Nice post Kevin, and great to see some life on our blog! I totally agree with the points you are making, although I would draw the line as hand-crafted vs data-driven. But either way, I agree that the best solution is some combination of the two. And, just as in animation, there is no winner-takes-all – some situations call for hand-animation, some call for motion capture, some call for physical simulation, and some a combination (such as tweaking mo-cap data by hand). These are all tools available to the artist. The problem in game AI right now is that we pretty much only have the hand-crafted options, so we need more tools in order for the discipline to grow to support new experiences (or better-support existing experiences).

Jeff

Good to see another post on here! :)

> The artists will argue that after roughly half a
> century of AI research, none of those techniques
> has been shown to be able to create the quality
> of behavior that we want to see in our games.

See, OK. This is a bit problematic, because you throw so many things in opposition to the "craft a behavior tree or HFSM" approach that it's difficult to make a clear-cut argument.

Specifically, you threw in things like player modeling, planning, and reasoning in the bucket along with machine learning.

It's easy for me to make the argument that machine learning isn't very useful for modern game AI.

But Jeff pretty much proved that planners CAN do an awesome job with his work in F.E.A.R., building an AI that kicked the ass of anything I've ever worked on. A certain level of player modeling has also been proven in some games, and Damian has argued very persuasively in favor of certain reasoning techniques.

So I think no one can argue that those techniques are useless. At most, they can argue that *SOME* games don't need them.

For me, the bottleneck really comes down to DESIGN.

Very few designers (maybe none other than Will Wright) really understand what's possible with AI.

That's the real bottleneck. That's what's holding AI back, and that's what we need to fix to move the industry forward.

It's that we very often design games around dumb AI because our designers don't know how to design for it -- they have no sense of what's possible, and they don't know how to design games in ways that make planning, reasoning, and player modeling possible, useful, and meaningful.

I'll spin the argument another direction. :)

I feel like the biggest bottlenecks are first producing volume content to present the behavior and secondly the amount of time it takes to specify the rules that drive the behavior.

I see the rule specification problem as a breath issue more than anything once you start focusing on believable characters; the number and depth of social rules that govern human interactions, the number of potential actions people can take (and the methods of interacting with objects).

The behavior domain for a believable baker is vastly different from that of a boxer or gunslinger. Humans are great pattern matching machines which makes the little failure in protocol break down the system quickly.

I'm not sure what the best way to encode all of this is, nor do I feel like the academic has been making rapid advances in such broad areas. In general, most real world problems to which AI techniques can be applied often seen more algorithmic than representation based.

As for Pauls point about dumb AI, I'm not sure I agree that it is designers underestimating what can be delivered. Every design I've worked with has wanted more than the AI group (or any group for that matter!) could produce.

I put the challenge of demonstrating interesting, marketable applications of AI tech squarely on us (and on the market demand). If we can showcase showcase an compelling/marketable new experience leveraging AI tech, with little risk to the schedule, I doubt we'd see much pushback.

What if Spore took it a step further and allowed the player to create (or at least tweak) animations? You would need a simplified editor in the same fashion as the creature editor and similar compression/distribution mechanisms. At that point you're getting free animation data from your customers. Granted, a lot of it would be crap but some of it would be good. You could play gatekeeper and decide what does and does not make it into the game. Allow players to rate the content and the good stuff gets pushed to the top. Of course, you still have the challenge of figuring out which animations to play and when. How would you go about solving that?

We've seen how successful games with mod tools/editors can be. And most of them are so complex that they take months and months to master. Not to mention learning Maya/Max. 99% of the people that fire up an editor probably quit fairly quickly. If you could keep those people around (by simplifying), just imagine how much more content your game would have.

@PaulT: What's the solution? Educate? Eliminate? or Provide them with better tools? I vaguely remember reading a list of the top 25 designers and almost all of them were programmers at one point in their careers.

I think the distinction is blurring even for traditional games and animation pipelines, not just Spore and GTA.

If you look into modern Skeletal Rigs, they include a wide variety of procedural tricks that simplify hand animation, almost to the extent that Ken Perlin uses. Ken is creating a system to procedurally animate characters, and my argument against him in the past was that we can't all afford Ken on our projects, and he's the only guy that can do this -- so far!

However, with the skills that technical artists have to animate bones procedurally in the Skeletal Rigs, I don't think it's that far fetched. Animators rely on this technology all the time now... The question is, where is it at runtime?

Alex

If, as Paul suggests, "our designers don't know how to design for" what the AI can do, perhaps AI people need to write a few more indie games showcasing what can be done with the AI that is available?

> @PaulT: What's the solution? Educate? Eliminate?
> or Provide them with better tools?

Eddie, I think the only real solution is "all of the above."

