent. Well I am no video game." Other things like the fact that if another civilization is dependent on you for a large amount of trade income, they won't just randomly attack you because it would hurt them too.

Re:Game where computer seems like it is thinking

[Kingrames]

90% of the time when it seems like the computer is smart, it's more because the game and the AI merge well together.

If the game has lots of bugs, bugs you might not normally see, the AI will suffer.

Oftentimes the A.I. of the game doesn't make the computer smarter. Making a game that's smooth and supports a good A.I., however, will make a much bigger difference.

With a game like chess, the A.I. program is huge and immensely sophisticated.
With a game like tic-tac-toe, you can make an A.I. that can't be beaten, simply because the game is simple and allows for that.

It's important to keep in mind that the actual A.I. algorithm can only accomplish so much. Putting Deep Blue into the seat of your tic-tac-toe opponent gives you the same result as the program you wrote that doesn't break a page.

That being said, a few more examples to look for for good A.I. that merges well with its game would be Kohan and Kohan 2. The AI in that game blew me away.

For an older game, check out emperor of the fading suns (you can get the full game for free) http://free-game-downloads.mosw.com/abandonware/pc /strategy_games/games_e_f/emperor_of_the_fading_su ns.html
I still have fun playing this one. It's interesting how the computer will actually send you money for nothing in the interest of making you like them more. There are a few other subtle details.

Incidentally I've heard that Gal Civ is based off of EFS, so if you haven't had a chance to check it out, enjoy.

Game AI? Why bother! True AI has been solved.

[Anonymous Coward]

So-called "Game AI" is not even AI. It is just s few fancy tricks to make the game program look, but not actually be, a little bit smarter.

Mind.html recently became a True AI that reveals the deep thought process in a tutorial display mode. You can interact with the AI Mind and watch it thinking, as spikes of excitation spread by associative tag from concept to concept in the knowledge base of the genuine artificial intelligence.

Mind.Forth AI for robots is written in Win32Forth for installation in autonomous mobile robots and has spawned at least one independent offshoot on the Web as the true AI evolves and speciates into multiple branches of live-or-die AI in the Darwinian jungle of survival of the fittest.

Franks AI Mind is the "son-of-Mind.Forth" AI with advanced features such as the ability to send e-mail and to read Web pages.

Civilization IV

[Kuciwalker]

With Blake's http://apolyton.net/forums/showthread.php?s=&threa did=159157A Better AI which Firaxis actually included in the latest patch, it's gotten pretty impressive.

Perfect Dark N64

[oliphaunt]

OK, it's pretty old by now, but I was a big fan of Perfect Dark. The AI robots had difficulty settings AND personalities. If you were running deathmatch games against the AI, you could set AI's to have various behavior attributes:

• PeaceSim:As the name implies, this Sim hates violence. In fact, the PeaceSim will go around hoarding weapons so people don't pick them up, and disarm people for their weapons. Therefore, they'll drop a payload of weapons when you kill them. Just don't let them sneak up on you...
• ShieldSim:Like some human players I know, this Sim is a shield addict. It will always go for the shield, even if it has no weapons! In fact, if you damage its shield in the least bit, it will retreat to get another shield! My advice is don't let it.
• RocketSim:This is the pyromaniac of the Simulants! The RocketSim will always pursue the explosive weapons, and will set them off, even if doing so would spell death for itself! Avoid this Sim, or kill it before it can get an explosive.
• KazeSim:This is fearless, suicidal menace. It will make suicidal runs, even with no weapons, to try and destroy you. It fears nothing, and that makes it a dangerous enemy.
• FistSim: Unlike the PeaceSim, the FistSim is violent. Like the PeaceSim, though, it will hoard weapons and try to engage you in hand-to-hand combat. It won't use weapons, but it will do good damage with its hands.
• PreySim: This Sim truly feels that honor is a minor detail in a fight to the death. The PreySim dislikes competition, so it will hunt down the easiest targets to gain an easy kill. Its favorite targets include weakened opponents that are unarmed or armed with a weak weapon, and enemies that have just spawned. The PreySim also loves to cloak, so beware.
• CowardSim: This is the SimWussy. It flees to safety at the mere sign of confrontation, and will only confront you if it has a superior weapon. Carry a big gun, and you will rarely meet this Sim. Hide out and try to catch the coward off its guard.
• FeudSim: Stay out of this Sim's way! If the FeudSim goes after you, it will hunt you until the end of the game! It will mercilessly hunt its target, even if you kill it.
• SpeedSim: As the name suggests, this Sim is extremely fast. It's definitely faster than you, so it's difficult to hit with standard weaponry. It's impossible to flee, so stand and fight like a man.
• TurtleSim: This Sim is the opposite of the SpeedSim. It moves at a much slower rate than most players, but it has a shield that is twice as strong as the standard shield! Fortunately for you, it has restricted mobility due to its shield.
• VengeSim: This is a psychopathic Sim! This Sim will completely ignore other players just to attack the player that last killed it! It attacks with a vicious rage; so to avoid its rage, just leave it alone.
• JudgeSim: This is the only decent Sim. The JudgeSim acts like the judge of the battlefield, going after the winning player to even out the odds. That means if you are an expert playing against some young rookies, expect this Sim to come after you!

The variety of personalities gave the game infinite multiplayer replay value, and made it easer for beginners to get into the game. You could pick simulants that would ingore a newbie human player and attack only the players with more kills, so the good players can run around slaughtering AI's on the difficult setting in the same game that a newbie is just exploring the level and figuring out how to reload. The experts still have fun while the newbies don't get instantly killed every time they spawn.

