The way I see it, the closest thing to "AI hardware" would be something allowing for massively parallel computation. Something like a card that is simply an array of CPUs, all of which have all the functionality of any other CPU (even if they are RISC instead of CISC; come to think of it, maybe that would be better). Even this system shows a bias towards certain types of AI... specifically NNs.

Here''s the deal:

Graphics cards work only because you can take a tecture and throw it into vram, so every time you want to draw said texture to the backbuffer, the cpu only has to pass the video card enough information to tell it which texture and where. It''s a proven technique, and it works like a charm. The less going through the card bus, the better.

What Carmack was probably trying to envision with his "polygons bad, solid models good" was the idea of putting not only the textures into vram, but also the geometry of the object being rendered, so all that''s going through the bus is a pointer to the model in vram and a dozen or so bytes about where and how it should be drawn. Basically, he was talking about geometry acceleration, and he was right, it came to be, but not in the form he was expecting.

Now, with AI, the ram that holds the "intelligent" entities has to be somewhere where it can quickly tell the video card where it is and how to draw itself. The memory doing the "thinking" for this entity also has to be somewhere where it can quickly be copied/compared to said entity so it knows what it''s next move is going to be, and what to do about it.

An AI accelerator card wouldn''t work without a unified memory architecture simply because there would be too much data flying back and forth across the bus to make it worth while.

Now, what we CAN take off the CPU, thanks to geometry acceleration and nvidia''s nFinite FX engine, is physics, which takes a ton off the CPU for anything with a physics engine comparable to a Quake game (read: even the shittiest of physics engines would benefit).

See, nFinite FX let''s you "program custom accelerated effects." That basically means you can write a function and put it on the video card to be run by the graphics processor. This is REALLY cool, because this function can manipulate anything in vram without touching the CPU. With geometry acceleration, the vertices that make up your world are on the card, so if you write a function that takes in a pointer to a model in VRAM, and information about a force acting on that model (vector, magnitude, spread, etc.) and put it on the video card, then all the CPU has to worry about is the couple dozen bytes that make up the parameters and return value of that function. Those parameters give the function all the information it needs (if the model in vram is detailed enough) to alter the geometry of that model to reflect whatever impact it just received. To be nice, the function could return a vector representing the changed trajectory of that model, and VOILA! your car has a realistically smashed and crumpled hood that looks like a painstakingly hand-crafted model even though it isn''t and is spinning away from whatever hit it on it''s center of gravity, offset by whatever friction with the road there may be.

Meanwhile, on the CPU side of things, the opponent driver AI can go through a "I just hit something" routine and a little model inside can run through a scripted grin/flip off animation so all the kids at home can say "WOW! The AI in this game ROCKS! The driver''s know when they hit something!"

(sorry, I''ve been under the opinion that game AI isn''t)

Maybe there could be some sort of hardware that accelerated logic operations, such as inferring facts from existing facts and rules. You''d send it little scripts in something resembling Prolog, if not Prolog itself, and every so often you could send it a query of some sort which it could answer in super-fast time Predicate calculus seems mathematically formal enough to be able to put onto hardware quite easily - the question is whether games could find much use for it (and thus, drive the market).

[ MSVC Fixes | STL | SDL | Game AI | Sockets | C++ Faq Lite | Boost ]

On the idea of using Predicate Calculus...

First Order logics just don''t cut it in AI. Certainly they can add some depth to current Game AI but ultimately, you wouldn''t want this sort of knowledge base in a complex world as they are too restrictive and inference can be extremely complex even in simple worlds. Since a game is a dynamic object, you would need as a minimum the Situation Calculus and then you have the Frame Problem to deal with in all its ugliness.

What I could envisage is a non-monotonic logic - like belief networks - being placed on a card. All processing can be handled in hardware on the card,: inference, querying, value of information, decision theoretic computations, etc. All that would need to be sent back and forth was information about which variables to query, or which observations needed to be added to the network.

The benefit of dynamic belief networks is that they are a general representation of Markov processes so you can handle all sorts of non-linear modelling for dynamic systems. I can envisage coupling a stochastic model of the environment to a physics engine for some very realistic simulations!

Anyway, just ramblings and ideas...

Cool thread, keep it up!

Timkin

I like the idea of dedicated AI boards. Though I did have a good laugh with a friend of mine about the idea of mandatory AI boards for games. With graphics you can turn down the level of detail and remove features to deal with lack of processing power. I can only imagine a little "intelligence" slider for your AI elements.

Humor aside, because true AI is so general it would probably be best to use a generic model for dedicated AI boards as opposed to algorithm specific AI board. The basic problem comes down to the fact that graphics are about as low level as you can get. The world commands the graphics. AI''s command the world. While pathfinding and similarly often used algorithms could be very useful the level of complexity in game world design makes it difficult to generalize the elements in said game world. My point is this: game world design changes dramitically from game to game and I''m afraid that one pathfinding algorithm which works for one game wouldn''t work for another game due to major design differences.

Orion