The results of the study are published in the top journal Nature. Can’t read the research paper because the DOI link is not active yet, but it should be fine by tomorrow. However, there’s the news article, which is conveniently not behind a paywall.
Thoughts?
I particularly liked the part about “has completed the StarCraft II challenge” by making it into Grandmaster, and NOT by beating Serral. That’s pretty rich for an AI that learned how to play from a million human games and still hasn’t figured out how to build a tight wall at its main.
Hope they’ll have a showmatch at Blizzcon, and get recked.
How popular is alphastar? You’d probably get more replies if you explained it a bit or provided a link in your OP.
Pretty well know. I think almost everyone here knows about it.
Ah well I’m a few years out of date. I’ve been googling it though.
Since GM and up are allowed to have opinions on balance, I asked AlphaStar which of the 3 races he thought was strongest. And I quote:
100101010101010101011101101010101010100101010101010101110101010101101101010111001010101011010010101011010101111010101001010100001010101010101011101011101111111010100101010110010101010101010101110110101010101010010101010101010111010101010110110101011100101010101101001010101101010111101010100101010000101010101010101110101110111111101010010101011001010101010101010111011010101010101001010101010101011101010101011011010101110010101010110100101010110101011110101010010101000010101010101010111010111011111110101001010101
There you have it folks. ZERG OP CONFIRMED.
The first number in any set of binary numbers is always 0
Spoken like a true zerg apologist
You’ve alphad your last star this time
https://deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-StarCraft-II-using-multi-agent-reinforcement-learning
I’m not going to be impressed with AlphaStar until it plays the same game as the rest of us. It plays an easier version of the game and until that is fixed their results are skewed. Many of the “bots” have huge strategic flaws in their play that can be exploited and that just completely break the bot’s “brain”. Right now the bots don’t play many games, don’t play repeat games and play under a barcode.
Probably the biggest aspect of the game is understanding your opponent, how they think, play, how they will react to X information, etc, which they have completely shielded the bot from.
Give the bot a name. Have it play 10k games on the ladder and I guarantee it won’t be anywhere near the rank they claim it is at.
It’s probably too hard to teach the bot what balance patches are… They’d probably have to start over from scratch.
The bot is just a map of inputs to outputs. That’s all it is. It’s matrix math, but with REALLY big matrices. It has no understanding of the game. They would have to retrain it on new maps, new balance patches, and basically for any new scenario that the bot doesn’t have a map for.
Not from scratch exactly. The existing bots would become somewhat outdated and drop in performance, but they could use the same procedure to create a new set of agents trained on the new patch.
But, aren’t there tons of barcodes in GM anyways? Besides, people were sniping alphastar knowingly and with intent and still dying to it, so your claim falls flat.
Well you just don’t understand how these things work. Strategies in SC2 are like competing microbes. Microbe A might have a 70% success rate vs microbe B, but have a 10% success rate vs microbe C. So what strategy is best depends on what composition of microbes there are. If there are lots of X microbes, the microbes that counter X will do well, but as those are on the rise, the ones that counter them will rise off of them. So it’s constantly shifting. This is how the meta works. Until AlphaStar is an active participant in the meta, nobody is going to cater their builds towards it. Nobody is going to say, “I am going to use strategy X which will have a 99% win-rate vs AlphaStar” when 99.99999% of their games are going to be vs someone other than AlphaStar.
Wow, batz is actually right about something for once. I guess this is what happens when he’s not talking about balance.
That’s kinda what people did, since alphastar replays appeared all over the internet. Granted all the players who were actually solid GM were quite comfortably able to stomp alphastar with the power of blind counter. But now it’s kinda the human who is cheating, is it not?
Nope. They claim the bot is playing with the same restrictions as everyone else, while they shield it from being profiled. Yet profiling is THE biggest aspect of strategy. They are saying “Sorry, you guys can’t use strategy vs AlphaStar, you have to go against it thinking it’s a random human opponent, even though we’ve trained it to beat the current human meta.”
If they let it play a bunch of games, and start being a meaningful contributor to the meta, players will lose mmr to it and they will have to start considering it’s playstyles in their own choices. Right now there is no incentive to do that so it gets to sit and game the meta with a decisive strategic advantage.
Uhm… Forgive me if I’m wrong here, but don’t you get matched up against random people on ladder? So you may not know who you are fighting against.
Pros already have barcode accounts for this reason, the well known people stay anonymous and the average folk aren’t well known anyways.
No, memorizing account names is not the biggest part of strategy. Outside of GM, it rarely amounts to anything because there are too many players, and in GM everyone has at least half a dozen strategies. Strategy SHOULD come down to watching what happens IN GAME, not because “oh I know this guy and he literally only ever does 1 build?”