Flash on DeepMind: "I think I can win" - Page 7
Forum Index > SC2 General |
Draconicfire
Canada2562 Posts
| ||
WinterViewbot420
345 Posts
| ||
chipmonklord17
United States11944 Posts
EDIT: Not saying this is poorly received, but imagine the hype if it was announced Google was getting into esports | ||
rockslave
Brazil318 Posts
On March 11 2016 10:40 Superbanana wrote: Imba Ai goes 3 rax reaper every game no matter what and wins every game Don't say "solved". Chess is not solved, Go is not solved. You're right about that. I should've said "they beat Kasparov without a flying penis" Checkers is solved though. | ||
Jonoman92
United States9091 Posts
| ||
Hypertension
United States802 Posts
| ||
b0lt
United States786 Posts
On March 11 2016 11:08 chipmonklord17 wrote: Hey Google, instead of making an AI to beat a starcraft player, sponsor a starcraft team. It would cost less and probably be better received. EDIT: Not saying this is poorly received, but imagine the hype if it was announced Google was getting into esports And it'd be completely pointless? | ||
beg
991 Posts
On March 11 2016 11:08 chipmonklord17 wrote: Hey Google, instead of making an AI to beat a starcraft player, sponsor a starcraft team. It would cost less and probably be better received. EDIT: Not saying this is poorly received, but imagine the hype if it was announced Google was getting into esports But that's the cool thing about Google... They're not doing things to polish their image, but to innovate. They're pushing the boundaries. Sponsoring a team wouldn't really do that, hm? Sponsoring a team is just for PR. | ||
ZAiNs
United Kingdom6525 Posts
On March 11 2016 10:35 rockslave wrote: Everyone is missing the point (including Flash). Go is already a game with an impossibly big search tree for brute force. Even chess is. The classical approach of heuristics coupled with brute force solved chess, but it was never even Platinum in Go. The only reason for AIs starting to beat Go players is a somewhat recent innovation in AI: deep learning. From 10 years ago or so, there were several advancements to machine learning that made a gigantic leap in many fields for which computers always sucked. For instance: character recognition used to be a PitA, but nowadays you can write Python code that gets it right 99% of the time in a few minutes (the breakthrough was a particular optimization technique called backpropagation). Even if you cap micro a lot, StarCraft isn't too much different from a combination of Go and a bunch of pattern recognition. That is precisely what machine learning solves. It's not easy though, there is a lot of clever training and parametrization to be done... But if they put it in their roadmap (with enough money), it will happen. Oh, and imperfect information is not a problem at all. Even with a more standard (backtracking / brute force) approach, you only need to throw some probabilities around. It's rather easy to write programs that play Poker well, for instance (discount the poker face though). Deep learning needs a dataset for the AI to be trained though. For AlphaGo they trained two separate networks (one designed to predict the next move, and the other designed to predict the final winner) on 30 million discrete moves from games played by human experts. After that it trained itself by actually playing Go against itself a ridiculous number of times. A Go game can be perfectly modelled by simple list of positions describing which square had a stone placed on it each turn, it's going to be very hard to get enough useful data (replays) to significantly help with the training. And without the initial training it's going to have to learn mostly by playing against itself which will be difficult because of the ridiculous number of game states. At least that's my understanding of things, I could be wrong, but it seems to be a lot harder than Go. | ||
evilfatsh1t
Australia8526 Posts
gives me chills just thinking about that possibility. that said though, i dont know how deepmind is programmed enough to comment on its ability but i do know that go is at its roots a game that could in theory be solved by maths. the only advantage pros had over ai in past years was there was no ai that could calculate every single possible move until recently. im not sure if this is how deepmind works now, but if the ai is able to calculate every single variable in a game that follows mathematical rules then a human shouldnt be able to win. starcraft however doesnt follow these rules so i dont see ai being able to defeat the decision making of a pro for a long time | ||
beg
991 Posts
| ||
BronzeKnee
United States5208 Posts
On March 11 2016 10:35 rockslave wrote: Everyone is missing the point (including Flash). Go is already a game with an impossibly big search tree for brute force. Even chess is. The classical approach of heuristics coupled with brute force solved chess, but it was never even Platinum in Go. The only reason for AIs starting to beat Go players is a somewhat recent innovation in AI: deep learning. From 10 years ago or so, there were several advancements to machine learning that made a gigantic leap in many fields for which computers always sucked. For instance: character recognition used to be a PitA, but nowadays you can write Python code that gets it right 99% of the time in a few minutes (the breakthrough was a particular optimization technique called backpropagation). Even if you cap micro a lot, StarCraft isn't too much different from a combination of Go and a bunch of pattern recognition. That is precisely what machine learning solves. It's not easy though, there is a lot of clever training and parametrization to be done... But if they put it in their roadmap (with enough money), it will happen. Oh, and imperfect information is not a problem at all. Even with a more standard (backtracking / brute force) approach, you only need to throw some probabilities around. It's rather easy to write programs that play Poker well, for instance (discount the poker face though). The thing about Sc2 though is that it is different. In Poker, or Go or Chess, when you move, you move. That's it. And a computer can process that. SC2 is different. If I load up a drop and sit it outside your base, I don't have to drop. But I might. But the dropship might actually be empty. What do you do? What does the AI do? I might show extreme aggression, but be taking a hidden expansion. I could also show an expansion, but then cancel it or not make it and attack. Unless the computer wins with perfect micro and macro, I think it would struggle against non-traditional builds, timing attacks and mind games. | ||
