EPTUpdates on alphastar / deepmind?
| Forum Index > SC2 General |
|
MarlieChurphy
United States2053 Posts
| ||
|
ilax300
13 Posts
| ||
|
jy_9876543210
202 Posts
On October 12 2021 06:21 ilax300 wrote: even with a APM cap APM cap? you mean the 2000 APM cap?  | ||
|
Chris_Havoc
United States518 Posts
You can read about DeepMind/AlphaFold's work here: https://arstechnica.com/science/2020/11/deepmind-ai-handles-protein-folding-which-humbled-previous-software/ | ||
|
Husyelt
385 Posts
| ||
|
Slydie
1329 Posts
On October 12 2021 11:52 Husyelt wrote: AlphaStar's team got embarrassed by losing their first actual live match in a PvP against John MaNa Connor. The team quickly closed shop and declared that they "Beat StarCraft" 10-0 in matches. Once they actually limited the AI to the actual screen's vision, (aka as playing the actual game,) it was figured out in a single match. No B07 against Maru or Serral. No ladder version. Was that really the reason? Afaik, they stated they successfully reached GM level and went toe to toe with top pros with all 3 races, which is how far as they got with the resources and timeframe. I believe they could have gotten to an undisputed super-human level as well, but hey, beating the game was never the main goal. | ||
|
Ciaus_Dronu
1289 Posts
On October 12 2021 06:21 ilax300 wrote: Oh that project ended a long time ago, the purpose never was to make the perfect AI that would conquer the world. It was more to gather data and help write a paper about it all. It would also be too expensive to continue at as it would have to learn to play on a new set of maps each time. Shame they never did some proper showmatches though, all in all I felt the whole thing was underwhelming and even with a APM cap, perfect control was nearly all alphastar relied on to secure wins. Eh, I don't really agree with this. It was really good at judging what fights to take, and seemed to have a reasonable sense of the map and where stuff was too. I mean the one Zerg one won a masters level ladder game against a cannon immortal proxy rush by double-expanding and just out-econing the opponent in what was honestly one of the coolest holds of that build I've seen - even if the execution was a little eh. Watching its ladder games where it has to face weird and aggressive early nonsense was much more impressive than the showmatches. It actually sort of knew what do in a lot of those spots. | ||
|
NinjaNight
413 Posts
On October 12 2021 17:13 Ciaus_Dronu wrote: Eh, I don't really agree with this. It was really good at judging what fights to take, and seemed to have a reasonable sense of the map and where stuff was too. I mean the one Zerg one won a masters level ladder game against a cannon immortal proxy rush by double-expanding and just out-econing the opponent in what was honestly one of the coolest holds of that build I've seen - even if the execution was a little eh. Watching its ladder games where it has to face weird and aggressive early nonsense was much more impressive than the showmatches. It actually sort of knew what do in a lot of those spots. They were very limited in strategy, like the Zerg one every single game I watched over and over would always do some kind of roach queen ravager early timing attack. Terran one vs zerg would overwhelm you with constant marine tank pushes and some marine drop harass at main base | ||
|
meadbert
United States618 Posts
Also, AlphaStar was WAY ahead of the metagame. AlphaStar went Disruptors in PvZ which at that time was considered crazy. Also in ZvP AlphaStar did a lot of Roach/Ravager Queen walks years before it was popular. AlphaStar's 2 Adept into Oracle open against Zerg was deemed so good, that they responded by nerfing Adept built time by a few seconds. | ||
|
Husyelt
385 Posts
On October 13 2021 02:29 meadbert wrote: AlphaStar did beat Serral in PVZ (which is AlphaStar's best matchup) Also, AlphaStar was WAY ahead of the metagame. AlphaStar went Disruptors in PvZ which at that time was considered crazy. Also in ZvP AlphaStar did a lot of Roach/Ravager Queen walks years before it was popular. AlphaStar's 2 Adept into Oracle open against Zerg was deemed so good, that they responded by nerfing Adept built time by a few seconds. Serral tried to beat Alphastar by just macroing and playing standards. The fifth or six time he ended up winning. Once you beat a bot or find it’s weaknesses it’s game over. There’s no actual adaption at a certain point. The truth is that the team realized they had hit an AI wall and weren’t making any progress. They learned a lot, but the “close up shop and move on” was shady to say the least. | ||
|
-NegativeZero-
United States2115 Posts
On October 13 2021 02:00 NinjaNight wrote: They were very limited in strategy, like the Zerg one every single game I watched over and over would always do some kind of roach queen ravager early timing attack. Terran one vs zerg would overwhelm you with constant marine tank pushes and some marine drop harass at main base If you're playing best of 1s against random/anonymous opponents, that's exactly what you'd expect from an AI that works by gradually iterating and optimizing its performance. | ||
| ||
