AI in games

[Piemax2]

I thought this call-for-papers might be of interest given our recurrent discussions of Game AI. I'm also curious what the industry folk think of the idea of building games where the NPC's use machine-learning algorithms-is this a promising direction or hopelessly impractical for the time being?
---

International Journal of Computer Games Technology

Special Issue on

Artificial Intelligence for Computer Games

Call for Papers

Artificial intelligence (AI) in computer games covers the behaviour and decision-making process of game-playing opponents (also known as nonplayer character or NPC).Current generations of computer and video games offer an amazingly interesting testbed for AI research and new ideas. Such games combine rich and complex environments with expertly developed, stable, physics-based simulation. They are real-time and very dynamic, encouraging fast and intelligent decisions. Computer games are also often multiagents, making teamwork, competition, and NPC modelling key elements to success. In commercial games, such as action games, role-playing games, and strategy games, the behaviour of NPC is usually implemented as a variation of simple rule-based systems.ions, machine-learning techniques are hardly ever applied to state-of-the-art computer games. Machine-learning techniques may enable NPCs with the capability to improve their performance by learning from mistakes and successes, to
automatically adapt to the strengths and weaknesses of a player, or to learn from their opponents by imitating their tactics.

However, many more AI techniques have been developed in academic research, and relatively few have actually been transferred and are used in the development of commercial video games. In this special issue, we invite submission of papers from current research in AI proposing new approaches specific to computer game requirements including but not limited to:

o Knowledge-based systems and management in games
o Machine learning in games
o Adaptable agents in games
o Multiagent models in games
o Constraint-based model for games
o Tactical and real-time pathfinding
o Application cases to neural network for games
o Fuzzy logic for games
o Goal-oriented action planning in games
o AI controlled storytelling
o AI scripting
o Evolutionary computation in games
o AI and game design
o Teamwork, coordination, and communication
o Planning in highly dynamic environments
o Multiagent learning and adaptation
o Human-agent interaction
o Opponent modelling
o Believable and emotional agents in games

Authors should follow the International Journal of Computer Games Technology manuscript format described at http://www.hindawi.com/journals/ijcgt/. Prospective authors should submit an electronic copy of their complete manuscript through the journal Manuscript Tracking System at http://mts.hindawi.com/, according to the following timetable:

Manuscript Due: May 1, 2008
First Round of Reviews: August 1, 2008
Publication Date: November 1, 2008

Guest Editors:

o Abdennour El Rhalibi, School of Computing and Mathematical Sciences, Liverpool John Moores University, Byrom Street, Liverpool L3 3AF, UK; a.elrhalibi@ljmu.ac.uk

o Kok Wai Wong, School of Information Technology, Murdoch University, 90 South Street, Murdoch, WA 6150, Australia; k.wong@murdoch.edu.au

o Marc Price, BBC Research, Kingswood Warren, Tadworth, Surrey KT20 6NP, UK; marc.price@rd.bbc.co.uk

[Aeon221]

http://www.hindawi.com/journals/ijcgt/si/aicg.html link to the original source.

[Piemax2]

Sorry, I didn't have that link, this was sent to me by email. I didn't even realize that the call for papers had been posted, though I guess I could have realized that it probably had been. Did politeness require that I go looking for possible links before posting? And should you perhaps have PM'd me to see if I had seen the link before offering politeness lessons?

[Piemax2]

PS I'd rather not derail this into an etiquette thread so I hope someone posts on the substance.

[Pogo]

What does this mean in layman's terms? I'm assuming most games now use if-then statements in complex manners to achieve more complexity. What do algorithms do that cause machines to "learn" from their mistakes, and in that sense know what a "mistake" is?

[MarinusWA]

What does this mean in layman's terms? I'm assuming most games now use if-then statements in complex manners to achieve more complexity. What do algorithms do that cause machines to "learn" from their mistakes, and in that sense know what a "mistake" is?
Most simply put, you have an algorithm (heuristic) that determines the value of a (game)state and acts accordingly. In a static system this will always lead to the same action. State X -> Worth Y -> Action Z. In a learning/dynamic system past experiences will add an additional modifier on the calculation of that value. State X -> Worth Y + XP -> Action ???

