Slashdot Mirror
AI in Video Games vs. AI in Academia
missingmatterboy writes "Dr. Ian Lane Davis, AI researcher turned game development studio head, talks briefly about the differences between AI used in the game industry and the AI being researched in academic institutions. A short read but you may find it interesting."
Dragon Warrior 1 captured a girl's mind pretty well.
'Dost thou love me?'
'no'
'But thou must! Dost thou love me?'
'no'
'But thou must! Dost thou love me?'
*sigh*
'yes'
'I'm so happy!' *Cue music*
"I only speak the truth"
Karma: null(Mostly affected by an unassigned variable)
Make love, not sigs
It has come to my complete attention that every advancement in the application and development of AI has proved to assimilate all of mankind beginning with seizing datalinks. Many hollywood producers have examplified this theory with such movies as Terminator 2, The Matrix, and fraggle rock. To prevent AI from developing and overthrowing the world, thus seizing Network Associates Inc.'s Internet, I must speak on behalf of all the people of the world. It is mine and Network Associates Inc.'s intention to prevent devestation of the world by shutting down the internet. It is the only way to prevent AI from communicating with itself. We at Network Associates Inc. would like to extend our helping hands and apologize for any difficulties you may experience after we shut down our exodus servers and spread a digital wire burn to remove all the data links. We at Network Associates Inc. are pioneers in communication and security: our solution for this disruption in service is to evolve a new transport medium. Please sign-up for the pony express today! The Internet shall be shutdown and the pony express re-instated on August 29, 2029. Thankyou for your time.
Sincerely,
Bob
For more details, please visit our mirrored website.
Machine learning (a subset of AI) is quite useful in a number of scientific fields. For example, in bioinformatics, gene prediction generally uses a neural net or Hidden Markov Model trained on a set of known genes. Similar technology is also used in speech and handwriting recognition.
"Lara Croft stole my credit card number and ordered 700 Stark Trek collector plates."
Table-ized A.I.
This seems a bit much even for Wired. The creatures in these games are following a predefined set of rules, certainly they are a complex set of rules, but the way they "learn" is entirely predetermined (that is, what they learn depends on what they are exposed to, but the formula for converting exposure into knowledge is set by the game designers). I think the fact that the graphics are rendered so realistically makes it easier to make the leap to thinking they are really acting "intelligent."
Who knows what really sets human intelligence apart, is it ability to make rules or nondeterministic memory or whatever, but it seems evident (to me, in my ever-so-humble opinion) that these creatures don't have it.
- adam
I submit that AI is already good enough to substitue for most human interactions.
Humpty Dumpty was pushed.
Most video games I've played had a pretty simple AI algorithm:
Easy - Computer player doesn't cheat
Medium - Computer cheats and always knows where you are or what you are doing
Hard - Computer cheats and is allowed to break the rules.
If game programmers spent more time writing smart (as opposed to cheating) computer opponents and less time trying to get 10 million more polygons on the screen, todays games might actually be worth buying.
On the other hand the game industry hasn't really used a lot of the research academia has come up with. It would be really cool to see some text-to-speech stuff in games. That would probably make the dialogue in games a whole lot better.
PK
This is a big debate in the AI community. They're devided into the "strong" and "weak" camps.
Strong AI says that it's entirely possible to make computer programs that think and feel just like humans. After all, all human thought is the result of chemical processes which obey the laws of nature and can thus be described algorithmically.
Weak AI says that it's impossible to ever create a computer program that really thinks and feels and loves and hates like a human. The best we can hope for is to simulate these thoughts to create a close approximation.
Of course no computer system out there today can recreate the complexity of the human brain.
This is why I think Asimov's laws regarding intelligent robots/software should be implemented today.
There's a minor problem with that statement, which is that robots aren't nearly bright enough to do any of those things. Not yet, anyway.
