computer - AI
Forum Index > SC2 General |
no.1
516 Posts
| ||
Empyrean
16926 Posts
| ||
Eggy-X17
United States133 Posts
Logistically, it would be very costly in resources. | ||
jtan
Sweden5891 Posts
Maybe they include some free ai-editing in the editor though, that wouldn't be a hard thing to do. | ||
no.1
516 Posts
| ||
LosingID8
CA10824 Posts
| ||
Eggy-X17
United States133 Posts
The best they could do in the next 5-25 years would be a single profile on someone's computer that learned only against 1 opponent. We're not talking hardware advancement here. We're talking software advancement. And I don't see software getting that good for a while. Although, if I think about it, I'm sure there are ways to optimize this. Maybe the AI can organize players into certain types of 'styles'. And the AI's could self spawn and learn as iterations, like software revisions against different styles. Hmmmm... /me wheels in head turning... | ||
davidgurt
United States1355 Posts
On May 25 2007 13:16 Eggy-X17 wrote: It could only be done 1on1 if they decided to implement it. You couldn't make online profiles composed of multiple players. IE profile x contains players 1 2 3 and profile x is designed to adapt and get better against teams composed of player 1 2 and 3. Imagine the amount of nodes involved in such an arrangement. It could be as high as N^k or as low as N assuming every one played singles, which wouldn't happen. It'd be exponential in nature if more people played with one another. And we're not even counting mixed teams of AI's and humans. The best they could do in the next 5-25 years would be a single profile on someone's computer that learned only against 1 opponent. We're not talking hardware advancement here. We're talking software advancement. And I don't see software getting that good for a while. Agreed. There's no way Blizzard could implement something that advanced and still keep inside the resource budget. They can, however, make the AI much better than it currently is. | ||
CharlieMurphy
United States22895 Posts
On May 25 2007 13:15 LosingID8 wrote: they should just make the computers have easy, medium, and hard settings (kind of like age of empires series). Computers have insane micro speed even though they may not use it properly. Computer settings on WAR 3 are newb easy medium hard and insane AFAIK. Plus computers have maphacks | ||
zdd
1463 Posts
| ||
jtan
Sweden5891 Posts
Imagine 10 saviors controlling 5 clumped mutas each :D edit:wouldn't be fun to play against though | ||
DrainX
Sweden3187 Posts
| ||
Jyvblamo
Canada13788 Posts
On May 25 2007 13:25 jtan wrote: Wouldn't be impossible to make a pretty much unbeatable computer opponent though. Take sc1 zvt for example, make it do some pretty safe 3-hatch muta build, then implement a badass algorithm for muta micro and make it have like 5 clumped mutas/group and kill marines and shit all over the place at the same time while constantly adding mutas. Imagine 10 saviors controlling 5 clumped mutas each :D edit:wouldn't be fun to play against though That would be boring. | ||
Eggy-X17
United States133 Posts
On May 25 2007 13:25 jtan wrote: Wouldn't be impossible to make a pretty much unbeatable computer opponent though. Take sc1 zvt for example, make it do some pretty safe 3-hatch muta build, then implement a badass algorithm for muta micro and make it have like 5 clumped mutas/group and kill marines and shit all over the place at the same time while constantly adding mutas. Imagine 10 saviors controlling 5 clumped mutas each :D edit:wouldn't be fun to play against though I'm all for stuff like that. Maybe, someday. | ||
Judicator
United States7268 Posts
While the example I provided is very different from the one you are thinking of and from an entirely game/map, the point is lets hope they offer an open/customizable AI where the community can really modify it to suit their purposes. | ||
Pika Chu
Romania2510 Posts
On May 25 2007 13:13 no.1 wrote: just imagine how awesome this would be.. computers x times faster than savior, nada and the other pro's to train with And then, we will obs comp vs comp play all day long, no need for programers and vods, not even replays, HURRAY! | ||
AC3
Canada337 Posts
I thought that this was cool. | ||
jtan
Sweden5891 Posts
On May 25 2007 13:57 AC3 wrote: Hey guys, wasn't there a programmer that played Boxer in WCG a couple years ago who had programmed his BW AI to be like all of the famous programmers. He had it execute the timing correctly and would vary through different of that programmers strats, scout his base and react, etc. I thought that this was cool. blackman, but he denied it himself and said that it was just a rumour. | ||
Haemonculus
United States6980 Posts
| ||
IIICodeIIIIIII
China1101 Posts
Normally, AI needs to see things in real life so it's hard for it to recognize things, like people vs desk. But since bw is all in sprites, it'll be easy to program the AI to recognize that this is a hydra and that is a marine or whatever. For the AI to work, it would probably be designed like this. Huge data matrix that simulates the Starcraft game - data matrix is continuously updated via observations - A virtual data matrix is continuously updated for both short term and long term projections (short term = micro, long term = build orders by the book, and macro, and timing, and etc) - The data matrix and virtual data matrix should be updated at least once every .1 seconds because that's kind of the pace that the game goes at. - The AI executes orders via keyboard and mouse. * note: to observe the game, I was thinking about inputing data directly from the computer screen into the AI. It's something similar to capture screen, store it in a bmp, analyze with AI, and repeat 10 times per second. * note: to have AI execute orders, I was thinking about the same way a key logger works. A key logger stores all the keys, then reexecutes them the same way it was recorded. The computer can't tell the difference whether you are typing the keys, or the keylogger is. The AI can just directly interact with the computer like a person does when observing, analyzing, executing. * data matrix is easy to implement. it's just a lot of code and processing, but it's not hard. * updating the virtual data matrix wouldn't be hard. The main difficulty here is finding a computer that can do enough analyses per second, or the code can be written so the AI only analyzes stuff when it's important to do so. * actually analyzing the virtual data matrix would be the hardest part. This is the actual decision making part. In a beta, it can just be a set build order, and some very easy micro things like stim marines, retreat hurt marines, etc. In some later versions, people may start adding stuff in like the AI creating custom build orders based on what the enemy is doing, and things like timing. * learning from mistakes is very hard to do for computers because mostly, AI is simulating actual intelligence with coding shortcuts to keep things efficient. In people, we are able to learn from past mistakes because we are able to think hypothetically, and relive things just in our mind without having to do them in reality. In our minds, we do things, make things work, and then do them in reality. for an AI to do this, the AI would actually have to save the replay of the game the AI messed up in, and keep replaying that game in simulation thousands of times while only saving the games that the AI did good in. Then in a real game, the AI will only do the things that worked best in the 1000 simulation games. ... When computers get fast enough, this would be more efficient, to have the computer do simulation games and only retain the style of play that wins the most games (a lot of chess computers do this already), but until then, it's best to just have a computer programmer edit the code. | ||
| ||