I actually wrote an <A href="http://minor piece (http://ai-depot.com/GameAI/GameTree.html%22]minor" target=_blank>http://ai-depot.com/GameAI/GameTree.html) back in 2003 on the requested topic, I might be compelled to submit a new piece.

[cliffski]

What does this mean in layman's terms? I'm assuming most games now use if-then statements in complex manners to achieve more complexity. What do algorithms do that cause machines to "learn" from their mistakes, and in that sense know what a "mistake" is?

holy cow, I'm sure I'm not the only AI coder whose work isn't IF THEN any more.

I'm sure learning is practical, but I guess it depend how you are defining it. it's pretty easy for you to code entities to 'rate' the success of a randomly selected strategy against a specific opponent, compare it with other similar data and give extra weighting to the most successful one.
I guess in an FPS, an enemy could choose a weapon to use to attack the player when the player is in position X, and over time, learn that the grenade launcher is the most effective strategy when someone is behind a barrel (for example).
The games I work on are more about simulating influence and emotion, rather than this stuff, but even in Democracy, the voters learn not to trust manifesto commitments from a government that has previously failed to keep them.

I guess in simple terms, IF and THEN have been replaced by a more granular variable that measures the influence of one thing on another. In other words, such stuff is more analog than digital these days.

[Lorini]

Just sent the call on to my JPL AI colleagues, thanks Pie!

[Mike O'Malley]

I want to get in on this, but 2 weeks!??!

Crap.

Edit:
International Journal of Computer Games Technology is an Open Access journal. Publishing an article in International Journal of Computer Games Technology requires an Article Processing Charge of €400 which will be billed to the submitting author upon acceptance of the article for publication.

Forget it, I'm not paying $800 (or convincing my advisor to do so) for this.

[cliffski]

they CHARGE you for sharing what you have learned?
HAHAHAHAHAHAHAHA

[AndrewM]

they CHARGE you for sharing what you have learned?
HAHAHAHAHAHAHAHA

The problem is that there are costs for making an academic journal. Where does the money come from? Traditionally, a big publisher publishes the journal and retains the copyright, and charges people money to access it. More recently, some journals have started using a new model, wherein the authors pay some small sum, and then the journal is available for free to anybody who wants to read it. So, in exchange for paying some money, your work will be more widely available. And 400 Euros isn't that bad. At some conferences, they charge you $200 per extra page above the usual limit (up to 2 extra pages, for instance).

Here's a Wikpedia article (http://en.wikipedia.org/wiki/Open_access), or maybe this one (http://en.wikipedia.org/wiki/Open_access_publishing) is more immediately relevant.

[datter]

He thinks that is mighty funny.

[Lorini]

I think this is basically a call to academia/corporate, not a call to small game designers/companies. "In this special issue, we invite submission of papers from current research in AI proposing new approaches specific to computer game requirements including but not limited to:...."

[ElGuapo]

I'd love it if more game explained, instead of devoting pages to vague lists of weapon descriptions, how their AI worked. How it thought, what you could expect, if it learned, etc. For instance, I was interested to know in Rainbow Six: Vegas that different weapons have different sound radii when shooting, and the sound of dry firing a weapon is the same as shooting it silenced. Also, what determined when reinforcements are called in? What determines where the bots/NPCs/AI spawns? Explain some game mechanics and I'll appreciate your product so much more.

[AndrewM]

Or appreciate it less. I think games get a lot of mileage out of people not really knowing what is going on, because then players might attribute to intelligent design what was actually produced by random chance.

[Zylon]

I'd love it if more game explained, instead of devoting pages to vague lists of weapon descriptions, how their AI worked.
Why do you hate immersion?

[Miramon]

IMO most games don't really have AI. Typically, some pathing and a very simple decision algorithm. Games with fixed scenarios (like historical wargames) tend to have scenario-specific scripting to make the computer enemy look smarter, and games with random scenarios just have low enemy intelligence, or else the computer enemy has a huge force imbalance or just flat-out cheats to put up a respectable fight.

Certainly there is little in the game industry that is similar to what AAAI members usually talk about.

There are a few exceptions, of course, but you're not going to see much old-fashioned strong AI goals and tasks and search spaces, much in the way of neural networks or induction, essentially no machine learning, and very little else of academic interest in the vast majority of games, not even rule-based systems, for the most part.

[Mike O'Malley]

This is a HUGE problem.