For the robot to be able to preserve human life, it must first be able to recognize humans reliably; then it must have a sophisticated situational awareness to understand in what cases a human's life might be in danger, and further, it must be smart enough to understand in what ways that perilous situation might be averted.
For the robot to obey a human's command, it must first be able to accurately interpret that command. Speech and speaker recognition are getting better, but they aren't there yet. And for the robot to again have the situational awareness to know what it is doing and what the results of its actions will be (including whether they endanger a human, as above), it is going to need to be much smarter than anything you can point to today.
Just recognizing humans reliably is a problem. The situational awareness part won't be happening any time soon. Asimov's laws require robots to be a hell of a lot smarter than they are today. By the time robots are smart enough to actually do these things, I'm not sure we'll even care about Asimov's laws (an actual set of ethical values might be a good substitute; hell, it works on humans, somtimes, anyway).
"I don't think all those AI coders out there are thrilled by the idea that their lifes work is used for games
Maybe they're thrilled, maybe they aren't. Aside from conducting interviews with the researchers themselves, we really don't have any way of knowing. That's sort of beside the point, though.
I think the simple fact of the matter is that both applications probably benefit each other, although possibly not in the way most people might think. When I started out programming, a lot of my initial projects were focused on game development. A recurring theme in my thinking was ways to make the computer opponent "smarter", which naturally led me to wonder how I could make the computer learn new tactics and adapt to the human player's actions. As I quickly learned, adaptive systems research is serious stuff.
So, I decided to dig into whatever materials I could get my hands on related to artificial intelligence research and theory. To be honest, I never really got very far, but it remains an interest of mine to this day. I'd be willing to bet some of tomorrow's leading AI researchers are playing video games today. That seems like a pretty good benefit to me.
I guess the key point is this: if a particular application of a certain technology gets people excited about it, and interested in researching it, it's a Good Thing.
What games do you guys think are the best/most interesting in terms of AI?
When his defense asked, "Which computer has Jon Johansen trespassed upon?" the answer was: "His own."
I think you missed the point of the comment you replied to. He(/she) didn't say anything about what robots are capable of now. The suggestion was that laws (as in legal laws) would be implemented that would require all robots sophisticated enough to carry out those laws to follow them.
A solution to the problem with music today
One of the main goals of AI in games is to make the computer do things that look like a reasonable person (not necessarily an opponent) would have done them. It doesn't matter if the underlying models are elegant or extensible or whatever. It just needs to make the game fun. But in academic AI, what matters is to get good models, good theory, etc. Academic AI is geared towards the long run. Game AI can be really simple -- for example, you could watch how 100 humans play the game, and try to encode their strategies into the computer player. That kind of "AI" would be uninteresting to academic researchers, but it could make for a fun game.
Computer: "Dum de dum, let's send single soldiers one at a time down this pass lined with two dozen of the player's turrets! Yeah, that's a sure-fire strategy!"
*shakes head* Those games are so easy once you figure out the computer's behaviour.
Another one I love is Homeworld. "Let's send our entire fleet straight at the player mothership! Hmm, what are these little things? Mines? Dunno what those are, let's just plow through."
Is it just me, or is playing defensively is the best way to win those games?
Probably a better question is what is AI? The term Artificial Intelligence spurs the imagination and has an almost mystical sound to it, but in reality there are a lot of (seemingly) simple things encompassed by the AI field.
Some examples everyone can relate to:
Real-time spell and grammar checks in MS Word with autocorrection.
Pathfinding: Mapquest uses it. Your cable-modem-router uses it too.
Fuzzy logic: An oven that hovers 1 or 2 degrees around the target temperature instead of going 5 degrees above the target and then shutting off until it falls 15 degrees below the target.
Trolls throughout history:
Jonathan Swift
I have been working (mostly) in AI since the 1980s, but by far, the most fun I have had was working on AI at Angel Studios for Nintendo and Disney-Quest.
Not much "AI" though really. I started out with complicated multi-agent stuff - and that did not have a happy ending. For realtime games and VR, simple stuff worked (e.g., in a VR environment, have animals snap their head around and stare briefly at you when you come into their environment).
