|
Nature paper : storage.googleapis.com
Explainer video : www.youtube.com
Blog post : deepmind.com
(featuring coments from Mana, TLO, and Kelazhur)
Furthermore - Oriol, the tech lead of AlphaStar at DeepMind, is coming to give a tech talk at Imperial College in London (open to all, free registration www.eventbrite.co.uk), where I research reinforcement learning. If you have any questions about the inner workings or the future of AlphaStar, please post them in this thread and we will try to accomodate them during the Q&A session.
Thanks all
|
This thing is so exciting to me. I view it as, effectively, the fourth generation of machine learning. Obviously generations within machine learning is very subjective. This is the fourth time that a world class AI is going to do battle with the best in the world at a game. This same concept beat the best humans at chess, then Jeopardy!, then Go is making its way up to the big leagues against the best in the world at StarCraft II. Obviously it still has some work to do, but making grandmaster as all of the races is a significant step.
I can't wait for the moment in the future when it squares off against whoever is the best in the world at that time. My money is on AlphaStar.
|
Can someone please tell me what parameters are set for Alpha Star? APM, reaction time to incidents etc...
|
|
|
I have a high IQ but can not escape diamond.
AlphaStar has less intelligence than a dog but is GM.
Conclusion:
I must balance whine harder. Blizzard is clearly at fault.
|
I'm really interested in what the unique strategies AlphaStar employs are--anyone have any info?
|
Nice, can Alpha Star play at WCS tournaments?^^
|
So when's Alphastar going to play Serral?
I feel like they need 3 series of bo5, one for each race!
|
On October 31 2019 06:32 Duckman wrote:
I'm really interested in what the unique strategies AlphaStar employs are--anyone have any info?
From what I saw earlier, not too much, but there were speed banshees and an insane roach/ravager 3base push.
It is not easy to compete with perfect macro+decent micro and decisionmaking.
There were some holes like walling off, countering air armies and transitioning to the lategame which can be exploited by players who know what they are up against.
|
On October 31 2019 06:41 Big-t wrote:
Nice, can Alpha Star play at WCS tournaments?^^
they need to be at least 16 to be allowed to participate. Baby AS is not yet eligible.
|
So... uh.... .how much data have alpha star skewed the balance team's data point ?  Is any of the recent zerg op issue was partly caused by Alpha star's terran and protoss is way stronger than normal human at the same level ?
|
On October 31 2019 06:45 loppy2345 wrote:
So when's Alphastar going to play Serral?
I feel like they need 3 series of bo5, one for each race!
This is some Euro ZvZ I'd love to see.
|
I know that alphastar played Serral in TvZ, Serral won.
Its casted somewhere on Basetrades Youtube
|
Deepmind must be burning through cash on this Starcraft project. Their losses were $570 mil in 2018, and $370 mil in 2017. I feel they could well run out of money very soon, unless they manage to find a commercial use for this type of research.
|
On October 31 2019 06:06 Jathin wrote:
I found the footnotes to figure 2 really funny:
B: MMR rating of AlphaStar Final per race (from top to bottom: Protoss, Terran, Zerg) versus opponents encountered on Battle.net (from left to right: all races combined, Protoss, Terran, Zerg). Note that per-race data is limited; AlphaStar won all Protoss versus Terran games.
Now you're gonna see a post on Blizzard forums "PROOF ONCE AND FOR ALL!!!"
If I'm understanding that correctly, for some reason there's only 4 PvT games of alphastar final vs players on battlenet in their sample. Which is an extremely low number for a quantitative paper so I may be misunderstanding that.
Edit: Though... It is very interesting that it won 25/30 games vs players as protoss, and only 18/30 for each of Terran and Zerg.
Edit: Yes, it actually didn't play that many games versus humans. So we REALLY can't read that much into it. Like, it also won four games of TvP (out of eight) and four games of TvZ (out of seven). Which just goes to show that we really don't know what it's PvT numbers would have looked like if there had been more of them. With that said however, the 25/30 as protoss overall definitely says something I think.
For the supervised and midpoint evaluation, each agent began with a fresh, unranked account. Their MMR was updated on Battle.net as for humans. The supervised and midpoint evaluation played 30 and 60 games respectively. The midpoint evaluation was halted while still increasing because the anonymity constraint was compromised after 50 games.
Edit: Also confirming the above. The numbers given there (25/30 as Protoss, 18/30 as Terran, 18/30 as Zerg) match the replays they attached. Those are numbers for the final version of their agent.
18/30 as Terran, 18/30 as Zerg? I would hardly call that "mastering" the game. Though given the calibre of players those final matches were against, it's still impressive.
|
Do wa have any idea of the MMR of AlphaStar ? I'm a bit surprised they're saying that they mastered SC2 with so few published data.
|
On October 31 2019 07:17 heqat wrote:
Do wa have any idea of the MMR of AlphaStar ? I'm a bit surprised they're saying that they mastered SC2 with so few published data.
6000 MMR, but it appears to be from a VERY limited sample size in terms of actual players they faced with the final version (see my above post). And Alphastar's numbers as Terran and Zerg are not really that impressive. (Winning only 18/30 games as each).
|
What does AlphaStar get "tricked" by the most often?
|
Bisutopia19295 Posts
I wish they would add AlphaStar as an AI option. I'd love to 1v1 and try to out think my opponent knowing my mechanics are below its level.
|
Hi MyLovelyLurker, I'd like to know:
1. When selecting multiple units, human players need to drag a box, which limits how they can micro, for instance in things like splitting marines against banelings. To be fair, could you make AlphaStar drag boxes too in the future?
2. When a dropship shows on minimap, there's a ~1/4 chance a best human player won't notice it. Is this implemented in AlphaStar?
3. When a cloaked unit moves across the screen, there's a ~2/3 chance a best human player won't notice it. Is this implemented in AlphaStar?
4. When clicking on something small, there's a ~5% chance a best human player will misclick, I guess this is already implemented?
5. Will you emulate a virtual mouse in the future, so that we can watch AlphaStar's first person view in the ultimate AlphaStar vs Serral match, which is bound to happen sooner or later?
Thanks!
|
|
|
|
|
|
EPT