Wrath
3174 Posts
2. The APM most likely will be restricted to around 200. AI's APM is equal to its EPM, it does not waste clicks like progamers and those who spam boxing or clicking to increase their APM. So for guys like EffOrt who can go to around 450 ~ 500 APM, what is the actual EPM of them? Does it go beyond 200? That is what we need to consider for AI. | ||
CursOr
United States6335 Posts
I would love to see an AI that dropped in different places, tried to deceive opponents, did real different build orders, and played map specific strategies, just as a person would. | ||
ETisME
12083 Posts
We may even see a whole new meta developing | ||
ZAiNs
United Kingdom6525 Posts
On March 11 2016 14:47 beg wrote: @ZAiNs: Aren't there many BW replays? Also, DeepMind is capable of learning from reading the graphics, so they could try using VoDs too. AlphaGo was fed 30 million moves and apparently the average number of moves per game is 200, meaning they gave it around 150,000 high-level games. Getting that number of BW games is impossible, and even if it were, I'm quite sure you'd need drastically more replays to get training results on par with AlphaGo's initial training set. I don't think VODs would even be useful because they show such little information about the game state at any point in time, I think a replay is needed so it can observe the entire game state at every point in time. | ||
Grumbels
Netherlands7028 Posts
On March 11 2016 15:09 ZAiNs wrote: AlphaGo was fed 30 million moves and apparently the average number of moves per game is 200, meaning they gave it around 150,000 high-level games. Getting that number of BW games is impossible, and even if it were, I'm quite sure you'd need drastically more replays to get training results on par with AlphaGo's initial training set. I don't think VODs would even be useful because they show such little information about the game state at any point in time, I think a replay is needed so it can observe the entire game state at every point in time. It would be nice if wherever Koreans play BW would automatically save the replay, scramble the names, and send it off to google. Or imagine people at google becoming frustrated because for once they do not have big data sets available for everything. | ||
lpunatic
235 Posts
On March 11 2016 15:09 ZAiNs wrote: AlphaGo was fed 30 million moves and apparently the average number of moves per game is 200, meaning they gave it around 150,000 high-level games. Getting that number of BW games is impossible, and even if it were, I'm quite sure you'd need drastically more replays to get training results on par with AlphaGo's initial training set. I don't think VODs would even be useful because they show such little information about the game state at any point in time, I think a replay is needed so it can observe the entire game state at every point in time. AlphaGo got off the ground with a big bank of games, but recently it's been improving purely through self-play. I think if the DeepMind team put their effort into BW, they'll be able to achieve superhuman performance in a few years time. There are some ways that the problem is harder than Go - partial information, real time and a much more complex raw game state. On the other hand, there are some clear advantages an AI will have over people (APM, multitasking) which are not present in Go. It seems to me that if you can get an AI that makes decisions like a half decent human player, it will be able to press its advantages well beyond human competition. | ||
lpunatic
235 Posts
On March 11 2016 13:01 ZAiNs wrote: Deep learning needs a dataset for the AI to be trained though. For AlphaGo they trained two separate networks (one designed to predict the next move, and the other designed to predict the final winner) on 30 million discrete moves from games played by human experts. After that it trained itself by actually playing Go against itself a ridiculous number of times. A Go game can be perfectly modelled by simple list of positions describing which square had a stone placed on it each turn, it's going to be very hard to get enough useful data (replays) to significantly help with the training. And without the initial training it's going to have to learn mostly by playing against itself which will be difficult because of the ridiculous number of game states. At least that's my understanding of things, I could be wrong, but it seems to be a lot harder than Go. On the other hand, evaluating a stone in Go is a very hard problem - it may depend on the position of every other stone on the board. For starcraft, the value of a base or a zealot is pretty simple to evaulate in comparison, and while zealots in a good position are better than zealots in a bad position, the positional relationships aren't anywhere near as complex as in Go. Point being, you maybe can get away with a simplified game state representation. | ||
Gluon
Netherlands338 Posts
On March 11 2016 15:02 ETisME wrote: Actually it makes me wonder what would two deepmind do if they were to play against each other. We may even see a whole new meta developing Exactly this. With the way the AI learns, the most interesting development will be in the fact that it will not be constrained to any conventional build orders. It could semi-randomly develop completely new builds for specific match-ups on specific maps. I'm really looking forward to that. Other than that, Deepmind should eventually win with stellar macro and micro, just by going 3 rax every game | ||
| ||