Wouldn't it be interesting if the computers had a difficulty setting. Ex: Easy, Mid, Hard, Very Hard. Were the setting Very Hard is Real Pro level (Koreans). With today's tech it should be easy to write a supper strong AI.
Are you kidding me? How many people would like to play with a crazy super strong AI? And if you are that hardcore you could write your own AI just like IdrA did (I think it's him). And anyways, you play online, not with an AI. I'm sure that the insane AI in Starcraft 2 with it's 2x resource gathering is hard enough.
Currently in starcraft 2, there are 5 AI difficulty levels, Beginner, Easy, Normal, Hard, Insane. It has been stated that all of the difficulties must scout to see you and that the Insane difficulty mines more minerals per round then you, making it take a incredible amount of skill to defeat. In regards to the "Korean level" of AI, it is impossible to mimic exactly how a human will react to any situation so thus it is impossible to code a machine that will act like a "Korean" that and each Korean pro player has there own unique style and tactics that would either be impossible or take a very long time to mimic with code. Hope this was useful to you, now back to my vacation -Seradin
You could make a map hacking, super AI I would think. Imagine, not a single Idle worker.. when units in big groups attack, each unit has it's own micro, timing is dead on, building placements, everything is calculated to the point where not a single mistake is made. Every possible move is simulated in advance generating millions of possibilities in a second. And from that select a move to counter all the possibilities to loose and select a path to win no matter what. Sure, the AI would loose. But only if you put it in unfair map situations. Like no minerals at AI spawn location.
I've watched Browder in one of his recent interviews commenting that AI at insane "knows everything" which means that it doesn't need to explore.. Just like BW..
RTS is a complex game in terms of choices available at any one time. Strategy goes much deeper than considering 1 movement ahead. If we had an AI in a chess game that only thought one move ahead, you'd easily win it. So you have to make it think deeper than that. How many moves ahead? 5 minutes worth? 10 minutes worth? Of course you don't have to anticipate micro skills that far ahead, maybe 10 seconds ahead is enough there. Recon is less critical than micro, so the clock can tick slower for considering updates to that. The economy and macro would need thinking a long way ahead, like 10 minutes or so. But the clock can tick quite slow for that. Maybe 1 consideration every 15 seconds would be enough. I think it is very possible to do what you are asking. But AI is hard to develop, requiring very intelligent people, and it isn't a big priority considering that most players are happy competing with their peers online. Also a complex analytic AI could take a big chunk of the cpu, and games designers prefer to put more emphasis on good graphics and supporting big numbers of units. AI isn't eye candy. It doesn't sell.
hey, you know what else is a good idea? if we had a THIRD RACE! oh wait, the original starcraft had that already. seriously, why do people keep making "suggestions" that the original game from 10 years ago already has?
jasmine, I've already said that. You said "5 minutes or 10" moves ahead. But I meant that it should calculate all the outcomes of the game (win or loss) based on the situation and then adapt. So each milisec. the AI would process every possible step ahead and predict weather the game will be a win or not, and then choose the path to a win.
The point I'm making is that it's a complex game. Lots more complex than chess. A chess computer CANT think 25 moves ahead -- there would just be ridiculous amounts of information that couldn't be sifted through in any reasonable time. A computer AI which thinks that deeply would run out of time considering even its first move. It would lose because it was too slow in making its moves. In RTS, the results have to come in before the next clock tick, otherwise you're predicting a game state that has already become history. The same is true of weather forecasting. You can run a really detailed simulation of the weather to predict what it will be like at 9.00 tomorrow morning. But if it takes your computer 2 weeks to run that simulation, then the result becomes worthless before the calculation is even completed. Computation time multiplies up (at least doubles) with each step into the future you consider. And with RTS, the limit would occur much earlier than in chess because the game has a higher computational complexity. No supercomputer in existence today or the foreseeable future could do anything close to what you are asking. But theoretically, yes it can be imagined. What I proposed was a compromise that could just about be done on today's hardware if someone was clever enough to program it. By splitting the game into different layers, running each on a different clock and anticipating each one a computable distance into the future. Some aspects of the game are time critical but there is negligible benefit to anticipating them far into the future, like micro skills. Some aspects require very long term planning, like build orders, and considering expansions, but these can be considered at a much lower temporal resolution.
A computer from the 1980's could process thousands of calculations just to keep your computer screen up. Your thinking of a human mind not a computer. A computer can do billions of calculations a second, and I am talking the low cost computers. Some can even do trillions of calculations a second, so I don't think a computer planning twenty five moves ahead is impossible neither is it to expensive for us to imagine in a video game. As for planning the movements in starcraft I think we could have it plan out a strategy in such a way as lobster was describing.
Please familiarise yourself with computational complexity and intractable algorithms. If we're planning on thinking 25 moves deep -- which is nothing in an RTS game. And if in each move you can give only 1 of 10 possible orders -- which is a gross underestimate if we're talking precision orders of scores of units. Then you'll be having to sift through 10000000000000000000000000 game states before giving your first order. Trilions of calculations per second is nothing. If you could process 1 trillion game states per second... your algorithm would take 10000000000000 seconds.
Let me give you something to think about. Chess cant be solved, it would take the fastest computer in the world over a billion years to fully calculate every possible out come in a chess game. Heres something to consider. Chess is infinite: There are 400 different positions after each player makes one move. There are 72,084 positions after two moves. There are 9+ million positions after the first three moves. There are 288+ billion different possible positions after four moves. There are more 40-move games than the number of electrons in our universe. As an experienced chess player that played in hundreds of tournament i can reacher you that chess or starcraft for that matter, cant be solved. But witting an incredibly strong AI is possible, It would need a sophisticated in dept micro managing database.
I'd assume this topic concerns more the original StarCraft, since the ideal metagame strategies for SC2 will probably take at least a couple of months before people reach a consensus. I'm not too familiar with exactly how the original StarCraft's AI is programed, but this is actually a really interesting topic because I think a Korean-level AI (that doesn't cheat) is probably very possible nowadays. Most games have really come down to execution rather than creative strategies, and obviously an AI would be able to multi-task far better than any Human. In terms of Chess. I think one of the big differences here is that APM / uber micro alone can win you matches in StarCraft. Even with an improper build, uber micro with a stack of Mutas or Vultures can win games. I recall Relic said so themselves that they introduced a margin of error on purpose for the Dawn of War 2 AI so it's more fun to play against. Blizzard will probably do the same for SC2.
it would be impossible to create an AI in game that would work and be as strong as one of us because AI rely on triggers, and do not actually think the only thing they could do is make more advanced triggers they will never have the instinct or strategizing of a human player
Triggers are how it is is done now. Triggers are a quick and dirty way of making an AI. The trouble with anything scripted is it's purely reactionary. There's no strategic analysis. No planning and contingency. That's where it fails. I did outline a method above for how to do AI without triggers that is analytic. It does take more cpu time though, and as I said: it's complicated to make, and doesn't sell more games. AI isn't eye candy.