We need to educate designers about what's possible (not least through the AI Summit).

We need to get more people who understand AI and how to design for it in design leadership roles (and kick out the ones who think only multiplayer matters, who think that design is an extension of art, or who only want to dumb games down).

And we need to give them better tools and more stable platforms, too, to make it easier to design good AI.

> I vaguely
> remember reading a list of the top 25 designers
> and almost all of them were programmers at one
> point in their careers.

That may be true, but I'm not sure it's relevant; even a lot of programmers unfortunately don't understand what's possible with AI and what it can contribute to design.

There are programmers who are tremendous at physics and rendering and still don't understand where AI is at and what it can contribute. I remember reading an interview with John Carmack a few years back where he went into so much detail on his new super-cutting-edge ultra-high-end rendering tech. And when it came to the subject of AI, he said, "Oh, well, we just use scripting." Kind of heartbreaking, really.

Ask a top tier AI engineer about their AI systems and chances are you'll get an extended description covering techniques used, research, etc. Ask them about their rendering tech and they'll probably say 'Oh, we use shaders'. ;) I view Carmack as a rendering technologist in the same way many here are AI technologists.

I may be out of place here, but I get the sense that the sort of AI applications Paul is interested in are ones that would likely be associated with new genres. If so, that moves into a different sort of conversation as (in the big budget game space), that becomes more of a publisher/marketing sales job than anything.

I think the comments earlier about indie games and small budget games are right on. This could be a way to experiment with leveraging AI approaches in a less market driven manner. In many cases, designers are delivering what publishers want.

Great post. I don't agree on the distinction between artist and "proceduralist" though - after all, there are Procedural Artists.

Good procedural animation/behaviour requires as much creativity (if not more) as traditional animation.

I don't think there is one definite solution for the problem you propose. Some designers may choose to directly create all content (and make great games), other may choose to do it all procedurally.

Personally, I'm anxiously waiting for an adventure game with zero pre-defined content, in witch everything is procedurally generated (of course, according to rules defined by the designer). This would require much more investment in realistic character behaviour (players still value much more graphics and action aspects of gameplay).

The best acomplishment so far in this sense to me is still Facade (www.interactivestory.net/) - wich is still very limited, although promising. BTW, If anyone has any other suggestions/references about games/projects of this kind, please post here!

ps: sorry for my limited english!

> Ask a top tier AI engineer about their AI systems
> and chances are you'll get an extended
> description covering techniques used, research,
> etc. Ask them about their rendering tech and
> they'll probably say 'Oh, we use shaders'. ;) I
> view Carmack as a rendering technologist in the
> same way many here are AI technologists.

Well, yes ... what I was trying to get at in that case was that Carmack was basically stating that scripting was all that was needed for AI, and doing anything more was pointless.

So, I really should have explained that more fully. I fail at making points.

Perhaps the missing voice in this debate is the end user, the player? My feeling is they want fluid, non repetitive animation that makes for engaging characters that don't break their suspension of disbelief.

There is a part in that description for procedural and hand/mocap techniques. Each have their strengths but blended together could be the best solution currently. Procedural can make characters fluid and non repetitive on top of which the artist can add their special touches.

Another voice to be heard is the business manager, something that was mentioned implicitly, in the sense that hand animation and manual content creation in general is simply no longer financially tenable. Content creation costs are increasing every year while sales decline, something has to change, content has to go procedural so games can keep moving forward.

I think the way forward is to not shoot for the moon , trying to have intelligent characters "out of the box". We create characters that out of the box just look alive and attentive and can react, simple goals that can make a powerful difference but its like plumbing, it is the base on top of which everything else is built.

Personally, I think the debate over Procedural and Handmade AI techniques is similar to the 2D pseduo-3D vs 3D debate of the mid 90s. Due to limited processing power, 3D models of the time were very poorly drawn and often slowly rendered. 2D arists could create the illusion of 3D very simply by maintaining a very careful sense of perspective as assets were designed (ala "Diablo"). Eventually, 3D took over as the primary method of drawing because graphics cards could process models and make them emulate the hand drawn quality as time went on.

It's similar to this because we'd prefer Procedural AI to Handmade AI - why do any work that the computer can do for you? However, increased CPU, GPU, and memory resources aren't going to allow developers to have a fully procedural AI for quite a while, if ever. This is because most improvements in hardware are still being used to help optimize subsystems like Physics and Rendering to act on the ever-increasing size and complexity of game assets. Also, because Procedural animation is rather new commercially in the games industry, a lot of the algorithms in use are unreliable, buggy, and slow.

I think the best we can hope for is a kind of hybrid system that uses Handmade AI for realism and Procedural AI for interactions that would be too memory consuming for keyframed animation (like grabbing a gun that could be facing in a hundred different directions correctly).