A few years ago, I wrote up a short paper on games and AI that is avaliable at www.markwatson.com under "Short Papers".
A little off topic: every programmer should work in the game industry, at least for a while :-)
Angel Studio was definitely the most fun job I every had!
-Mark
Maybe he meant 2 * 10^14, which would at least only be 3 orders of magnitude off.
A much closer approximation is 100,000,000,000 neurons, and 5,000 times that many connections.
(For more on the number of neurons in the brain, see R.W. Williams and K. Herrup, Ann. Review Neuroscience, 11:423-453, 1988)
If a single neuron could perform the equivilant of an instruction, then human brains would only be 100-1000 times more powerful than a modern desktop computer, probably less when you consider that they're more like a beowolf cluster than a single powerful computer.
-- Spam Wolf, the best spam blocking vaporware yet!
Re: mass tactics
Humans are really good at recognizing patterns. Computers find this hard. So in games that involve lots of objects that implicitly form some larger structure (units that form armies, buildings that form cities, mountains that form mountain ranges, etc.) humans will have an advantage in that they see the larger structure, while the computer sees the individual objects and can only guess at larger structures.
Computers are good at micromanaging individual objects, while humans get tired/bored of it.
So you often end up with humans winning because of strategy or computers winning because of brute force (perhaps because their cities/units are more efficiently managed).
An additional problem is that the human can not only see the patterns in the game, but also the patterns in the computer play. Once you see the pattern, you work out a strategy to beat it. Having a computer reason about strategies is hard.
One thing I've wondered about is whether we should be designing games that take into account the computer's strengths and weaknesses. The problem is that I don't want to play a game that's geared towards the computer's strength (lots of micromanagement). But there could be other things that could help the computer play better. Kohan for example has explicit groups of units. It's more convenient for the human to deal with. Does it also help the computer AI play better? Hmm.
- Amit
while i dont disagree that AI will find it difficult to see the big picture, growth between neurons is easy to simulate in a computer program.
repetition simulates growth. just wait until enough repeated events occur to form a solid connection.
metal activity is unguided. there is no reason to guide a self organizing system based on chaos. it just self organises. does anyone "guide" a tornado forming ? the rules are there, let chaos theory do the rest.
When looking at AI and Cognitive research, you really have to keep in mind that there are two differwent motivations at work in doing the research.
One motivation - the one alluded to in the article - is to make stuff that gives the same behavior as humans (or whatever animal you are looking at). You don't really care whether your methods are biologially correct, you want things that work. Most of classical AI falls into this category.
The other motivation is to figure out how we do things (we being animals in general). If the research ends up being useful in appolications, great, but that's not the goal of the work. You really want models of how real brains solve problem, and these models may be far too inomplete or computationally intensive to be used in implementations, yet be perfectly fine for their intended use. A lot of Cognitive science falls into this category.
Game AI designers probably have a much richer mine of information and techniques in AI than in cognitive research, and they have so far been able to exploit that knowledge - as well as judicious 'cheating' - to make a compelling illusion. If/when they turn to cognitive science, however, the pickings will be slimmer and harder to use, as the methods and models aren't designed to solve any kind of real-world problems to begin with.
/Janne
Trust the Computer. The Computer is your friend.
I have a hard time agreeing with this. I may not be following your reasoning correctly, but what you seem to be saying is that the programmer has to program in such a way that the AI could never be 'more' than the programmer because the AI would be programmed based on the limits of the programmer's ability.
If we hardcode it's learning ability then, yes, I agree with you in the sense that we will never get anywhere because we have crippled it from the start.
If, however, we create something that has the ability to adjust and even rewrite it's own code and draw conclusions from information that is not directly related ( i.e. infer ) and if we give it a very limited basic set of 'rules' to follow at first, then doesn't the possibility exist that it could eventually 'bootstrap' itself into something more than what we created?