1. Conflict of interest. How strict is their peer review process if they're gaining revenue from publication?

The problem is that there are costs for making an academic journal. Where does the money come from? Traditionally, a big publisher publishes the journal and retains the copyright, and charges people money to access it. More recently, some journals have started using a new model, wherein the authors pay some small sum, and then the journal is available for free to anybody who wants to read it. So, in exchange for paying some money, your work will be more widely available. And 400 Euros isn't that bad. At some conferences, they charge you $200 per extra page above the usual limit (up to 2 extra pages, for instance).


2. It's NOT FREE (http://www.hindawi.com/subs.html). It's $195/year. So they're making money on both ends. If they're charging for one- as asinine as this peer-reviewed vanity press idea is- they shouldn't charge for the other.

[Derek French]

int(rnd() * 4) FTW!

[AndrewM]

This is a HUGE problem.
1. Conflict of interest. How strict is their peer review process if they're gaining revenue from publication?

Sure, that's a problem, but if a journal is terrible, then who is going to pay $400 to have their stuff published in it? And open access journals are neither the beginning nor the end of places that take anything you submit to them, and consequently nobody cares if you get something published in them.

2. It's NOT FREE (http://www.hindawi.com/subs.html). It's $195/year. So they're making money on both ends. If they're charging for one- as asinine as this peer-reviewed vanity press idea is- they shouldn't charge for the other.

It costs money to get a print subscription. It looks like you can download the articles for free.

[Kunikos]

int(rnd() * 4) FTW!

Your random number generator sucks!

P.S. Yes, learning AI is completely possible for games but most game teams don't even have a dedicated AI person (or one that is academically trained) who knows how to even start such a thing. A decent engineer can throw together a simple state machine that can execute scripted sequences based on inputs from the game state, but will likely fail to recognize that you can use fuzzy-logic to reduce erratic behavior. A more competent one will likely research what they can despite not having formal education, and may be able to puzzle out with some reference material how to do a neural network or perhaps even a genetic algorithm.

[Miramon]

Your random number generator sucks!

P.S. Yes, learning AI is completely possible for games but most game teams don't even have a dedicated AI person (or one that is academically trained) who knows how to even start such a thing. A decent engineer can throw together a simple state machine that can execute scripted sequences based on inputs from the game state, but will likely fail to recognize that you can use fuzzy-logic to reduce erratic behavior. A more competent one will likely research what they can despite not having formal education, and may be able to puzzle out with some reference material how to do a neural network or perhaps even a genetic algorithm.

And a really competent one probably won't use any of those techniques directly because they are not going to be very effective for most games, though he or she might incorporate snippets or hints of them in more conventional modules.

You need very deep knowledge of these things to get them to work well (apart from fuzzy logic, which is trivial and also produces trivial results) and also a big investment of time and effort. If you don't know for example, how hard it is to get a neural network to do something as simple as a logical xor, you might waste a lot of time in development when you could have just done without it.

A lame analogy, but it comes to mind:

I once knew an artist slightly who made these incredibly detailed landscapes for bookcovers. Tiny detailed leaves all over the place. Beautiful. Looked like they took weeks to do them. Then I saw him actually doing the work while he was bored at a sf convention. He smeared out some yellow, green and brown, rolled up a sheet of plastic wrap and crumpled it, and smacked it down on the canvas. 30 seconds later, leaves everywhere.

The point is, that techniques like those mentioned are like painting in the leaves with a one-hair brush, and the results would be no better. For a game, you need quick and dirty results that look good, just like for a movie, it doesn't matter if the sets have no back side, you just shoot the front side.

So I would say it's great for the engineers and designers to understand AI techniques, but not so great to actually apply them according to the usual academic or even industrial approach. Of course there are exceptions, but it's hard for me to imagine all that many cases where a neural network or an machine learning algorithm or a sophisticated task space search system would really be a good idea.

[Chris Nahr]

And a really competent one probably won't use any of those techniques directly because they are not going to be very effective for most games, though he or she might incorporate snippets or hints of them in more conventional modules.

Precisely. First there's the requirement of getting the AI done in a game development cycle of 1-2 years. Then there's the requirement that the AI needs to respond fast or in real time, even on slow systems that barely manage to generate the (far more important) graphics. Then there's the problem of changing rules.

