Anyone agree that A.I. in gaming needs attention?

[jpr27]

I've tried to get some feedback on this forum and others about the evolution of A.I. It seems that with people hearing about cell, dual core processors, future GPU, next gen consoles and most recently the notion of a dedicated PPU for physics (AGEIA for ex), that much of the focus is mainly focused on graphics and physics.

A big concern of mine is that A.I. is taking a back seat or overshadowed and it seems a bit troublesome especially since the quality of graphics and physics (to a lesser degree) is really outpacing A.I. development (To much so in my opinion). Although Fuzzy Logic, Neural Networks and Genetic Algorithms show promise, hardware still has some growing to do before we can fully utilize them.

I know M$ has stated that XNA is putting focus into A.I. in its makeup but as we have seen time and time again that many games and developers that promises revolutionary A.I. have yet to really deliver. Dont get me wrong I dont think that A.I. is total crap right now. I just hope in the near future that developers take a more proactive approach to A.I. development. Maybe a company thats sole purpose is to develop game A.I. (such as Havok does for physics) might be the answer. To be honest I dont know of any big development companies that do dedicated A.I. Most are in house or maybe cross liscensed from other developersor incroporated into Liscensed engines etc.

I know that A.I. is a very difficult area of gaming but please can we shift more focus to it? I think the next "real" revolutionary game will not be one consisting of just stunning graphics or indepth story line, but a robust and believable A.I. system.

Just my two cents

 

[AlNom]

A developer focused on AI like Havok with physics is just not too feasible IMO. Situations vary per game let alone per type of character so it's not some universal AI code that would automagically work.

Take Doom 3 and Halo 2 for example. You've got Zombies, zombie soldiers, pinkies, hellknights, various demons. Each type acts all differently. Now go take a look at Halo 2, which doesn't feature demons or zombies. You've got little grunts and then jackals and elites who all act very differently again and nothing like human characters.

There is very little in common between just two games.

Contrast this to physics where most games take place in Earth-like environments. Even then, not everything is accurate; developers haven't exactly made sure that every object reacts properly due to friction or wind. The level of detail with that isn't "needed" normally.

AI is just too specific for needs of the game to have such a company.

 

[ERP]

Wrbong forum for this really. But since it's here

While I agree with you superficially, the big problem with AI isn't even making it better or more believable. It's making it better with in the context of gameplay.

Lets say I have 100 levels (or whatever in my game) and I've tuned say 40 of them, when I start working on level 41, I discove AI behavior I don't like (or is nonsensical), have it fixed, then discover the first 40 levels are no longer fun.

The problem with AI (at least any decent solution) is it's inherently unpredictable this makes working with it as a designer is anything but open shooter type environments difficult.

Even if it was "great AI" could you tell? You have mininal communication from an AI, only what it can display on the screen, there simply isn't the resolution or time to display complex emotional state even if it was modelled. If you did have such a model in a fast paced game you'd probably put it down to bad AI and randomness. Anything outside of I did this and the AI reacted this way is pretty much lost.

Example - I rushed the monsters and my AI buddy just stood there and did nothing. You write it off to bad AI, perhaps though the AI buddy had decided from your previous actions that he didn't like you. Without some obvious way to communicate that it's lost.

I'm not saying that AI doesn't need work in games, just that we have to be careful that the cause and effect can be communicated to the player, or we end up in a frustrating game experience.

In some ways physics is in the same boat, it's unpredictability makes it difficult to work with in traditional gameplay, it has a great visual pay off though so people find ways to wedge it in.

 

[V3]

Example - I rushed the monsters and my AI buddy just stood there and did nothing. You write it off to bad AI, perhaps though the AI buddy had decided from your previous actions that he didn't like you. Without some obvious way to communicate that it's lost.

Like the enemies can taunt you, why can't the buddy AI said something like "I hate you, you're on your own." Its what we human use for communication, I am sure its not that hard for AI to do that, with some pre recorded phrases.

 

[BoardBonobo]

AI as used in games seems to be more of a basic action reaction system with the agent reacting to their environment but never learning from the result. A few simple rules would produce what looks like quite a complex behaviour but is in reality quite dumb.

