Next gen AI ?s

[pc999]

For (especialy next gen consoles but probably it is only a question of time till PC catch it up) next gen AI we already heard to things: learning and predict the player action.

I remember that some time ago (2 years?) MS has searching for people to dev AI that predict (In a sport game I belive) for here would the player pass the ball or for here he would go, and that kind of actions, and then use this predictions, sorry I cant find a link (I think there is one in the dead console forum, if anyone have it please post).

In this http://www.psinext.com/index.php?ca...mp;PHPSESSID=355a98f5b2badf13cd0b1ea2719a7b2d

He cleary talk for a AI in independent thread that constantly learn, and act with a new base.

Do you think this will really be possible anytime soon (and with next gen consoles)?

If it is possible how much performance (minimun), should need, and what kind of features would benefict more (Arnaud_M seems to think that a separeted thread/core? really benefict the learning one)?

What kind of problems will it have (for the devs), it sould be very hard to programe such thing once that the only AI (that I know) that predict is chess and it is very different from a real time app?

If there is any info (easy to understand) please link it to me.

Thanks in advance, and if need please correct me.

 

[ERP]

For (especialy next gen consoles but probably it is only a question of time till PC catch it up) next gen AI we already heard to things: learning and predict the player action.

I remember that some time ago (2 years?) MS has searching for people to dev AI that predict (In a sport game I belive) for here would the player pass the ball or for here he would go, and that kind of actions, and then use this predictions, sorry I cant find a link (I think there is one in the dead console forum, if anyone have it please post).

In this http://www.psinext.com/index.php?ca...mp;PHPSESSID=355a98f5b2badf13cd0b1ea2719a7b2d

He cleary talk for a AI in independent thread that constantly learn, and act with a new base.

Do you think this will really be possible anytime soon (and with next gen consoles)?

If it is possible how much performance (minimun), should need, and what kind of features would benefict more (Arnaud_M seems to think that a separeted thread/core? really benefict the learning one)?

What kind of problems will it have (for the devs), it sould be very hard to programe such thing once that the only AI (that I know) that predict is chess and it is very different from a real time app?

If there is any info (easy to understand) please link it to me.

Thanks in advance, and if need please correct me.

I disagree with some of what's in the article, learning needn't be expensive. It's largely not used because it's undesirable. Or at least difficult to integrate into a compelling experience.

I do agree that high level strategy is largely missing in games, but that's more about development effort than it is CPU load.

Most AI's "speculate" or "predict" although some of it is implicit in the ruleset. In a real time game it's not particularly useful to try and predict long move sets, since the context of the action is relatively short. i.e. What I did 10 seconds ago has minimal bearing on what I'm going to do next. I guess if you could predict the players high level goals then you could change your high level strategy to compensate, but this problem is more about being able to describe goals and strategies than it is CPU load.

What the CPU does buy for you is the ability to make a larger number of queries on the world, having a more complete view of the world to make decisions on.

I still think the local memory size is going to make moving AI onto an SPE difficult.

 

[corysama]

Putting AI in a separate thread is not beneficial in itself. If it was then that would be the way we always did it. Instead, it is a painfull necessity if you have multiple processors want to utilize them. Having systems in separate threads makes communication between the systems much more difficult than it is in a single-threaded setup. If you are smart, you can use the extra work you have to do setting up the communication as an opportunity to make better interfaces between systems. If you are lazy, you will end up with a multithreaded system that is trying miserably to have the same old single-thread-style interface. That is a recipe for much pain.

Any system that is going to try to take advantage of the SPEs must fit within the narrow domain that SPEs excel in: chewing through a whole lot of vector math without making many decisions along the way. Most game AI is still accomplished with state machines and decision tree expert systems -neither of which fit well to the SPE's feature set. Neural nets are probably the best match of the mainline techniques to the SPEs but they are not general purpose and are notoriously difficult to set up. Maybe fuzzy logic systems would be a viable replacement for decision trees? When it is faster to do a lot of math than it is to make a single branch, it can be more effective to accumulate the contributions from all options than it is to pick one.

 

[pc999]

Thank you for the replys .

BTW my interess in AI go beyond games (phylosophy mostly) anyone knows here I can get a good(deeper) info how it works, although in a not very complex level (programing wise). Thanks again.

 

[Mate Kovacs]

Artificial Intelligence: A Modern Approach
I've got it, be warned: it's 65 mm thick.

 

[pc999]

Thanks .

I had found this http://www.cs.berkeley.edu/~russell/ai.html

but it seems that is linked there too.

I will start now.

 

[jpr27]

I think (and hope) that we will see a more robust A.I. this coming gen. I for one have been asking and requesting (on this board as well) that A.I. should be getting more attention especially now that multicore processors have come.

I have been working on A.I. myself in some capacity and though I'm more focused on what my game needs (from my point of view), I really see the possability of expanding it . Its a delicate balance of sorts and although my ideas are solid, implementing takes time and patence :?