The rules and statistics for a computer game can change at any time up to release, and at any time after release in patches or user mods. And the computer player is required to immediately offer a strong opponent within the new system. So any AI methods that require any amount of learning or training at all are unusable, except maybe for optional refinement.

But such optional refinements don't give a good ROI when the average customer plays only a few matches against the AI, or perhaps only plays a campaign where the situation and opponents are fully scripted anyway. Their implementation would have to be nearly free to make them worth the bother.

And then there's the lack of demonstrable benefits. If those academics who like to berate working professionals about their "simple" algorithms lacking "real" AI want to be taken seriously, the burden is on them to produce such a "real" AI for a real game whose complexity and variability is comparable to commercial games. Not just some simple classic board game or card game, or a nearly empty 3D space with a few simple objects, but a realistically complex environment. Implement that game, then implement your AI, then show me how long it took, how well the AI handles changes, and especially how much it improves the player's experience.

Once you've done that... then maybe I'm actually interested in what you have to say. Until then STFU and go away. I'm sick of pretentious fools with no field experience who think game programmers are stupid and neural networks are a silver bullet.

[Naeblis]

Or appreciate it less. I think games get a lot of mileage out of people not really knowing what is going on, because then players might attribute to intelligent design what was actually produced by random chance.

This. Lots of people would be surprised the few factors affects really the AI.
Or how sometimes when it appears the AI flanked you, it didn't really, it just there are only two paths to go from their location to yours, and so 50% of the enemies chose path 1 and the other 50% chose path 2 which results in the "flanking manouver".

[marxeil]

What is the feasability of an AI engine (ala HAVOC)? How easy / hard would it be to adapt something like that for each game?

[Charles]

Your random number generator sucks!

P.S. Yes, learning AI is completely possible for games but most game teams don't even have a dedicated AI person (or one that is academically trained) who knows how to even start such a thing.

That's not true at all anymore. It used to be true 5-10 years ago. But almost everyone has multiple dedicated AI programmers nowadays. It'll be a rare company who doesn't (and they'll probably make crap).

Maybe japanese devs can still get away with it. But traditionally they have some of the dumbest enemies around.

[Charles]

What is the feasability of an AI engine (ala HAVOC)? How easy / hard would it be to adapt something like that for each game?

All AI is extremely game specific, so doing a generalized AI engine isn't likely to ever happen.

[MarinusWA]

Multiple dedicated AI programmers? So far I can count only a handful of companies where I can actually notice this. Stardock & Associates for example. AI is not just about the tech it's also about how you use it. Any decent coder can write an optimised A* search algorithm but if your heuristic is ass your AI is going to be ass. And that second part tends to be the problem.

All AI is extremely game specific, so doing a generalized AI engine isn't likely to ever happen.
I don't see why not. Decision making is State -> Process -> Action -> State -> .... Seems pretty straightforward to generalize (probably massive overhead though).

[skedastic]

Lots of academic journals have small (usually about $50) fees to submit a paper for consideration. This is to discourage frivolous submissions, not (primarily) to raise revenue. Charging $800 is ridiculous: that's not an academic outlet, it's a vanity press.

[Miramon]

That's not true at all anymore. It used to be true 5-10 years ago. But almost everyone has multiple dedicated AI programmers nowadays. It'll be a rare company who doesn't (and they'll probably make crap).

Maybe japanese devs can still get away with it. But traditionally they have some of the dumbest enemies around.

Turbine fired their only "AI" programmer, and he had no AI background at all, he just wrote a mediocre pathing system -- so now they use a commercial package instead. Monster behaviors are a hack, basically, and that's it for AI in a MMO, anyway.

[Charles]

Multiple dedicated AI programmers? So far I can count only a handful of companies where I can actually notice this. Stardock & Associates for example. AI is not just about the tech it's also about how you use it. Any decent coder can write an optimised A* search algorithm but if your heuristic is ass your AI is going to be ass. And that second part tends to be the problem.


If you think AI is still just simple pathfinding, you don't know what you are talking about.


I don't see why not. Decision making is State -> Process -> Action -> State -> .... Seems pretty straightforward to generalize (probably massive overhead though).

Okay. So you simplify it down to that and then.... the vast majority of the implementation is still on the main CPU. What have you gained?