Having elements inthe game that utilise ANNs or maybe GAs to develop a more complex behaviour would IMO be a little bit of a waste of resources. Most games are relatively short and very serialised (and yes there are a few exceptions). There is no real need for a character to develop that far and be so complex.

A simple reaction scale and look-up table of results would give just as an impressive number of reactions and utilise very littlke overhead. Interactions between the environment can also be plotted using such a system. Add the two together and introduce a little element of randomness and IMO that would be as much as you need.

Making a game that uses more AI techniques would only be good if that game used those techniques to the full e.g. An adventure\interactive game where the main proponents learned from their\your previous interactions and from those of characters around them. Basically a game where you would be pitting yourself against several learning systems each with their own goals.

 

[digitalwanderer]

AI will improve dramatically in games once PPU boards are available to free up the CPU.

(Yeah, and then a whole bunch of monkeys will fly out of my butt and throw flowers to all the happy children. *roll* )

 

[ERP]

Example - I rushed the monsters and my AI buddy just stood there and did nothing. You write it off to bad AI, perhaps though the AI buddy had decided from your previous actions that he didn't like you. Without some obvious way to communicate that it's lost.

Like the enemies can taunt you, why can't the buddy AI said something like "I hate you, you're on your own." Its what we human use for communication, I am sure its not that hard for AI to do that, with some pre recorded phrases.

Maybe.... But is that enough and wouldn't it be frustrating anyway?

Once your AI goes beyond simple state machine it's not always trivial to determine (and describe usefully) what an AI's state is anyway.

The point really is nothing more than AI is only useful to the point that you can communicate it's "intelligence" and in some cases good AI does not parrallel with good gameplay.

Traditional gameplay is very much about deteriministic rulesets that a player understands. "I go here and I get shot, I sneak up this way and I can make it....". It's simple to understand cause and effect, trying to make a game fun and not frustrating when characters in it are behaving in complex ways is a challenge.

 

[Frank]

There is a large difference between the AI of an FPS, where you only see a limited set of enemies for a few seconds while they try to kill you and some static NPC's to interact with, and an open gameworld where you can freely move around and revisit locations and NPC's.

For the throw-away enemies, you need no real AI. Just some simple triggered scripts are good enough. But for walking around in a believeable world, filled with changing people, you do need one. You want all the NPC's to go along with their daily life, and you want your interaction to have a clear impact upon their behaviour. Like they're doing with Oblivion.

If you want to allow the player to have a broad choice of actions, it becomes quite hard to script all possibilities up front and you want objects to be as independent as possible. And those actions should be based in part on the possibilities (physics) offered by the surroundings. So, realistic physics and AI go together.

 

[ET]

A developer focused on AI like Havok with physics is just not too feasible IMO. Situations vary per game let alone per type of character so it's not some universal AI code that would automagically work.

That's true, but only to the extent that you could say that about graphics. You have to create 3D models and animation and shaders for each games, just like you have to tune AI for each game. The number of different AI algorithms is not that huge. A large part of the problem is doing things efficiently, since the low level AI algorithms are pretty slow. The high level behaviour is less costly than the low level navigation and "visibility" (determining what's around you).

If you're looking for AI middleware, check out http://www.biographictech.com/. There are a couple of others, but granted, the field is not as advanced as physics.

 

[ERP]

There is a large difference between the AI of an FPS, where you only see a limited set of enemies for a few seconds while they try to kill you and some static NPC's to interact with, and an open gameworld where you can freely move around and revisit locations and NPC's.

For the throw-away enemies, you need no real AI. Just some simple triggered scripts are good enough. But for walking around in a believeable world, filled with changing people, you do need one. You want all the NPC's to go along with their daily life, and you want your interaction to have a clear impact upon their behaviour. Like they're doing with Oblivion.

If you want to allow the player to have a broad choice of actions, it becomes quite hard to script all possibilities up front and you want objects to be as independent as possible. And those actions should be based in part on the possibilities (physics) offered by the surroundings. So, realistic physics and AI go together.

FWIW I agree, that there are really intriguing gameplay possibilities that come from solid AI, but you have to go with the dynamic element and move away from "scripting" and "linear storytelling".

I just think it's a huge design challenge, that hasn't really been attempted outside of the Sims. I know of a couple of games in development trying to do it, and a couple more going part of the way.