I must admit though as much as I thought using multicore would help its more of a challenge then I thought it would be :? In the end I think it will be worth it

 

[SanGreal]

I remember that some time ago (2 years?) MS has searching for people to dev AI that predict (In a sport game I belive) for here would the player pass the ball or for here he would go, and that kind of actions, and then use this predictions, sorry I cant find a link (I think there is one in the dead console forum, if anyone have it please post).

I don't know about sports, but Forza has a "learning" AI designed by Microsoft's Machine Learning & Perception team.

Here is a video where they talk a lot about it: http://channel9.msdn.com/ShowPost.aspx?PostID=23040

Also, www.drivatar.com has some info

 

[pc999]

In the end I think it will be worth it

I am sure it will

I don't know about sports, but Forza has a "learning" AI designed by Microsoft's Machine Learning & Perception team.

Here is a video where they talk a lot about it: http://channel9.msdn.com/ShowPost.aspx?PostID=23040

Also, www.drivatar.com has some info

I had never had the chnge to try it (although I already knew about that feature), would be very interesting know more about these drivatars.

Thanks for the links, as soon as I can I will see them

 

[eastgate2]

As long as players don't want to change their habits, it's unfortunately otiose to talk about "next-gen AI". I haven't seen any improvements in AI as compared to graphics technology lately. People just want enemies that can duck and cover while firing, and when they see this they claim "wow, nice ai". To me the ai in unreal1 and doom3 is not different at all. One was released in 1998, other in 2004.

Can anybody forward me to a game that has better ai then thief or boiling point? I'd like to try those games if they exist.

 

[Remi]

Although I have trouble calling the NPC (non-player characters)'s conduct in games "AI", I remember having fun with NOLF 2's NPCs (NOLF = No One Lives Forever). Their system is rather basic but allows some freedom in play-style: they have a list of priorities, including their daily taks, and they just do what is at the moment their #1 priority. For a gard outside, that gives behaviors like moving to stay warm, lighting a cigarette, moving, going to drink some hot beverage inside, (iterate a few times), heading with haste inside toward the toilets (which - how evil of you ! - you might have actually broke!), and rushing out outside to the nearest tree to finally find some relief... (at which time you can actually help them to relax even further with a gentle good blow on the head).

Very far from a learning processus (it's just another state machine), and yet rather effective to make the world a bit more alive...

 

[jpr27]

Reply Eastgate2:

Thats been my whole issue with the evolution with A.I. over the years you mentioned. Take Half Life 1 for example, for its time I thought it was an excellent A.I. and served the game very well (though some bugs did exist). That along with graphics and physics for the time gave me great hopes on the direction A.I. would be heading. But over the years I have been disappointed by the evolution of A.I. Not exactly "terrible" but in comparison to physics and graphics A.I. was at the bottom of the list for development. IMO even Half Life 2 didnt excel with its A.I. to the extent I was hoping for.

I must say Halo 2, Brothers in Arms have pretty good A.I. that fit well into their respective roles in the game design. I just think A.I. is behind in terms of evolution.

Now it has been mentioned (and I do agree to some extent) that you dont want an A.I. that is so robust that it puts the player to shame On the other hand with next gen consoles, dual core processors, and hopefully more open ended game play (less linear), A.I. will be able to streatch its legs. I think improved A.I could help the gamer experience varied game play so that each time you play through there will be less predictability.

 

[eastgate2]

Jpr27; I completely agreee with you. I sat and thought very much after HL2 about what was missing etc...I don't remember exact details now but there were moments in HL2 where I thought HL1 ai was even better.

Though I am not a coder, my guess is that good ai doesn't require lots of ram or myriad cpu power. Afterall imitating intelligence is also a linguistic & narrative issue, not just a technical one. When a person in real-life "says the right thing at the right time", we go "wow, clever dude"... For instance in unreal1, on a p3 machine, you can make an ai say the different things and play different animations and follow player and/or attack him in different pathnodes and multiply this possibility by 8x8x8x8x8x....so on without slowing down the machine, I tried that I know...

"Saying/reacting at the right thing at the right time" thing could be in games also, I've read enough in game-ai resources to believe more diverse ways of faking intelligence is a matter of system management, not just a matter of excellent coding techniques. So my conclusion is that developers just don't wanna spend more time on designing ai since there is less demand for it. E.g, Valve didn't update their ai between hl1 and hl2,only immense graphics shift, yet hl2 sold over a million and they earned enormous $$$ and I've never read a single review about complaining on ai in Hl2. So why would valve spend more time on ai for HL3? I don't think so.

Games like boiling point and stalker(hopefully) has a different approach to ai, worth to check. Very simply, the environment in boiling point aims for a total, breathable living world simulation instead of individual ai cases.

 

[jpr27]

I think many people seem to be "satisfied" with what A.I. gives us at this point in time. And to be honest it really burns me to no end. I'm just suprised that someone hasnt taken the role of a middleware position (similar to Havoc with Physics) and make A.I. its main focus. I think people will totally be blown away once a developer really raises the bar on A.I. and shows what an advanced A.I. can bring to a game

I know some people like myself are working on ways to improve A.I. and now that multicore has come upon us (which for me right now Im trying to get accustomed to ) I think (and hope) that A.I. will get the attention it deserves for the next generation of gaming.