[Charles]

Turbine fired their only "AI" programmer, and he had no AI background at all, he just wrote a mediocre pathing system -- so now they use a commercial package instead. Monster behaviors are a hack, basically, and that's it for AI in a MMO, anyway.

MMOs are a completely different beast from single player games.

[cliffski]

I don't see why not. Decision making is State -> Process -> Action -> State -> .... Seems pretty straightforward to generalize (probably massive overhead though).

The overhead is the problem. especially for real time games. Any game that isn't a by the numbers copy of an existing title is going to need some creative thinking in coding it's AI. there is almost certainly nothing in the AI code for a quake bot you could reuse in a game like civilisation. Ditto the Sims, or Total War.
People have very high expectations for computer AI, and to get the results you want you need to cut a lot of corners, do some clever hacks and a lot of smoke and mirrors. In some games, the ai entities emotional state is everything (the sims, kudos...) in others, it's irrelevant. the same goes for pathfinding. people seem to equate AI with pathfinding or controlling attacks in an RTS, but it's way mroe than that. Selection of text used by NPCs is part of AI. Making the movements of background characters in a 3D RPG seem lifelike is also AI.

When computers are so fast that CPU time is basically meaningless for a real time game, maybe some VERY generic AI system will help ease the burden, but until then, a lot of game AI will be coded from scratch, and there will be a lot of scripting.

[MarinusWA]

If you think AI is still just simple pathfinding, you don't know what you are talking about.
I was trying to make the opposite point actually, there is far more to AI then just the tech aspect.


Okay. So you simplify it down to that and then.... the vast majority of the implementation is still on the main CPU. What have you gained?
With performance? Not much. Mind you, I'm personally of the opinion that before we are going to see any decent AI in games it's going to need hardware acceleration of some sort. But of course that ain't going to happen if there are no games that need it. Circle complete!

there is almost certainly nothing in the AI code for a quake bot you could reuse in a game like civilisation. Ditto the Sims, or Total War. With regards to State Search or rule based systems this is correct. But honestly don't even see those as real AI (especially not rule based). Any generic solution should have some kind of neural network core or possibly a genetic algorithm.

[Charles]

With performance? Not much. Mind you, I'm personally of the opinion that before we are going to see any decent AI in games it's going to need hardware acceleration of some sort. But of course that ain't going to happen if there are no games that need it. Circle complete!

But that's the point. You can't have hardware acceleration because it's not generic enough. Hardware acceleration for AI is going to mean "this code is properly multithreaded and the computer has many cores."

[beloved one]

Selection of text used by NPCs is part of AI. Making the movements of background characters in a 3D RPG seem lifelike is also AI.

You really feel like that qualifies as AI? NPCs are all just bots reading off of scripts. ACT I, morning: "Boy, this is one glorious morning." ACT II rain, "Wow, it's really coming down." ACT III nighttime: "Why don't you get some sleep?", ACT IV tragedy: "You KILLED her, oh my god!" Audience choice scene 1: "You insulted me!" or scene 2, "Let's go back to my room."

That sort of behavior strikes me as the domain of a playwright and not that of a creator of artificial intelligence.

[MarinusWA]

That sort of behavior strikes me as the domain of a playwright and not that of a creator of artificial intelligence.
It's not. Well technically it's weak AI but as far as I'm concerned stuff like that like most rulebased/scripting is intelligence by proxy. With the coder being the brain.

[Charles]

I just realized I misunderstood talk of an AI Engine earlier to mean hardware acceleration. Ignore that part. Carry on!

[ravenight]

The overhead is the problem. especially for real time games. Any game that isn't a by the numbers copy of an existing title is going to need some creative thinking in coding it's AI. there is almost certainly nothing in the AI code for a quake bot you could reuse in a game like civilisation. Ditto the Sims, or Total War.

Of course you aren't going to use the code for a quake bot to program Civ IV AI, but that doesn't mean you don't gain something by codifying AI routines. For starters, you can then make any quake bots you need with your simple quake bot AI. The same goes for hardware acceleration - there are certain things AI needs to do that are at least as generic as shader pipelines, T&L and texture memory, and a card designed to do that stuff quickly would alleviate some of the CPU burden of complex AI, making it possible to do more. It's not like graphics cards eliminated the need for graphics programmers - of course you'll still someone to adapt to the specifics of the game your making. Even if you use packaged products you'll need someone who learns what they do and how to integrate them. That person just doesn't have to be capable of writing a good state machine or goal system from scratch.

