EPTDeepMind sets AlphaGo's sights on SCII - Page 11
| Forum Index > SC2 General |
|
aQuaSC
717 Posts
| ||
|
Endymion
United States3701 Posts
| ||
|
Jono7272
United Kingdom6330 Posts
| ||
|
ZigguratOfUr
Iraq16955 Posts
They posted this blog. I'm happy they mentioned an APM cap. | ||
stuchiu
Fiddler's Green42661 Posts
| ||
|
SCHWARZENEGGER
206 Posts
| ||
|
stuchiu
Fiddler's Green42661 Posts
On November 05 2016 04:21 SCHWARZENEGGER wrote: what if AI will be too smart and refuse to play terran? shit it really is smarter than all of us. | ||
|
Hider
Denmark9442 Posts
On November 05 2016 04:21 SCHWARZENEGGER wrote: what if AI will be too smart and refuse to play terran? What if AI figures out the best way to win is to complain about balance on balance forums in order to manipulate blizzard into patching the game in a certain way. Or what about developing memes such as Gom TvT in order to make sure blizzard doesn't buff terran/nerf zerg for the next 2 years no matter how needed it is. | ||
|
Fran_
United States1024 Posts
| ||
|
ZenithM
France15952 Posts
I think the "incomplete information" part of the game is what's going to be tough. But if they at least give the computer perfect mechanics, it should be feasible to win against most pro-players most of the time. How often will be the measure of this AI project's success. | ||
|
Dumbledore
Sweden725 Posts
On November 05 2016 05:57 ZenithM wrote: It's not the "real time" aspect that's hard, real time for powerful computers is nothing more than a very fast series of turns (I can assure you that's literally how that's going to be implemented ;D). I think the "incomplete information" part of the game is what's going to be tough. But if they at least give the computer perfect mechanics, it should be feasible to win against most pro-players most of the time. How often will be the measure of this AI project's success. Learning algorithms thrive in situations with randomness, so it will probably thrive with unknown information. What makes it difficult is the massive state space that Sc2 has compared to board games. | ||
|
Nebuchad
Switzerland12472 Posts
On November 05 2016 04:28 Hider wrote: Or what about developing memes such as Gom TvT in order to make sure blizzard doesn't buff terran/nerf zerg for the next 2 years no matter how needed it is. Reality: prime developer of memes since 2010 | ||
|
ZenithM
France15952 Posts
On November 05 2016 06:01 Dumbledore wrote: Learning algorithms thrive in situations with randomness, so it will probably thrive with unknown information. What makes it difficult is the massive state space that Sc2 has compared to board games. I think they will rather easily find ways to smartly discretize the search space. What I meant by "incomplete information" is that it may happen that in a certain state of balance of SC2, Build A beats Build B no matter what you do. It would look silly if that AI got eliminated by a couple of cheeses that somehow hard counters its build 2 times in a row. But I think SC2 might be stable enough that there aren't many of these scenarios, if at all. Edit: And in my humble experience (I don't know what's your own experience with machine learning), I wouldn't say with a straight face that "learning algorithms thrive in randomness situations". The less random my test set is, the happier I am :D | ||
|
MyLovelyLurker
France756 Posts
On November 05 2016 06:01 Dumbledore wrote: Learning algorithms thrive in situations with randomness, so it will probably thrive with unknown information. What makes it difficult is the massive state space that Sc2 has compared to board games. Not really, the 'partially observable' part definitely makes reinforcement learning harder. Two bits strike me as really really hard. The infrequent rewards part ( if you define a reward function as per the number of adverse units killed minus yours, for instance ) makes it hard, as this function might not move for the first five minutes of the game, so you don't get that many training examples. The planning aspect as well, since the AI will have to infer series of increasingly complex concepts, from 'click the mouse' to 'fake proxy gate into expanding'. Great that they're outsourcing this, excited to try some algorithms when it's out. | ||
|
MyLovelyLurker
France756 Posts
On November 05 2016 06:11 ZenithM wrote: Edit: And in my humble experience (I don't know what's your own experience with machine learning), I wouldn't say with a straight face that "learning algorithms thrive in randomness situations". The less random my test set is, the happier I am :D Agreed, this is why regularization was invented. He probably meant that gradient descent is robust to initialization & noise. | ||
|
coverpunch
United States2093 Posts
| ||
|
ZigguratOfUr
Iraq16955 Posts
On November 05 2016 07:36 coverpunch wrote: I'm curious - is the native AI in the game limited in APM? No. But it sucks so no one cares. | ||
|
MyLovelyLurker
France756 Posts
On November 05 2016 07:36 coverpunch wrote: I'm curious - is the native AI in the game limited in APM? If Deepmind proceeds as they have in the past for Atari games AI, their SC2 AI would automatically be limited to 2 or 3 actions per frame, so at 60fps, that's max 180APM. | ||
|
oGoZenob
France1503 Posts
On November 05 2016 07:36 coverpunch wrote: I'm curious - is the native AI in the game limited in APM? no but they use their apm differently, as they cant box units, and have to move them individually. So an insane AI, which is about plat/diamond level, is generally playing at more than 1000 apm average | ||
|
jimminy_kriket
Canada5536 Posts
On November 05 2016 07:42 MyLovelyLurker wrote: If Deepmind proceeds as they have in the past for Atari games AI, their SC2 AI would automatically be limited to 2 or 3 actions per frame, so at 60fps, that's max 180APM. wouldnt that be 180 actions per second? That would be cool to see. | ||
| ||
