i don't think it's possible to build an AI in the near future that can beat a human with human level micro because starcraft isn't a game with discrete game states like chess or go, and even chess isn't strongly solved yet, and chess has far fewer possible positions than go, let alone a real time game
"Deep Blue" for SC2? - Page 2
Forum Index > SC2 General |
rauk
United States2228 Posts
i don't think it's possible to build an AI in the near future that can beat a human with human level micro because starcraft isn't a game with discrete game states like chess or go, and even chess isn't strongly solved yet, and chess has far fewer possible positions than go, let alone a real time game | ||
Bwaaaa
Australia969 Posts
On November 11 2012 14:45 EtherealDeath wrote: So little faith ~ Go is probably similar. Sure is hard as hell to program an AI for it, but we've got a 6dan AI so far. But in go/chess you only ever play on one map and you have 100% map vision. It would be so much easier to write an AI which could beat every human if you programed it for a only one simple map and gave it map hacks but there are so many variables in a game of starcraft that come from scouting that it would be near impossible to do. Also for chess there is a set amount of moves you can do at each point in the game that will be exactly the same no matter how many times you play it and whether you are vs a computer or a human(chess has 18 moves at move one and those moves are not affected by time like starcraft). In starcraft due to the abundance of maps and huge amounts of build-order variations a computer would have to account for so many variables that I doubt it could be done. Anyway if it could be done it would be impossible for zerg due to how the larva management works and how players need to use logic and common sense to decide whether to build drones or army. The decision between toss and terran is a hard one. On one hand toss is pretty deathbally and an AI could have perfect FF/storm/Feedback/Pheonix control but terran has the extremely microable MMM ball and ghost control?(pre-snipe nerf this would have been insane). Edit: I am sure you could build an AI who can win 100% TVZ and TVP vs standard openings by proxy 2 raxing everygame but what hapens if the zerg 6pools or the protoss proxy 2 gates? Now the AI can be beat by someone in gold league because it would not have the flexibility to play a standard game. So to all the people saying just to program it for a super cheesy win before the 4th minute build consider that those build have super hard counters that if you know they are coming(know the AI has one build) no amount of perfect micro will save the AI. | ||
dcemuser
United States3248 Posts
If somebody picked up a team of talented programmers and threw a million dollars and 2 years time at the problem, they could probably make a Grand Master level AI. Not one that would always win versus the top in the world, but it could probably go 50/50. | ||
mishimaBeef
Canada2259 Posts
just train the network on like a month of data from a progamer | ||
ScaSully
United States488 Posts
| ||
Probe1
United States17920 Posts
| ||
To3-Knee
Canada100 Posts
Of course, this was 2 years ago and I'm sure the research has advanced. I doubt very far however. Any decent planner for RTS is likely to be rocked by any diamond+ player. | ||
Ragnarork
France9034 Posts
This has been done for Brood War, and may not work for SC2 (their choices wouldn't really work, but the way they proceeded to chose may...), but still gives an idea of how far we've already been with these kind of challenges : (Warning, pretty long read) http://arstechnica.com/gaming/2011/01/skynet-meets-the-swarm-how-the-berkeley-overmind-won-the-2010-starcraft-ai-competition/ And there might have been better AI since, but not afaik. | ||
Prillan
Sweden350 Posts
How do you do that? The biggest problem though is that SC is a game of imperfect information. That makes it really difficult (I'm talking insanely difficult, borderline impossible) to create an AI with optimal play. You can't play optimal if you don't know what the optimal play is. The most practical thing IMO is to use machine learning and feed it tons of strategies. But then it wont always play optimal. | ||
Tobberoth
Sweden6375 Posts
| ||
vorxaw
Canada245 Posts
check out the vods http://eis.ucsc.edu/StarCraftAICompetition | ||
ElMeanYo
United States1032 Posts
However, it would take the most powerful computer ever made and millions of man-hours in programming and testing. Getting it to work on your average Joe's PC without horribly lagging the game is another matter entirely. | ||
SupLilSon
Malaysia4123 Posts
| ||
Doomwish
438 Posts
AI's base and most AIs will just ram up against the FF trying to get back in the base. Or as Terran if you know its just doing the same timing every game camp your choke with stim bio and tanks. Perfect macro/micro don't really matter when you can't think strategically. This game is about destroying the enemy buildings not winning engagements. In base trade type scenarios AI's are so far extremely dumb. People in this thread are just thinking in terms where u line up and have a straight up fight , it is so much more complex than that with army movement around the map and such. | ||
Incognoto
France10234 Posts
that's a really cool picture. and calvinball just made it perfect | ||
Derrida
2885 Posts
On November 12 2012 05:42 Ragnarork wrote: You can do some pretty neat things already. This has been done for Brood War, and may not work for SC2 (their choices wouldn't really work, but the way they proceeded to chose may...), but still gives an idea of how far we've already been with these kind of challenges : (Warning, pretty long read) http://arstechnica.com/gaming/2011/01/skynet-meets-the-swarm-how-the-berkeley-overmind-won-the-2010-starcraft-ai-competition/ And there might have been better AI since, but not afaik. holy shit, what a great article... very inspiring. | ||
JKM
Denmark419 Posts
That said, I think you could program an AI to perform some rather imba early game attacks using perfect marine micro and similar, but it would only work once against most players meaning it couldn't win a BoX series. | ||
AnachronisticAnarchy
United States2957 Posts
| ||
Enearde
France265 Posts
| ||
Scholera
United States166 Posts
On November 11 2012 12:56 Warlock40 wrote: This. I could see Automaton 2000 having a ludicrously high win rate within the first ten minutes of the game by executing a marine timing attack with impossible micro. no, you could always fuck with it. For example: if you have a group of lings with higher mobility than their marines and threaten to counter attack, will they push to your base or try to defend? | ||
| ||