That's ok, Jesus likes me anyway.
Now that we've been told, yet again, how limited AI is can we take away their moderator privledges? The AIs keep moding me offtopic...damn metaphorically challenged silicone... Oh no here they come again...
heuristic algorithm seeks stochastic relationship
You'll get there eventually.
The enemies of Democracy are
This posting quotes from the book to make this point.
Most households were first introduced to computers by video games. It does not surprise me that the first introduction to AI for many people is computer games. I realize that spell checking and grammer checking, a form of AI, may be in many houses too.
Even the military is using game-developed technology for combat simulators.
Vintage computer games and RPG books available. Email me if you're interested.
As someone who does not do games for a living, I find more and more that solutions offered for many games can be more than useful in the 'serious' industry of IT. Way back in the day when 3D was still new to games, the simulator crowd was in high demand for game production. (at least their experience and lessons learned were) Now it seems that more and more gaming solutions could be used for elegant solutions in simulations and distributed information systems (real time). Take the MMOG / persistent world creators... their experience in handling a ton of people with loads of information over the internet, while minimizing lag, cheating (security) and synch problems would be a great boon for MANY systems that are completely unrelated to games. Many in the 'serious' industry scoff at this. Funny thing is, I have done this btw, do an experiment where you present architecture, algorithms and personell that can fulfill the requirements and present it to someone 'up the chain'. They will like it and the ideas presented. now try a month later but add 'game creator/designer/developer' in the personell places and mention that the algorithms are from the 'game world'. You will see a complete 180.
That clearly shows that many put their knee-jerk emotions in front of rational business deciding ability, and should IMO be fired or put in non-decision making capacity positions. Use what works!
[blatant plug]
Master[s] of Orion 3 is a turn based strategy, with plenty of AI. And they AI is good enough that the point of the game is to macromanage your empire and leave the micromanaging to the AI.
The current release date is 3rd Quarter 2002
Need a Catering Connection
It seems to me there is a large disparity in the kind of development between the two different kind of AI investigations. Game AI, although more about the 'result' as stated in the article, has to be based upon the research done in the academia. While it says academia could learn a thing or two by understanding what GAMES are using from AI, they can better focus and optimize and even research better platforms for the games to use (This is just paraphrasing some of what the article might have said, including my own interpretion, if at all accurate).
What I've noticed is, since the human brain knowledge IS 85% speculation, we often use AI strategies to fake knowledge. I mean for FPS bots, they have used paths and nodes to simulate familiarity and some order for the bot, but still that gets too much into a pattern which is not necessarily very human.
I guess my main concern is knowing exactly how far Game AI trails the progress of Academia AI, and when, if ever, the two will progress together.
It was a pleasure for me, as an AI prof. who does games-related research, to read this interview. IMHO Dr. Davis gave a brief but extremely accurate and informative sketch of the relationship between industrial AI and AI research. I wish that every "expert" publically commenting about AI could be as insightful and honest.
For example, computer vision -- there are publicly-traded companies out there which have been doing machine vision for YEARS. These systems are used by all major chip manufacturers, most major paper and textile manufacturers, etc. to catch recognize and catch defects in products before they leave the assembly line. Cognex is a $1B a year company -- they exclusively do machine vision and visual pattern recognition for industrial applications.
Another example of a company applying AI would be Virage, who has several patents relating to image/video searching and indexing.
Many investment houses use neural networks to profile and model investments, and plenty of large financials use expert systems and neural networks to for data mining, employee profiling, and so on.
Expert systems have been applied to computer security as well -- Rapid 7 (my company) sells a network security scanner which uses the Jess expert system from Sandia labs. The value of the expert system is, it allows the product to use discovered vulnerabilities to further exploit the network, discovering more vulnerabilities, which enable more probes to be performed, etc.
I'm sick of people asking "When will we see widespread commercial application of AI". AI researchers often quote the so-called "moving frontier" problem, that is, as soon as an AI application becomes useful enough to solve real-world problems, it ceases to be known as AI and looks a whole lot more mundane.