[Miramon]

But that's the point. You can't have hardware acceleration because it's not generic enough. Hardware acceleration for AI is going to mean "this code is properly multithreaded and the computer has many cores."

Agreed.

Hardware acceleration makes no sense at all to me. If AI was generally useful, I'd say that your reason made a lot of sense, but I find it hard to believe there is much AI at all in games either way. I've certainly seen no evidence of it myself in the vast majority of games.

AI doesn't actually take that many cycles or that much memory compared to something computationally hard like rendering hi-res graphics or even just blasting vast amounts of video data from disk to memory to card.

The problem is just that IMO AI techniques aren't actually that useful compared to conventional algorithms for most applications, and they're way harder to use. Partly this is because they are intended for different purposes, partly because they just aren't that good inherently except for some particular special applications.

Water flows downhill. The most cost-effective techniques will generally be used for any given application. If neural nets made any slight sense for say TBS strategy games, or FPS action games or whatever, they would certainly be used consistently. But I claim they don't make sense. Nor does simulated annealing, genetic algorithms, induction, knowledge-based systems, boolean logic systems, search spaces, task engines, blackboard systems, fuzzy this or fuzzy that, or almost all of the other various things AI researchers have come up with over the years.

There's a reason that AI just up and died as an application field back in the early 90s. It just turned out not to have that much value to industry. Try searching for "AI programmer" on monster.com, and see what you get in terms of listed positions. Last time I checked, in a 20 mile radius from Boston, there was a total of three positions at one company, a government contractor with a special purpose application suite.

This isn't to say that there aren't some select good AI applications, nor that AI as a field is unworthy and shouldn't be investigated further, just that it's not a panacea or even a generally useful area for conventional development at its current level of usefulness.

I say this as a no longer embittered former AI application developer, by the way, for what little it's worth.

[Charles]

The main hardware acceleration benefit for AI would be raycasting for collisions. That's one of the main time eaters of AI right now. That's the only truly generic part of AI, besides pure pathfinding, but pathfinding itself tends to be very game specific.

[ravenight]

I think it's more that no one wants to try to push the edge of it, because it is a very risky edge. Adaptive systems could make for some really interesting games, but you have to be willing to float a lot of money to try to make something work with no guarantee you'll end up with a playable game.

Look at The Force Unleashed, for example. They are using a package AI technology to make the characters in their game respond in a lifelike manner to stimuli. Do they need to do that? Not really, they could just as easily have released Jedi Knight III with only minor graphical improvements and probably made plenty of money. But once they have a handle on integrating DMM, Havoc and Euphoria, they may be able to make a whole series of games based off the same engine, and the things it does that no other game does will be a reason for some to buy it that wouldn't have.

The same is true of Spore. Is there any reason to use procedural animations? They gain a tiny bit in terms of making it easier to create an editor and distribute custom creatures, etc, but they could have made up for that by just hiring more artists to crank out content and making players suck up longer download times that they are already completely used to.

So, basically, the argument "I don't see any need for real AI in games" is short-sighted and meaningless as a counterpoint to the statement "games should be made to utilize AI better to do cool things they can't currently do".

[Charles]

I think it's more that no one wants to try to push the edge of it, because it is a very risky edge. Adaptive systems could make for some really interesting games, but you have to be willing to float a lot of money to try to make something work with no guarantee you'll end up with a playable game.


In fact, I'd say that it's almost a guarantee that you end up without a playable game. AI isn't what you want in games, after all. You want AS -- Artificial Stupidity. If the AI was actually as intelligent as everyone wanted, games would be absurdly difficult.


Look at The Force Unleashed, for example. They are using a package AI technology to make the characters in their game respond in a lifelike manner to stimuli. Do they need to do that? Not really, they could just as easily have released Jedi Knight III with only minor graphical improvements and probably made plenty of money. But once they have a handle on integrating DMM, Havoc and Euphoria, they may be able to make a whole series of games based off the same engine, and the things it does that no other game does will be a reason for some to buy it that wouldn't have.


Last I checked, all Force Unleashed had was a fancy automatic animation system that wasn't AI in the least. Can you provide examples of this magical packaged lifelike AI?