Falcon 3

[Ullteppe]

Anybody remember Falcon 3? While all the other sims pretty much had scripted missions (many still do), Falcon tried to run the whole war in the background. And when you ran into other planes, they acted pretty convincingly. They were hard to beat as well, I remember the Mirage F1s especially as being pretty tough in a dogfight. Pilots of different planes acted differently according to their planes strength/weaknesses. They used 6 months to patch the game sufficiently that it didn't crash all the time, kind of understandable with the complexity.

Re:S.T.A.L.K.E.R.

[Anonymous Coward]

Ive played that mission too, they dont come avenging, they follow you back to the village, but thats still damn impressive. One of the greatest things about this game is watching the AI work without your interference. Watching the Bandits shooting at STALKERS, then watching the animals decide they want in on this fight, or watching dogs fight over a mutilated corpse.

Even better is when youre wandering through a deserted trainyard late at night, only to realise that youre being hunted by a mutant on the other side of one of the platforms, creepy.

Re:Kart Racing

[Cecil]

It's not a star, not the same way you get a star from a question block. It's an invincibility special ability. It's not quite the same as a regular star as it doesn't make you go faster, improve your handling, or allow you to drive with reduced penalty on grass/mud. It's just invincibilty + collision damage.

All of the AI players have their own special abilities, which fall into two basic categories. Mario and Luigi get invincibility, everyone else gets a tossable/droppable item (Banana, Fireball, Shell, etc) which they can use repeatedly.

It's not the same kind of cheating when a car mysteriously catches up to you at seemingly supersonic speed. In the former case, it's an obvious game mechanic. In the latter case, they're trying to be subtle about it and use it to cover up AI weakness, hopefully without you realizing that they're doing it.

Re:Perceived Intelligence - Simple is better?

[Eivind]

That's typical of the AI possible today: It can't actually perform better than a stupid program following simple rules.

Had a similar depressing experience in my class on AI. The task was to build a neural network that could guesstimate the sex of a first-name. A quite complex neural net, trained on 300 random male and female names could thereafter guess the correct sex of a name about 65% of the time.

Which seemed impressive until someone pointed out that a trivial table-lookup of the most common 100 female and male names, and random guesses for everything else is enough to reach about 70-80% (depending on the country the names are from, some countries have more variation than others) and even something as simple as "if ends in a, guess female, otherwise guess male" is enough to reach similar "accuracy" as the neural net.

Re:ummm, Galactic Civilizations II?

[ditoa]

I just went to the GalCiv2 website to have a read and saw this

No CD copy protection. Once you install, you never need your CD again. You can even use the included serial # to re-download the entire game from us years from now.

That is very refreshing to see these days. I have given up on most PC games these days because of their copy protection systems.

Re:Perceived Intelligence - Simple is better?

[NekoXP]

AI in games is mostly triggers and actions.

It's still AI if those triggers are hardcoded or based on statistical analysis.

It *is* intelligent for an NPC to realise a grenade was thrown by the player into it's general area, and throw it back at the player. That's not to say it shouldn't be ranked highly or that it is a crappy AI.

If an enemy hops over a wall next to him for cover it's the enemy realising that he needs cover and there is an easily hoppable wall. He could just as well duck behind a barrel, but the barrel may be made of soft shitty wood, the wall is nice and made of brick. It makes an education decision just like we do. They may well be SCRIPTED.. if near wall hop over wall else if near barrel duck under barrel else if player is actually in effective firing range, run the hell away else sit there and taunt them into the minefield..

Think of your thought process when you would be in the firing line and you have the choice between ducking behind a worm-riddled barrel to avoid gunfire, or a solid brick wall that you can hop over. How many seconds does it take? Can you remember every microdecision you made? No, you think "ohshit I'm being shot at" and probably duck behind the barrel under stress, when it starts to splinter then you scramble over the wall and realise you dropped your gun hopping over the wall..

Does that make it crappy AI that it did not break out scientific analysis of the situation and count variables or do complex physics?

The Thief Series was Groundbreaking

[BrendaEM]

The AI could see not only see in a view fustrum, but also by how much light you were lit by. The AI could hear depending on what surface you walked on, you could affect the light by putting out torches, affect the sound my mossing the floor. The AI notice bodies and things out of place, such as a climbing rope. The AI also had different alert states. I think that they are pretty good for a game made almost a decade ago.

There are over 400 Thief series fan missions. Last night, I played "Ominous Bequest"
http://southquarter.com/?p=131
One level change for 6 hours of gameplay! Yes, Thief 1 and 2 graphics are quite dated, and low-poly, but if you can forgive that, the gameplay and atmosphere compensates for it.

Lately, there's been a reinsurance in Thief 1 and 2 add-on missions. There's a updated version of 680mb Thief2X add on. Missions like "Ashen Age," and a "Night in Rocksburg" have breathed a little life into T2 visually.
http://forums.eidosgames.com/showthread.php?t=5526 3
http://southquarter.com/?p=131

Someone is working to refresh the Thief Series Engine, but the source code would help. I wish that Eidos would release the source code--for an eight year old game.
http://www.ttlg.com/forums/showthread.php?t=102289 &page=11

Re:Civilization III

[Taelron]

All of the Sid Meirs Civilization games cheat. To prove it, get one of the trainer and save game editors. Save the game each round and look at the the AI players citys and units. They will produce two units at once and instantly without spending money to "buy" them. The cities dont suffer ill effects of to many troops or to long of a war. The higher the level you set the game at the more the AI cheats.