EPTDeepMind/AlphaStar Discussion Thread
| Forum Index > Closed |
|
jy_9876543210
265 Posts
| ||
Nakajin
Canada8989 Posts
On the APM argument I don't know where too stand, sliding the apm up or down won't change the fact that it should always be perfect so it's kinda weird. But the standard camera is good step to make it fair. | ||
|
swissman777
1106 Posts
| ||
|
jy_9876543210
265 Posts
On January 25 2019 05:41 Nakajin wrote: That some Buzzfeed thread name, you should probably rename it too be more clear. (if you really think it need it's own separate thread) Well, I think it's important for them to notice that the AI micro is quite inhuman, but yeah that title was not clear. Sadly I can't edit it after posting it. Hopefully some mod can find a good title? Please fix it, thanks. | ||
|
TBO
Germany1350 Posts
| ||
|
Nakajin
Canada8989 Posts
On January 25 2019 05:46 jy_9876543210 wrote: Well, I think it's important for them to notice that the AI micro is quite inhuman, but yeah that title was not clear. Sadly I can't edit it after posting it. Hopefully some mod can find a good title? Please fix it, thanks. Fair enough | ||
|
swissman777
1106 Posts
On January 25 2019 05:46 jy_9876543210 wrote: Well, I think it's important for them to notice that the AI micro is quite inhuman, but yeah that title was not clear. Sadly I can't edit it after posting it. Hopefully some mod can find a good title? Please fix it, thanks. I think we have to accept that AI micro will be amazing, given that'll be the only thing that AIs would be good at theoretically. However, I think we need to carefully tone how amazing AI micro would be, since the whole point of this project is to mimic human creativity and adaptability. | ||
|
frontliner2
Netherlands844 Posts
| ||
|
Charoisaur
Germany15979 Posts
On January 25 2019 05:49 TBO wrote: their apm graph they had where the compared it to Darios APM was also a bit misleading because of the way the "steal from control group and assign to new control group" action he extensively uses messes up the apm counter. how does it mess up the apm counter? | ||
|
jy_9876543210
265 Posts
On January 25 2019 06:04 frontliner2 wrote: nice title... I think the fact that an AI can compute and counter strategies at this level is actually commendable Yes, it already showed some huge improvements in understanding the economy system (which may benefit human players) and attacking skills comparing with AI before, considering that it learnt sc2 from scratch instead of scripts written by human, but it's not enough. When base was under attack, it retreated the entire army back home which caused it to lose the live game vs Mana. If it doesn't have such inhuman micro, maybe the progamer can win it much more easily, not like a 0-10 score. We just hope to see that next time it plays with a progamer, it will win with strategies, not inhuman micros. | ||
|
Rowrin
United States280 Posts
For instance, the AI goes blink stalkers very consistently (probably because of the high value they get when paired with increased ability to micro). Knowing that the AI considers this strategy "optimal" or at least favored, we can build a counter strategy etc. It's decision making against the warp prism harass was another weakness that displayed reproducible decision making that MaNa could and did abuse (at least in that game, who knows if it'll learn to adapt to that strategy after a few more hundred years of simulated learning). | ||
|
brickrd
United States4894 Posts
| ||
Waxangel
United States33432 Posts
On January 25 2019 05:49 TBO wrote: their apm graph they had where the compared it to Darios APM was also a bit misleading because of the way the "steal from control group and assign to new control group" action he extensively uses messes up the apm counter. that's actually pretty interesting to hear so you'd say MaNa's max of 700APM is more realistic? ![[image loading]](https://www.teamliquid.net/staff/Waxangel/tea.png) | ||
|
Charoisaur
Germany15979 Posts
On January 25 2019 06:39 Waxangel wrote: that's actually pretty interesting to hear so you'd say MaNa's max of 700APM is more realistic? pretty sure that's only when Mana uses rapidfire warp-in. I don't think more than 500, maybe 600 "real" peak apm is really realistic | ||
|
jy_9876543210
265 Posts
On January 25 2019 06:37 brickrd wrote: actually, SC2 fans should understand a thing: the purpose of deepmind entering starcraft is not to provide us entertaining bots to satisfy our intellectual curiosities about the nature of competition in RTS. it's to promote and showcase machine learning. they're doing something historically and technologically significant whereas we are playing games. we should be thanking them for taking any interest at all and for giving SC2 visibility rather than criticizing people who are professional researchers working in their own damn field. Whatever, I can make a program that showcase "AI can beat human in Counter-Strike" and valve should be proud that I choose their game to show that. Right now the SC2 AI didn't show any better strategies than human, just better micro which is not surprising at all. So if this continues it means nothing if it beats all human players, maybe it's not more clever, it just has 1000 more fingers. And every researcher still believes they created an AI that is better than human in strategies and say "look the apm of AI is lower than human players", which is absolutely misleading. | ||
|
Poopi
France12901 Posts
On January 25 2019 06:37 brickrd wrote: actually, SC2 fans should understand a thing: the purpose of deepmind entering starcraft is not to provide us entertaining bots to satisfy our intellectual curiosities about the nature of competition in RTS. it's to promote and showcase machine learning. they're doing something historically and technologically significant whereas we are playing games. we should be thanking them for taking any interest at all and for giving SC2 visibility rather than criticizing people who are professional researchers working in their own damn field. Actually their goal is to progress towards general AI, and they think "solving" a game such as SC2 will help them get to that point. So the question of how fair does the AI play is very important, because for any breakthrough to happen, they'll need to be able to beat top humans consistently in the most fair way possible (so that it wins through intelligence). If they manage to create such an AI, it'll have to be robust on its own (by fast adaptation), one way or another. | ||
|
No_Roo
United States905 Posts
| ||
|
StormsInJuly
Sweden165 Posts
But I agree it's interesting to place restrictions on it and see if it can still beat pros with a hard cap on 200apm, then 100apm, etc. Seeing it get juked hard by warp prism play makes me think it's a bit too early to start handicapping it like that though. | ||
|
Cuce
Turkey1127 Posts
It woudl have been a much more important achivement with more limited micro I think limitation is the whole point. we dont need alpha ai to beat pro players on micro. we had automaton 2000 in beta fr that. the idea is to beat the pro player in decision making not micro, without limitations there is no point to this experiment. | ||
|
Charoisaur
Germany15979 Posts
On January 25 2019 07:32 StormsInJuly wrote: Asking for fairness, or trying to get an AI to play as human-like as possible by imposing restrictions on it, is sort of missing the point. But I agree it's interesting to place restrictions on it and see if it can still beat pros with a hard cap on 200apm, then 100apm, etc. Seeing it get juked hard by warp prism play makes me think it's a bit too early to start handicapping it like that though. no it's not missing the point - that is exactly the point. The point of this deepmind bot is to prove machines are smarter than humans we already know that they can have better macro/micro with a million apm. | ||
| ||
