AI Acceleration?

[beardiebloke]

Has anyone here seen this website - aiseek.com aiseek.com ?

They say they have a dedicated processor for AI - it's *supposed* to speed up things like LOS checks and path finding A LOT.

One thing that I'm not sure of is how a (real time) games designer will cope with players having vastly different AI processing capabilities since it directly affects gameplay - unlike fancy graphics. Turnbased is of course a different story...

What's the chance of CMx2 supporting this when it comes out?

[Cpl Steiner]

Originally posted by beardiebloke:

One thing that I'm not sure of is how a (real time) games designer will cope with players having vastly different AI processing capabilities since it directly affects game-play - unlike fancy graphics.

I personally wouldn't have a problem with different game-play, as long as you had the option to disable the AI processor for comparison purposes. Any owner of a PC equipped with such a processor would presumably welcome a more difficult game-play experience due to better enemy AI - which in a game like CM would be offset by better friendly unit AI, one must not forget.

As graphics processor upgrades for PCs are reaching a kind of plateau now, as so much power is available to developers that little more is needed, AI and physics processors would appear to be a logical next step. The only problem with this logic is one wonders what the humble CPU will actually have left to do other than coordinate all the other dedicated processors. The CPU may end up being like Captain Kirk on the Enterprise, getting feedback from his host of experts such as Spock, McCoy and Scotty and tying it all together!

[Cpl Steiner]

Having just read the whitepaper on the site I'm impressed on how it might be applicable to CM style games. In particular the AI processor is capable of doing many line of sight checks at great speed - something CM would definitely benefit from.

[hellfish]

I think if they could plug this processor onto the Phys-X card you might actually have a product that people would buy. As it is now, I'm not gonna pay $300 for each of them.

[offtaskagain]

I'll second that idea.

[Battlefront.com]

Yeah, it's nothing we're going to support. In theory it sounds good, but until it is standard equipment with a solid API, it might as well not exist

Steve

[Bertram]

I think it is a shame (an a comment on the state of the average game AI) that doing LOS checks and path computing are termed AI. In my opinion these routines are (cpu intensive) basic procedures. With HT and duo/quad cores these processes should be able to be done in parrallel on the regular CPU, in sofar as you still have a regular CPU in a few years.

The AI should kick in after these procedures are done, to formulate strategy and tactics based on the results of these procedures. Writing a good AI is very labor intensive (and needs good Human Intelligence), but the result isn't necesairily cpu intensive...

Compare it to chess programs. The first ones just tried to compute every possible move a few steps deep, and then gave a value to the resulting position. The CPU intensive part was computing all possible moves, the smart programming was coming up with a value for the position that resulted. These programs could be tricked by thinking a few steps deeper, and luring them into traps, where an advantage in terms of pieces was compensated in position (position is difficult to formulate values for, even in chess, and more so in wargames with less ordered battlefields). More modern programs either use more computing power, to evaluate more moves (brute force, cpu heavy but intelligence light) or choose a few promising moves and develop these (and this needs intelligent design to recognize these promising moves).

War games need more of the second type of AI, as the possible permutations are so great that brute force will never give good results. The crux of good game AI design is cutting up/describing the situation in information units on which rules (as in "if ... then ... " statements) can be applied. Of these two the describing of the situation in terms that can be used in the rules is more difficult then actually formulating the rules themself...

(Disclaimer: my recent AI/expert system experience is more related to technical juridical problems. In these the knowledge is easier to describe, so the difficulty of describing the situation in a game might be in my perception).

[tar]

Hear, hear!

I would think that LOS computations would really be the realm of graphics processors. They certainly aren't an AI problem. The AI problem that dominates is the inabiliity of programs to come up with a good overall plan, the inability to recognize plans being executed by the human opponent, and the inability to try to mount deceptive operations. Terrain analysis tends to be another weak point, at least at the level where it helps to inform the actions of the AI as far as identifying good routes of advance and strong defensive positions.

[luderbamsen]

If there is anything that fuels my contempt for the majority of computer game makers these days (present company excluded), it's AI development, or rather lack thereof.

There are numerous game franchises that have developed dramatically in many areas over the last decade (notably graphics, of course), but has seen little or no development in terms of AI. In fact, at times it appears to be getting dumber! And in case you're wondering: No, massive scripting is NOT an acceptable alternative to better AI.

BFC are among the few who seem to understand this, and we should all be grateful for that.

[Redwolf]

What is named "AI" processor here has nothing to do what you would normally call AI.

These "AI processors" are only concered with LOS computation and very simple behavior algorithms like A* search for pathfinding, flocking and the like. Absolutely no planning or anything resembling tactics in involved, much less learning. There is no processing of any kind of predefined knowledge, not even about that units strength's compared to it's opponent.

This kind of "AI" is what you use to implement moves after you make decisions but it doesn't help making decisions. Atrocious decsion making is what makes computed opponents not fun to play after a while.

The graphics card is excellent for LOS computation and it even comes "free". But it only works from the player's viewpoint. In an "independent-view" 3D wargame like CM or ToW the LOS computations never go from or to the player's viewpoint, it is always between two units on the map at arbitrary angles towards the player. You can't use the graphics card for it. OK, you could require a second graphics card and constantly flip it between all units and "look around". Would probably look pretty cool

[Battlefront.com]

Psssst.... wanna know why game AI has never been a top priority of game developers? First of all, because to do it right usually takes as much programming as the entire game itself. Second, significant game AI can only be programmed at the END of the development cycle, which is usually when the project is already over budget in time and money, therefore there is little interest in making things worse. Third, and this is the most important, AI has never ever even once to the best of my knowledge ever been a make or break feature for a game. It sure creates bitch-a-thons galore, but this is from people that have already bought the game. Fourth, and perhaps even more important, since writing a really great AI is nearly impossible, people are always going to bitch and complain about whatever AI actually gets put in the game. So why bother to go through all the effort, expense, misery, and what not just to be bitched at by people that have already bought the game regardless of how good the AI is or isn't? Why? Exactly

Steve

[Vanir Ausf B]

You can't show good AI in a screenshot.

Having said that, GalCiv 2 is one game I can think of that has successfully used strong AI as a main selling point.

[Battlefront.com]

Vanir,

Having said that, GalCiv 2 is one game I can think of that has successfully used strong AI as a main selling point.

The sad thing is games use AI as a selling point whether it is crap or (relatively speaking) good. Yup, game marketers are so cynical that they not only know they don't have to put good AI into their games, but they can also get away with claiming right on the box that their game without decent AI has fantastic AI.

And so the mass market developers get richer...

Steve

[Bruce70]

"And so the mass market developers get richer..."

Sound like a good candidate for differentiation then.

[Redwolf]

Myth II also was a game sold on reasonable AI, and the AI was arguably more fun to play than in most other games.

The problem is that the lesson from those games do not carry over to wargames.

Those fantasy games are heavy asymmetric. You can tune the units available to the computed opponent to their capabilities. Instead of teaching the AI to do what it can with the units at hand, you go the other way round - you look at your "AI" capabilities and then design units and unit balance around it. A pretty dumb approach to attack you can look like a good AI if that approach is the best you can do with the units at hand.

That doesn't mean any kind of decision making was involved.

[ October 02, 2006, 09:31 AM: Message edited by: Redwolf ]