 

[bloodbob]

I just think it's a huge design challenge, that hasn't really been attempted outside of the Sims. I know of a couple of games in development trying to do it, and a couple more going part of the way.

Seriously I don't think the ai in sims is really all that good. Actually the whole point of sims is its supposed to have bad AI so you have to interact with them.

 

[ERP]

I just think it's a huge design challenge, that hasn't really been attempted outside of the Sims. I know of a couple of games in development trying to do it, and a couple more going part of the way.

Seriously I don't think the ai in sims is really all that good. Actually the whole point of sims is its supposed to have bad AI so you have to interact with them.

Depends what your comparing it to. The AI is dumbed down somewhat so that the characters don't play too well autonomously.

But my point about the sims is more about how the gameplay design is based around the AI, rather than the quality of the AI.

 

[Ragemare]

I think the area where A.I most needs to be improved is it's tactics. Alot of people complained about HL2's AI because it would run through doorways into your line of fire or would just stand out in the open instead of finding cover or even worse it would run towards you, into the open, when it had cover. Your team mates would also just stand in your way ect.

 

[jpr27]

Well I'm glad I chose to focus more on A.I. in my current project. It just seems to me that although huge leaps are not necessary at this time, we do need more focus on A.I. "awareness and reactive" states. I might be doing this all for nil but it certainly cant hurt

 

[rashly]

AI sucks now. I remember back when HL1 came out it was touted as the next big thing, but it seems like we haven't improved much since then. HL2 AI didn't seem much better than HL1 AI. Even games like Brothers in Arms had crappy AI, and they touted that as a great feature along with the squad based action.

Even RTS games like command and conquer don't feel much different. The enemies in Red Alert behaved almost the same as they do now in Generals.

I wish more games focused on AI. Better AI will have a better impact on gameplay than better graphics or physics can.

 

[AlNom]

Halo/Halo 2's AI were pretty good (driving too).

 

[KimB]

Well, personally I feel that the current best AI in action games is about as good as it needs to be. More interesting would be good AI in games where players interact with NPC's (instead of just shooting them). Oblivion seems to be taking this route, for example.

 

[anaqer]

You've got Zombies, zombie soldiers, pinkies, hellknights, various demons. Each type acts all differently.

And 90%* of these differences stem from the simple fact that some of them can fly, some of them leap, some of them have only melee attacks, others can throw projectiles, etc. So they have different capabilities, but the basic tactics is the same across the board - charge at you and attack at regular intervals whenever the LOS is clear. You could probably construct (and accelerate) an AI that parses a short list of available... assets, for the lack of better term, and recombines these on the fly into a series of sub-actions into a complex response to the current situation.


(* determined by world-class pnooma methodology)

 

[KimB]

If you want a look at dynamic AI, take a look at Spore (basically a tech demo showed off by the guy behind The Sims recently). There was some discussion of the (possibly) upcoming game on these boards.

 

[jpr27]

I think that is where I differ with you Chalnoth. I really dont think that the best A.I. now is as good as it needs to be. My main point is focus and evolution. Anager uses tactics but I break it down furthur with reaction and awareness. (Point here is in enemy, not squad based A.I.)

Take anagers reply to zombies, Imps, etc. Although given that Imps are not a perfect example of overall A.I., think about a few Imps in a room. You come through the door and one of them sees you (before you see them) becoming "aware". Instead of mindlessly throwing fire balls or just jumping into a mele attack and the other follows, would'nt it be better if they were able to misdirect and confuse you by say each screaming and moving in the shadows? Each one trying to get the best position (aware of the others position) and the first one with the opportunity... attacks. As you engage the Imp its buddy then jumps in from the otherside (mele or fireball). Sure they would still end up using mele or fireballs but would be able to "react" more as the player moves throughout room.

Now I know this is the A.I. in its most simplistic form but what Im trying to say is there is no huge leap in A.I. in the above example but a extension of it. (Now that I read my example again using Imps might not be the best example but I think it gives an idea of what Im getting at ). I think its the added immersive aspect that would benefit us at this stage.

Graphics and physics are evolving and I definately dont have anything against it. I just dont want to see A.I. get left too far behind because to me, the improved graphics and physics can really show a weak A.I. At this time its becoming more visible every generation.