[MarinusWA]

In fact, I'd say that it's almost a guarantee that you end up without a playable game. AI isn't what you want in games, after all. You want AS -- Artificial Stupidity. If the AI was actually as intelligent as everyone wanted, games would be absurdly difficult.
Most existing game types aren't fitting at all for actually good AI. Of course this makes me wonder if there are game types that can only exist with good AI.

[Charles]

Most existing game types aren't fitting at all for actually good AI. Of course this makes me wonder if there are game types that can only exist with good AI.

Sure -- anything where you have a sidekick. But it has to be cooperative.

[Miramon]

I think it's more that no one wants to try to push the edge of it, because it is a very risky edge. Adaptive systems could make for some really interesting games, but you have to be willing to float a lot of money to try to make something work with no guarantee you'll end up with a playable game.

Look at The Force Unleashed, for example. They are using a package AI technology to make the characters in their game respond in a lifelike manner to stimuli. Do they need to do that? Not really, they could just as easily have released Jedi Knight III with only minor graphical improvements and probably made plenty of money. But once they have a handle on integrating DMM, Havoc and Euphoria, they may be able to make a whole series of games based off the same engine, and the things it does that no other game does will be a reason for some to buy it that wouldn't have.

The same is true of Spore. Is there any reason to use procedural animations? They gain a tiny bit in terms of making it easier to create an editor and distribute custom creatures, etc, but they could have made up for that by just hiring more artists to crank out content and making players suck up longer download times that they are already completely used to.

So, basically, the argument "I don't see any need for real AI in games" is short-sighted and meaningless as a counterpoint to the statement "games should be made to utilize AI better to do cool things they can't currently do".

What does procedural animation have to do with AI? I ask for information only. I don't know anything about the Force Unleashed, so I can't comment really, but having characters respond in a lifelike manner -- really? Lifelike? In what way? In animation selection or in actual gameplay behavior?

When I'm talking about AI I don't mean "any kind of risky or interesting programming", I mean what AI researchers research, which has always been the humorous but essentially correct definition.

[MarinusWA]

Sure -- anything where you have a sidekick. But it has to be cooperative.
With regards to that and supportive AI there is still a lot to gain. Sins of a Solar Empire is doing it as well with their (almost) autonomous fleets but they are running close to falling into the replacement trap.

[ravenight]

What does procedural animation have to do with AI? I ask for information only. I don't know anything about the Force Unleashed, so I can't comment really, but having characters respond in a lifelike manner -- really? Lifelike? In what way? In animation selection or in actual gameplay behavior?

When I'm talking about AI I don't mean "any kind of risky or interesting programming", I mean what AI researchers research, which has always been the humorous but essentially correct definition.

Procedural animation is a separate topic I brought up as an example of other types of risky things that aren't really necessary but that get people excited. It has nothing else to do with AI.

If you look up Euphoria, you will see that it uses animation selection routines that are self-contained and based on stimuli, which is an AI problem. Their solution probably doesn't use a genetic algorithm or anything, but that doesn't mean that it couldn't or that having it use one (or a pre-trained neural net) wouldn't produce significantly better results.

My point was that you can look at current games that just hack in some scripting to substitute for "intelligence" and say that is all you need (or that it is better that way, as Charles argued), but you are just ignoring the idea that there may be games we don't currently have that would be more interesting with real AI. If enemies in a shooter were adaptive agents, or the NPCs in Oblivion had a true economy and blackboard algorithms governing their behavior, or the people in a city builder did (and most city builders do have some level of AI decision-making), then maybe you get a game that plays very differently from the current scripted ones. And maybe you create a new type of game based around interacting with AIs (it's not like this would be any more new than Portal or Puzzle Pirates were), or a watershed in an existing genre (like the way 3D changed FPS).

[Charles]

If you look up Euphoria, you will see that it uses animation selection routines that are self-contained and based on stimuli, which is an AI problem. Their solution probably doesn't use a genetic algorithm or anything, but that doesn't mean that it couldn't or that having it use one (or a pre-trained neural net) wouldn't produce significantly better results.

Uh, no. This isn't AI at all, unless you really stretch the definition. It's examining the environment and picking from a list, there's nothing intelligent about it at all. It would be like saying the climbing in Assassin's Creed is AI because it has to analyze the environment and determine where to put hands and feet.

