Replays are in replays.zip in supplementary data
Grandmaster level in StarCraft II using multi-agent reinforcement learning
11–20 of 321 posts
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#12> After 50 games, however, DeepMind hit a snag. Some players had noticed that three user accounts on the Battle.net gaming platform had played the exact same number of StarCraft II games over a similar time frame — the three accounts that AlphaStar was secretly using. When watching replays of these matches, players noticed that the account owner was performing actions that would be extremely difficult, if not impossi…
To play against alphastar, you have to opt-in. IIRC it states Alphastar will hide itself.
One of the things people noticed in replays was the lack of control groups and in the case of zergs, the ability to select larvas directly, which no player ever does. It could have been as simple as removing these quirks.
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#13How soon is this put into the hands of battlefield commanders? Not presuming it’s a bad thing, but I wonder when we’ll hear the outcome of a battle was aided by this kind of intelligence.
A bit further away I would say. So far it trains on a deterministic model by playing millions of games against itself. It can be thrown off by a single rule change. The challenge here in making an AI usable on the real battlefield is more on making a realistic simulation rather than on a better AI.
So, what I'm hearing here is that we should start a series of endless wars?
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#14> placing within the top 0.15% of the region's 90,000 players > 61 wins out of 90 games against high-ranking players This doesn't seem to be quite as commanding as it was in Go. Do we know what MMR it reached or if it consistently beat players like Serral?
It hasn't beat Serral once, from the article: "The AI wasn’t able to beat the best player in the world".
To play devil's advocate:
They didn't say anywhere they scored a win against the top 10 either.
To elaborate: to be among the "top 0.15%" of 90k is place ~14 (13.5) in the worst case, so nowhere near the "commanding" abilities mentioned by GP
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#15There has always been the issue of interface when playing videogames AI vs human. Either give the human a brain-computer interface or give the AI a mouse, keyboard, monitor, robot hands and a camera. Anything else seems inherently unfair.
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#16> After 50 games, however, DeepMind hit a snag. Some players had noticed that three user accounts on the Battle.net gaming platform had played the exact same number of StarCraft II games over a similar time frame — the three accounts that AlphaStar was secretly using. When watching replays of these matches, players noticed that the account owner was performing actions that would be extremely difficult, if not impossi…
These extremely difficult/impossible things didn't really give an advantage. For example, AlphaStar would sometimes click on an object at the border of the screen. For humans that would be almost impossible, because the screen would scroll when the mouse approaches the border.
Similarly, AlphaStar would not play with group hotkeys, but use a different technique. However, in none of the analyses, people noticed things that would give AlphaStar an unfair advantage.
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#17Earlier quoted context omitted.
It hasn't beat Serral once, from the article: "The AI wasn’t able to beat the best player in the world".
To play devil's advocate: They didn't say anywhere they scored a win against the top 10 either. To elaborate: to be among the "top 0.15%" of 90k is place ~14 (13.5) in the worst case, so nowhere near the "commanding" abilities mentioned by GP
You'll struggle to work it out, go look at the top 20 of any of the ladders IIiiIII1iI1iII!II
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#18Not being robust to strategies it hasn't seen before is a serious shortcoming in a real time strategy game. That also indicates an interesting flaw in how this model is trained - in the millions of games it plays against itself, how do you ensure that it tries every viable (and some inviable) strategies? Sure, it couldn't best Serral but I wonder how it would fare against Has, a player known for some pretty off the wall builds.
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#19> After 50 games, however, DeepMind hit a snag. Some players had noticed that three user accounts on the Battle.net gaming platform had played the exact same number of StarCraft II games over a similar time frame — the three accounts that AlphaStar was secretly using. When watching replays of these matches, players noticed that the account owner was performing actions that would be extremely difficult, if not impossi…
To play against alphastar, you have to opt-in. IIRC it states Alphastar will hide itself. One of the things people noticed in replays was the lack of control groups and in the case of zergs, the ability to select larvas directly, which no player ever does. It could have been as simple as removing these quirks.
For the type of beast AlphaStar is, they do not sound like simple tasks at all.
Re: Grandmaster level in StarCraft II using multi-agent reinforcement learning
#20> After 50 games, however, DeepMind hit a snag. Some players had noticed that three user accounts on the Battle.net gaming platform had played the exact same number of StarCraft II games over a similar time frame — the three accounts that AlphaStar was secretly using. When watching replays of these matches, players noticed that the account owner was performing actions that would be extremely difficult, if not impossi…
I'm not sure this is unfairness. IIRC they put in a fair bit of effort to put it on a level playing field with humans by limiting APM and not allowing it to observe multiple areas simultaneously by spam moving the camera.
It might have some minor unfair advantage in terms of being able to click with pixel perfect accuracy, but they're marginal and from watching this project evolve, it's pretty clear that the strategic planning aspect of alphastar has indeed become better than humans.
"Agents were capped at a max of 22 agent actions per 5 seconds, where one agent action corresponds to a selection, an ability and a target unit or point, which counts as up to 3 actions towards the in-game APM counter. Moving the camera also counts as an agent action, despite not being counted towards APM."
This is worse micro than top human players.