 

[pc999]

I agree with you, and considering that gfx is, and will each time deliver less impresive results/returns (I, actually, am more interescted in GPGPU) e.g. water is realy nice in XB it would not benefict much in havig 10x more instructions. People will pay more atention once that can see, even, small improvments in AI IMO.

Abount midleware I tink that UE3 will have some nice improvements (at least for what they say)

Artificial Intelligence
"You never want the game to feel like it devolves into 'Oh, I know the AI can't handle this tactic.' When you find a tactic that the AI can't handle, then the game is broken."
Through player-to-bot interactions, Epic wants to give the bots of UT2007 'more personality'
These personalities dictate the particular skills and weaknesses of bots, especially pertaininig to specific goals and tasks
"So if you're doing a flag run, you can hear what's going on back at your base because guys are shouting to each other where the enemy's coming from and what they're doing to prevent it" - Steve Polge
"You'll be able to ask [the bots] questions about their status: 'Is there anybody down in the cavern?', 'No, cavern's all clear'. Or you could throw out warnings like, "There's a sniper on the tower. Go get the sniper.'" - Steve Polge
"You want the AI to do things that surprise you so that you feel challenged in terms of your reaction time, your strategies, and your tactical skills." [QUOTE/]

http://www.planetunreal.com/envy/information.shtml
http://www.1up.com/do/feature?pager.offset=2&cId=3140119

Maybe (if they really delliver it), this wake up dev and dev, because IMO gfx will be done this next gen.

 

[ToxicTaZ]

Real Time AI on a Chip?

Hay Sup?

I got more info for you guys! I found some info on a new AI chip called the (AIS-1 chip) made by AIseek.

Take a look:
http://www.extremetech.com/article2/0,1697,1855163,00.asp

"If the AIS-1 makes it to market, the chip could provide yet another method for Intel and PC OEMs to differentiate entertainment PCs from their more practical business-oriented brethren. It's possible, Mendelson speculated, that the AIS-1 and the PhysX physics chip could be added onto an add-on card "to create a true gaming PC,"

http://www.aiseek.com/

I hope you guys like the info!

 

[Zaphod]

BTW my interess in AI go beyond games (phylosophy mostly) anyone knows here I can get a good(deeper) info how it works, although in a not very complex level (programing wise). Thanks again.

Turing's Man. The book is old, so some of the technical considerations are outdated, but the theme is still very interresting.

Here's a quick primer with references on one of the most canonical aguments in AI Philosophy/Cognitive Science. There's lots of interresting stuff to read on this subject, and Searle is a good starting point.

Edit: I'd also at look at introductory texts to both formal logic and game theroy (not as in computer games), before dwelling into the practical implementation of things - be it in computer games or elsewhere.

Another edit: Didn't notice that this was an old thread beeing bumped. Sorry.

 

[KimB]

Old post, but I don't think this has been mentioned.

I disagree with some of what's in the article, learning needn't be expensive. It's largely not used because it's undesirable. Or at least difficult to integrate into a compelling experience.

Well, I think there are some very simple situations where limited learning could be a great boon. Imagine, if you will, a capture the flag match. When I think of CTF, I usually think of the original UT. In most every map, when playing against bots, I'd find a route that works, and stick to it.

It would be really cool if after a capture the bots wise up and beef up their defenses along the route I took to the flag, and were also programmed to deal with simple diversionary tactics.

This isn't true "learning" AI, of course, but could definitely be compelling in a number of gaming scenarios. In a single player FPS, for instance, you could imagine an area where there are security cameras, and after taking an enemy's radio, you hear somebody speak over the radio, "He's using tactic X, redeploy to formation Y!" Would take some work, obviously, but would definitely make for a challenging experience.

 

[3dilettante]

Though I am not a coder, my guess is that good ai doesn't require lots of ram or myriad cpu power. Afterall imitating intelligence is also a linguistic & narrative issue, not just a technical one. When a person in real-life "says the right thing at the right time", we go "wow, clever dude"... For instance in unreal1, on a p3 machine, you can make an ai say the different things and play different animations and follow player and/or attack him in different pathnodes and multiply this possibility by 8x8x8x8x8x....so on without slowing down the machine, I tried that I know...

From what I've learned, the situation looks to be the opposite.

AI is a problem that requires serious resources or amazingly elegant solutions to get right. For computers, AI is getting hardware that is inherently stupid to appear smart using software that is inflexibly stupid as well.

For chess, the AI relies on massive look-up tables and serious node searching to compete with human players, and chess is a game whose fixed rules, discrete steps, and simple movements makes it well-suited for a computer to evaluate.

More complex games bring up unexpected situations, which software really can't handle. AI in games is simple because any more complexity leads to huge resource hogs that usually wind up failing catastrophically anyway.

 

[pc999]

Thanks for all the new info, people .