[Miramon]

Uh, no. This isn't AI at all, unless you really stretch the definition. It's examining the environment and picking from a list, there's nothing intelligent about it at all. It would be like saying the climbing in Assassin's Creed is AI because it has to analyze the environment and determine where to put hands and feet.

True enough.

However, there is a commonplace AI trap which is to say "Oh, I understand how the program does that, so it can't be AI".

For example, Douglas Lenat's* early work (and his best IMO) AM (and the associated Eurisko) was able to come up with all kinds of interesting mathematical hypotheses, exhibiting superficial behavior that appeared to show some kind of mathematical understanding. However, it was "really" only doing "syntactic" formula manipulation using some simple rules.

Lenat was criticized for overinterpreting the actual simplistic workings of the program as interesting results like Goldbach's conjecture. However, Lenat argued that the fact that apparently interesting results could come from such a simple mechanism was interesting in itself.

Anyhow I'm sure the cases you cite are so simple that they don't qualify as AI under any reasonable definition, but it's still true that even among AI researchers, a system that more or less works as intended in the boring predictable way that one generally hopes for from software, for example a conventional expert system, is not considered to be very interesting or even to be part of "AI" as such anymore.



*In the Hacker's Dictionary, a Lenat is the fundamental unit of bogosity, but like the Farad, the Lenat is too big for practical use, so generally bogosity is measured in microLenats.

[Charles]

I'll interpret that as "I got trolled."

[Piemax2]

Thanks for all the responses, folks, I found the discussion interesting. If machine-learning type AI takes too much computer power to be practical now, any guesses about when it might become practical?

[AndrewM]

Well, Hans Moravec (http://en.wikipedia.org/wiki/Hans_Moravec) thinks we'll get computers as smart as people by 2050, so...

[MikeJ]

Well, Hans Moravec (http://en.wikipedia.org/wiki/Hans_Moravec) thinks we'll get computers as smart as people by 2050, so...


Yes, but the question is: will they want to just barely lose to me in strategy games?

[Zylon]

If machine-learning type AI takes too much computer power to be practical now, any guesses about when it might become practical?
You're missing the point. A "learning" AI is useless in the overwhelmimg majority of cases, because most AIs don't have a sufficient lifespan to learn anything.

[AndrewM]

You're missing the point. A "learning" AI is useless in the overwhelmimg majority of cases, because most AIs don't have a sufficient lifespan to learn anything.

You could do some kind of offline learning. Kind of like taking all that data they collected for Halo 3 about where people went, and feed it into some kind of algorithm instead of just looking at it.

[Mordrak]

You could do some kind of offline learning. Kind of like taking all that data they collected for Halo 3 about where people went, and feed it into some kind of algorithm instead of just looking at it.

This is a really interesting idea, especially if it is the plot of the game.

Imagine a Sci-Fi MMO where the players are humans fighting off a borg like enemy as the premise. The borg could be a centralized intelligence that collects the combat data from its units as they die, upgrading its tactics along the way.

It's getting thousands of trial cases simultaneously. You'd probably have to have it update it's intelligence like once a day or once a week.

It sounds cool at least. I don't if it'd be that feasible (now or in 50 years).

[Miramon]

Thanks for all the responses, folks, I found the discussion interesting. If machine-learning type AI takes too much computer power to be practical now, any guesses about when it might become practical?

It doesn't take too much CPU, it just doesn't make games better for the cost of the investment. The application of machine learning to games has minimal value for the kind of games the industry makes and that people enjoy playing.

Hans Moravec is kind of silly, by the way in his more widely disseminated pronouncements, just like, uh, whatshisname, Drexler, the nanotech enthusiast. His 2050 date is based on some extreme enthusiasm about Moore's law, not on any real basis in the progress of the fields of AI or cognitive psychology. Mind Children is fun to read, though.

What I'd say the field has learned is that CPU power may well be necessary for certain apps, but it's useless if you don't know what to do with it. And for all the cool traditional AI goals, like passing the Turing test, or recognizing all the objects in a scene, or driving around town, or even walking naturally (that cool boston robotics video notwithstanding), we just don't know how to do it right, regardless of the CPU power we may have to throw at the problem.

So even if Moravec is right about how many cycles will be available in 2050, it still will mean nothing if the basic problems of cognition are only being solved at the same rate as they have been from 1966 to the present day.