We know strategy game AI is awefully stupid. It trickles units into our base, letting them get mulched one by one. It can''t understand that it should be building air units instead of ground defenses. It tries the same thing, over and over, no matter what has worked. Now, in all fairness, AI is so lacking because resources and processor power are scarce. In fact, according to a Game Developer article I read (State of the Industry AI, or something) it''s only been recently that AI''s gotten a full time programmer in most strategy games!!! But I have a (potentiall very sucky ) idea: AI is fast, accurate, but dumb as a brick. What about taking advantage of this? Two senarios I can see: One is that there''s an NPC side the player doesn''t get to play in a strategy game. Call them the Automateds. It''s the Automateds vs. the Humans. You can play different human sides, but you never play the self-aware Automateds. The Automateds are billed as stronger, faster, and completely accurate. All of their abilities take advantage of their traits. Example: They can target perfectly, predict your vector / position, coordinate attack multiple quickly, etc. The humans, on the other hand, are divided, slower, more clumsy, but a lot brighter, and all their abilties reflect that (no quick click units, or heavy micromanaging of resources, energy, etc.) The fight''s not exactly an even match, but, more like an RPG, you don''t expect it to be (plus you could have side deals w/ the AI, inter-human conflict and backstabbing, etc.) Other senario: Every side comes loaded with it''s share of AI usuable and player usable units / functionality. I don''t like this as much, because it implies an even contest and is sort of the way Starcraft is now. The difference would be in magnitude. The AI would be coded to make heavy use of quick moving, quick clicking units / features. But this seems a lot more unfair, and I''m not sure how the player could be persuaded to think it''s kosher. -------------------- Just waiting for the mothership...

isn''t that more or less just extending what''s there now? the AI (in most strategy games i''ve played, at least), makes up (or tries to make up) for a total lack of ingenuity with better accuracy, and especially short build times and more resources (or lower resource requirements)..
i generally consider that a cop out on the AI, and ideal AI would be the player''s equal with the same conditions, but as you''ve said, AI just doesn''t get the processor time to do that..
i like the automateds idea, because it would be honest about what it is, and i think alot of fun to play..

and A. Buza, for the link (in your sig i guess), you need to prefix the ''members.zoom.com'' with ''http://'', otherwise the dumb browser will try to find it at http://www.gamedev.net/forums/members.zoom.com...''

Hmmmm... Star Craft has a nice and challenging AI (ok, not perfect) and runs nice on a hypothetical Athlon with 200 mhz. I have a 700 and a 800 Athlon, this would make up to 600Mhz for the AI and I also have a bout 450MB RAM for this AI. 450MB and 600Mhz sounds enough for a realtime neuronal network with genetic algorithms and fuzzy logic. We shouldn''t target a robot like AI, we should try to get developers to use such features. There''s an algorithm that is based on how ants do their pathfinding. This genetic algorithm is blazing fast, ultra simple to code and so advanced that it could be used to savely route an unarmed vehicle trough the map while avoiding all threats. Every unit could be powered by a neuronal network and the whole class of ressource / attack / overall strategie problems could be solved by a big neuronal network that gets its wrong / right decisions by simple genetic algorithm. All this wouldn''t be to hard after all, but it would take some ressources. But be sure, this AI will beat every human player, not the slightest chance....

But 3D gfx and eye candy sells better than good AI so developers don''t care...


Tim

--------------------------
glvelocity.gamedev.net
www.gamedev.net/hosted/glvelocity

quote: Original post by tcs

All this wouldn't be to hard after all, but it would take some ressources. But be sure, this AI will beat every human player, not the slightest chance....

No offense to you, tcs-- maybe you've worked on this stuff. But __EVERYONE__ I've ever talked to or heard from... game AI coders in the biz, GDC roundtables, friggin brain splitting mathematics weilding comp sci Phds... will tell you that AI is a __DAMN HARD__ problem to solve.

Genetic algorithms, AFAIK, right now can only solve simple games in the best of cases. They can't even handle Chess.

NNs suffer from nasty I/O problems, again AFAIK. Getting useful input and output is very tough the wider and wider your domain gets.

Now, I'm NOT an AI programmer, but here's what (little) I know: You've got pathing, terrain analysis, tactics, and overall strategy. Pathing and terrain analysis will get better as the processor cycles improve. That's because the map is a much smaller problem domain. But the latter two (and probably anything else I've forgotten) are a royal b*tch. You've got a myriad of interactions between units, timing, pattern recognition, anticipation and future prediction... a lot of stuff we take for granted.

For example, take Starcraft: When should you expand your base? Well, that depends on your situation, and what your overall strategy is. What's your situation? Well, you'd have to have a mental model of the entire battlefield to assess that... and that model isn't just the current state of things that a cheating AI could read. It includes not only ***HOW THINGS ARE*** but the likely way things are ***GOING TO BE***. Now, how exactly are things going to be? Well, that would require some representation of predicting not only how things interact, but how they're likely to interact based on a TON of permutations.

Trust me, just look at Chess!!!! The interactions are much more predictable and easier to compute than your average RTS, and yet Gary Kasparov can ****STILL**** beat Deep Blue what, half the time? Especially, as I've heard, if he lures the AI into traps.

Believe me, I'd love to see more human like opponents, but with the technical hurdles that exist, I'm not sure that's going to be any time soon... which is why I propose playing up the AI's weaknesses as strengths!



--------------------
Just waiting for the mothership...

Edited by - Wavinator on September 27, 2000 7:32:36 PM

Correct me if I''m wrong, but didn''t Deep Blue beat Kasparov a few times?

---------------------------
"Until next time, boobie. Push the button, Frank."

Yep, Deep Blue won four matches to three, I believe. And IBM''s newest supercomputer project makes Deep Blue look like a kid''s toy...

-Ironblayde
Aeon Software

Chess is computationally an "easy" game, actually... it''s entirely deterministic, since you can build a tree of all the possible combinations of play. The difficulty for current computing lies within the depth and breadth of that tree, and IBM has proven that super-computers are only NOW able to generate that tree to a depth that can actually rival a human player. Quite a feat if you ask me!


People might not remember what you said, or what you did, but they will always remember how you made them feel.
~ (V)^|) |<é!t|-| ~