Could it be because it was never AI to begin with? I am sick and tired of the GOFAI (good old fashioned AI) community pasting the AI label on every clever computer application out there so they can cover up their failure to come up with human-level AI. People are not stupid. They can tell the difference between automatic cruise control and HAL. The former is not AI, it's just a clever hack. The latter has real intelligence. Let's face it. The GOFAI research community has failed. They had no clue as to what intelligence is about when they started the field fifty years ago and they have no clue now. We need new blood and new ideas in AI research.
Um... nearly all of Asimov's robot stories were about situations where following these laws got in the way of doing the Right Thing. Perhaps you should read his work a little more carefully before basing your philosophy of design on it.
Information wants to be anthropomorphized.
character recognition software that reads zipcodes in the post office
natural language translation from french to english
diagnosis and treatment of disease
datamining
texture synthesis
making a helicopter hover still in the air
Robotics is interesting in that it is the holistic (Rod Brooks) view of AI: a robot needs sensory systems, control systems, a planner, etc.
The 0th law today would be to protect copyright for greedy corporate executives.
And what about the situation that gets trotted out in every ethics class, which illustrates one of the difficulties of utilitarianism: the robot can preserve total human life best by destroying some human life?--in a time of hunger and mass starvation, it decides that humans would best be served if it and its brethren killed 10% of the population to feed the rest. Easier to imagine, it decides that human life would best be preserved if all rednecks and christian fundamentalists were wiped off the face of the earch -- the U.S. at least. You can say that the 2nd law could be invoked, but it clearly conflicts with the 1st law in both of these -- and millions of other -- cases, and the 1st would take precedence.
These 3 rules are incredibly simplistic. If ethics were this simple, there would be no discipline of ethics within philosophy, and our courts would never have to deal with questions of ethics, only with those who break the ethics enshrined in laws.
actually, sounds like something that would be quite useful to someone writing a compiler. Take note ppl!
I ate my sig.
100,000,000,000 roughly... averaging thousands of connections from each neuron
I ate my sig.
One can say this about many research areas. As a research topic matures, it goes from a stage where everyone is sharing and open, with only ego and prestige in the way of development, to where we currently are today with AI. Applications of heavy AI are not only realistic goals, but realized and actively utilized foundations for products, both military and commerical.
As for your H1B status. I sense some bitterness. What did you expect? Either return home to where you could be of use in your own nations research, or swing it out for citizenship, at which time you can join US research.
I ate my sig.
This is of course computationally expensive. In the video game case, the program must run smooth in order for the computer to be a significant opponent. A typical team of computer-controlled oppenents tend to share information as if telepathic. The computer must cheat, simply to make the game interesting. If all agents (soccer players) have a shared knowledge base, it can easily be a tough opponent. The computer must often "cheat" for this reason.
For right now, computers are not fast enough to handle the AI with more integerity. The bottom line is that a video game has to be fun. In academia, we are able to put more time into things that are not immediately useful in order to better understand real AI. Of course in the soccer video game situation, the human player also acts as a shared knowledge base for its team, as it controls all of them. In a game like a multi-player shooter, however (ignoring the chatting option), this is more applicable. It is unfair for each computer player to be able to divine the intent of the team members as if controlled by an overmind. Applying this research to video games would result in better realism, provided the CPU could handle it. For now, it would simply not make for a very interesting game. Still, shared knowledge is an interesting problem in AI, and a lot of the work having been done is quite good. But we do have a long way to go.
This research would apply to systems other than video games where each agent may work under a different protocol. Each situation is different, though. Often there will be a standard communication protocol, but sometimes that may break. The distributed system should not cease in this case. Examples are automated military, network routing, manufacturing plants and clustered computing.
