AlphaStar vs Serral - An Introduction
Вставка
- Опубліковано 16 чер 2024
- DeepMind's AlphaStar played against the World's #1 StarCraft 2 player, Joona "Serral" Sotala at Blizzcon.
In this video I will go over a brief history of AlphaStar, as well as address the most asked questions about fairness. Here's a brief summary:
1. AlphaStar's APM was human level.
2. AlphaStar has to switch screens just like a human.
3. Serral played on a mouse and keyboard that were not his.
4. This was not an official match, just something fun Serral did. He should be thanked for it!
5. Serral did get to rebind his hotkeys before the match.
The next 5 videos will be analysis of AlphaStar's games against Serral.
The 7th video will be an interview with the AlphaStar Lead, Oriol Vinyals.
Next Video: AlphaStar vs Serral Game #1
• AlphaStar vs Serral - ...
Also check out more of my content:
Twitch.tv/Artosis
Artosis - Ігри
Karoly from Two Minute Papers says hi
Hello dear fellow scholar!
Here from 2 minute papers
Yes, love this intro! This playlist is gonna be a great place to start to introduce someone new to watching SC2 in general.
artosis coming in with the definitive lecture on alphastar. welcome to his masterclass.
wow. I was really scared that we would never see these games! Thank you so much!
This is super cool. Can't wait to watch the games!
your professionalism is so far off the charts, Artosis.. excellent video and suberb presentation. pay this guys more money!
*jumping up and down like a toddler* ARTOSIS AND ALPHASTAR AND SERRAL!!! Yippeeee! Thanks so much for doing this Dan.
Great work, Dan!
@9:10 this analogy is so clean and demonstrative of why artosis is GOAT
very interesting video thanks artosis!
Hell yeah, just came back from work and i see this. Now lets watch this playlist gem :).
this mas deserves more recognition for his work
It seems like reaction time is a factor that's still being overlooked in terms of bringing the AI in line with human mechanical limitations. The average human reaction time to a simple visual stimulus (such as the screen changing in color) with a mouse click is around 200ms. Add the requirement of reacting to less obvious changes to the screen (such as change in direction of movement of a unit) as well as the fact the reaction will likely require moving the mouse before clicking, and you're pushing the average time between the visual stimulus and the completion of the reactive action to at least 300ms.
AlphaStar's reaction time needs to be limited to around 300ms at least before we can really consider it to be playing the same game as humans.
AlphaStar's instant reaction time with perfect mechanical accuracy means that there will be plays that it makes that are bad plays for a human but good plays for AlphaStar because it can do some thing on reaction. This hurts the ability of AlphaStar to serve as a learning tool for humans, since a human might try to copy a play from AlphaStar not knowing that its viability is entirely contingent on a level of reaction time unattainable by humans, and wonder why it keeps failing.
some of these "decisions" don't matter because of the evaluation function. simply put, alphastar only calculates win/lose/draw scenarios. in other words, will my next move force me to lose? if not, then I'm at least safe. The ironic part is that all of those TPU's are mostly focused on the state of play in the current moment. Whereas humans look at their opponents' plans and study their games, DeepMind is like, "yeah, everything looks good at this point in time.
@CptShrimps You'd be interested to know that AlphaStar actually does simulate reaction time! From what I could understand from the paper, it has two "reaction times". First it has a delay of about 110 ms between when something is observed and when an action is executed, and the second happens because it predicts what to observe next to about 370ms (though could be as long as a couple seconds) into the future, and cannot change that, meaning it could potentially react significantly slower than normal if caught off guard.
I don't think that its a perfect model to emulate human reaction time, but it does remove a majority of the computer’s “instant reaction time advantage”, and im glad that the researchers did address it.
@@Golinth i just never said the computer was instantaneous. In fact, You said the same thing i did about the AI focuses on the present a lot more than anything else. Against humans there is just no way of beating something that never gets tired, has eye strain, gets bored or dejected with their play unless we quickly take advantage of a flaw in the evaluation function before it fixes itself. the 110ms has to do with alphastar first being able to see the whole map and could micro you from anywhere without having to switch screens. The pros said that that wasnt fair and deepmind fixed it. cheers
Are you even aware that you just started that comment the same way that Deckard did in Blade Runner when doing a Voight Kampf test on a Replicant?
_"Reaction time is a factor"_
Your videos make me cry of joy.
Great video Artosis. Edit: great series!
Shouldn't you be comparing to human *EPM*?
Very interested to learn more about Alphastar
never played starcraft, more of an AoE guy. thanks for breaking everything down for me
Thanks so much for this!!! woooooooooo so cool.
thanks bro! After all these years I still want to protect this smile!
I'm so glad they toned down AlphaStar to make things more fair. Imagine a player with perfect micro, perfect construction, perfect economy and perfect sight.