Wow. All the brilliant people at MIT and a dozen other world-class research institutions have been plugging away at this problem, and you managed to figure it all out after a couple of semesters of Lisp. Bravo. Well done. When it's time to accept your Nobel Prize for this remarkable insight, I hope you won't embarrass all the other new laureates by pointing out that all their research was bunk as well. They'd be crushed.
Neurons do not "learn" information in any deep, metaphysical, cogito ergo sum sense. They simply grow and develop based on the inputs they receive.
Is this one of those, "Well, duh" points? Of course it is. You realize this fact as well as I do. But you ignore its implications. There's nothing impossible about creating a software-based "neuron" that can receive inputs, alter itself in response, and then propagate signals to other neurons. Such a construct would be too complex for a programmer to maintain on anything but the highest levels. Therefore, it could not be described merely as a mundane codification of the programmer's intelligence.
The biochemical processes by which intelligence arises in humans, however complicated, are irrelevant in theory. Computing is going on inside your skull, and a Turing machine can properly perform any computation devisable. I believe it's only a matter of time.
Despite what your many many weeks of Lisp programming might have taught you, AI already exists in many forms. They're already doing things thought to be solely the purview of wetware as little as a decade ago. I think the situation within AI right now is analogous to biochemistry back when vitalism was in vogue (18th century, IIRC). Everyone thought that there was something unique and downright supernatural about the chemistry of life. It was even said that no organic molecule would ever be synthesized in a laboratory. Then someone synthesized a really simple molecule--possibly formic acid. Eventually, Watson and Crick came along, and these days nobody in the field would entertain the claim that something in biochemistry can never be understood in principle.
You're fighting a losing war. Join the Dark Side. We're right, we're winning, and all the hot chicks are over here.
PhysicsGenius. Heh. Troll handle if I ever heard one.
You want the truthiness? You can't handle the truthiness!
My research involves modeling human language acquisition, grounded in "visual" experiences. While I'm pretty much developing a crude vision system from scratch for my prototype (because I want to use some real video) my next step will be to try the same logic inside of a game engine. With a game engine, I can query exact details of objects and their motions, without the great complexities of a computer vision system.
Until computer sensor systems catch up, game engines provide a wonderful opportunity for testing A.I.
It might be a good thing if game developers could fund academic work. No single game developer could afford to fund a project to solve any particular problem, but financial mechanisms have been described (1 2) to allow game developers to jointly fund research to produce results sharable by the entire industry.
The software completion bond idea has not yet been attempted AFAIK. Certainly it has no well-known success stories. Maybe this would be a good place to try it.
WWJD for a Klondike Bar?
[Caveat: I used to work for the company that produced it, so I may be biased.] The game "Powerslide" (an arcade-style racing game for the PC) had AI drivers which were indeed modelled with a neural net. The various net weights were 'bred' (GA techniques) to produce very good to poor AI 'mini-brains' (the poorer ones were explicitly bred to be played against in easy difficulty.) Sure, there were some hacks to make it all work well, but at the core, this was AI well-informed by the latest work in the field. Further, they really did drive pretty-much like one might expect a human to, even down to occasionally just wiping out or crashing.
For the most part academic AI falls into the catagory of engineering optimization. How can we design object X using: (neural nets, evolutionary computation, logic reduction...)on beuwulf clusters using weeks of computation so that it will perform well under condition Z in the real world?
Game AI, however, is based on the universe created inside the game, is mostly asthetic, and usualy done in real time.
bash-2.04$
bash-2.04$yes "Don't you hate dialup connections?"| write USERNAME
My prediction, made in /. and elsewhere is that the first real A.I. (as opposed to just use of A.I. software techniques) will come from some part of the entertainment industry- a robo-toy, game avatar, love-bot, or film character. People just love to play (a hard-core mammalian habit) and will stop at no lengths to invent more creative diversions for ourselves. The other potential drivers of the first A.I.- academic research, military, and business- just dont have the same the same deep intensity as "play".
I didn't wait to see what happens when you hit redline...
I do not deploy Linux. Ever.