After randomly getting interested in chess-playing AI again a couple of days ago, reading up and coming across DeepMind's work, I have to say that I've been blown away. When I first heard about Deep Blue, I had the impression that it was "thinking" more than it was. After watching Down the Rabbit Hole's video on Deep Blue and the match between it and Garry Kasparov, it seemed a bit less impressive to me, knowing how much of the opening moves and previous high-level chess games were programmed into it. However, AlphaGo and especially AlphaZero felt like the "real deal" to me, and I've had such a great time learning about them. AlphaStar really seems incredible away, since to me StarCraft appears infinitely more open-ended than games like chess and go. When I searched UA-cam for AlphaStar, I jumped right into the AlphaStar vs. Serral games, and having watched them, I feel like this explanation is even more impressive. My hat is most definitely off to you Artosis; thanks for a wild ride!
"there's something i ought to tell you. i'm not left-handed either."
I'd like to see how AlphaStar would do on different servers. Does it gain too much of an advantage if an areas average APMs is too high? Is any one region more adaptable than others?
Any way to get the AlphaStar vs Serral replays?
artosis, i heard beastyqt had some good games with alpha star and florencio had some good cheeses if possible would love to see those games
Is there a video on youtube where you interviewed TLO and DeepMind? Could you link that?
ua-cam.com/video/jowvgczuS-s/v-deo.htmlm30s
@@MTGandP broken link
Ty!
How do I find an event video which is shown 8:30 at top of screen?
How did your games vs. it go?
Artosis = the man! let's gooo
Greetings from Two Minute Papers.
No need to introduce yourself, Artosis
your job is fun :D
Oh boy, here's tonight's entertainment sorted out.
is this guy like the agadmator of starcraft?
I know they wanted it to be fair and interesting, but isn't it counter productive to limit alphastar?
Depends on what you want from your AI. If you just need it to work with digital data then let it go wild, but if you want it to eventually interact with real world stuff then you do need to limit it.
No, computers have been able to beat humans forever just be moving way quicker, search 'starcraft automaton' for a robot that uses 20000 APM or something ridiculous
If you gave the bot unlimited APM, it could just build really bad units, not understand any of the strategy, and win by controlling better than a human ever could - the point of this is that no AI (until now) has been able to beat a pro player via strategy only
I honestly didnt find anything cheaty about the initial "full-map" camera view of AlphaStar in January.
Because players have access to the same information, via the minimap.
I DONT LIKE NO REGRET
alphastar made every action different, while human pressed move ( M ) or attack ( A ) 20 times when once was enough, so basically, AI is waaaay more efficient than human. Maybe if humans started to make actions that only matter and not just repeat things for sake of apm, we would know the REAL apm to give to alphastar. 300 is way too much for an AI because every one of those 300 apm is going to be it's own thing.
Yeah, while top top pro players do have like 225 effective actions per minute. AlphaStar should instead have like 150-180 effective actions per minute max, that would be much more fair.
The best player that got destroyed in BlizzCon 2019, the only significant tournament he won (last year) so far. Tsk tsk.
at what criteria is Serral world best player again?
Aligulac
@@sehyo7403 why do we have world championships?
He's playing the current best race, and performing most consistantly against all of the other top players, having a weakness against a single player does not mean that players overall performance is above yours. Dark is the only player who can match his record currently, and the better of the 2 is unknown. So until that's clearly determined it's fair to call Serral the best, even if he's not the only one in that position.
@@j.e.6083 only the championships that serral wins count
Pro player consensus
wrinkle brain plays only
Reducing Alphastar to human level skills sort of defeats the point. It’s like asking Serrel to tape his fingers together because he is ‘too fast’ for other human players.
Disagree. The point is whether or not it is possible to make AlphaStar "smarter" in a sense, not just the best at micromanagement. We already know that in terms of precision, AI is better at us in every way. The goal is whether or not AlphaStar could learn the game and make sound strategic decisions.
Give this guy (Serral) to play AlphaStar how many times he want, god. Weird situation, I hear a little bit too much about it, of his defeat, and other's people excuses about it. So he can easily show AlphaStar is nothing to him.
It'll still win every time. Its just that good.
Well, to be fair, AlphaStar is a really good player. That should be said.
Though Serral has shown to be able to defeat AlphaStar on the ladder. And plenty other lower players defeating it too.
However, the main point I want to make is instead: that after say ~20 losses against AlphaStar, the Serral (a human) would be able to figure out a strategy defeating AlphaStar. I have no solid grounds on this statement, other than that was what happened with the DOTA AI, after a while pro-gamers defeated the AI much more efficiently.
stop telling us who you are.. you're fucking famous, everybody knows you !
"I was a professional gamer." Really? When? Top US amateur in SC1, zero achievements in SC2 as a player.
You are salty AF
He qualified for GSL in the beginning of Starcraft 2, with zerg. He then quickly focused on casting instead of playing though.
And yeah, he was bascially talking about both SC1 and SC2. If they make Tennis 2 with slightly different rules and maps, Roger Federer is not suddenly not a pro Tennis (2) player.