How AlphaStar Became a StarCraft Grandmaster | AI and Games #48

Support AI and Games and help the show grow by joining my Patreon:
/ ai_and_games
--
For the first episode of 2020, I return to arguably the biggest AI story of 2019: the success of AlphaStar. An exerpiment in deep reinforcement learning by Google DeepMind, AlphaStar is the first StarCraft AI player to not only defeat professional players but also to achieve Grandmaster rank in online play. In this video I walk you through the story of how it all works.
Chapters
[00:00] Introduction
[00:42] Why StarCraft?
[04:47] How AlphaStar Works
[10:28] Defeating Pro Players
[13:52] Grandmaster
[17:00] Closing
Some extra resources:
The StarCraft Survey Research Paper that estimates 10^1685 states
www.cs.mun.ca/~dchurchill/pdf/...
Google DeepMind AlphaStar Blogs:
deepmind.com/blog/article/alp...
deepmind.com/blog/article/Alp...
Nature October 2019 Paper on AlphaStar
www.nature.com/articles/s4158...
AlphaStar Replays (can be played in StarCraft 2):
deepmind.com/research/open-so...
--
AI and Games is a KZbin series on research and applications of Artificial Intelligence in video games. It's supported through and wouldn't be possible without the wonderful people who support it on Patreon, plus KZbin members and Paypal donations.
/ ai_and_games
/ @aiandgames
www.paypal.me/AIandGames
--
Get yourself an AI and Games t-shirt over on Teespring!
teespring.com/stores/aiandgames
You can follow AI and Games (and me) on Facebook and Twitter:
/ aiandgames
/ aiandgames
/ get_tuda_choppa

[@AIandGames]

I'm back! Let's kickstart 2020 with the biggest AI story of 2019: Google DeepMind's AlphaStar achieving grandmaster ranking in StarCraft 2 multiplayer. Sorry it's been so long, I was busy with a consultancy gig, then a conference, then I managed to be ill twice in one month! But that means I had plenty time to write new episodes for y'all to watch.

[@MattSipka]

AI and Games it would be interesting to see alphastar play like an actual human, down to typing BM back at salty players.

[@seamon9732]

Is it still cheating with inhuman EAPM bursts during fights despite having an acceptable average ( which means nothing )? If so... meh.

[@darosmaedafreitasassuncao5936]

nice one m8

[@oriolvinyals5565]

Thanks for the video : )

[@WhirlwindHeatAndFlash]

The title is misleading. Even a potato could reach SC2 grandmaster outside of South Korea. HueHueHue

[@centosan1]

Now we can let the best AI protoss, zerg, and terran duke it out in order to find out which race is truly unbalanced

[@Codabear89]

Actually not a bad idea at all

[@riverbucknor2222]

@Isaac Robinson Or they could just identify that off player statistics which is pretty much what the ai learns from anyway

[@musclecard54]

@@riverbucknor2222 not to mention balance is based off of players. if the stats are balanced for the playerbase, then the game is balanced. it doesnt matter how much a bot can exploit some obscure strength for a race if players cant do that. balancing based on bots would lead to imbalance for players

[@pingwingugu5]

@@musclecard54 It depends on how the bot is build. If it is designed to imitate human limitations (reaction time, max number of operations per minute etc.) then it can be used to find exploits that human player base didn't find yet but would be able to use.

[@hats1642]

Alphastar is better at Protoss than the other races, I think many of the agents had trouble figuring out depot walling as terran. It is pretty good at ravager micro though.

[@RFalhar]

The mention of version 1 having "normal" APM is misleading. True, it's average APM was low. But it was able to spike it's actions for few seconds. Which is why for they second version, they made APM limits even stricter, and added APS (actions per second, or was it 5 second?) limit also. It is obvious to anyone knowledgable about StarCraft, that in match with TLO, AlphaStar was able to micro it's Stalkers to inhumane levels. Juggling them around with Blink to the point of getting massive value from each individual unit.

[@rompevuevitos222]

Funily enough, this was the most realistic version, normally players tend to do a constant 300 APM with spikes of 600+ due to how APM works Basically if you hold a command key and have a command being spammed each trigger will count as an action, meaning that in some situations (specially with Ravagers) the APM can get to 1000+ with the press of a button

[@seamon9732]

@@rompevuevitos222 APM /= EAPM, alphastar's APM = EAPM, a normal player's EAPM however will always be lower than its APM. As long as AS can cheat with inhuman EAPM micro, I'll still call bullshit on their ridiculous claims.

[@xZerplinxProduction]

Besides, Alphastar's clicks are of much higher value per click. Humans spam click as an alternative to try hitting precise timings and to keep their hands warm and ready.

[@palebears8913]

@@seamon9732 Have you seen any of the replays? Alphastar's micro is actually not the best when playing in GM. Human players gain advantage in micro is pretty common.

[@BrokenSymetry]

@@palebears8913 Fully disagree, AS micro is what helps it achieve, white it's long term strategies, planning and improvising are it's weak points.

[@kevingriffith6011]

16:40 Obviously the most important information to take from all of this is that zerg is OP and protoss needs a buff (sarcasm, obviously.)

[@MattSipka]

Kevin Griffith I agree and Terran needs a nerf. And I play Terran.

[@kevingriffith6011]

And I just realized you read them from lowest to highest, so I got it entirely backwards!

[@Healermain15]

@@kevingriffith6011 Well obviously Protoss are OP, ramble ramble deathball ramble attack-move ramble

[@eduardowada358]

Alphastar's zerg didn't know how to deal with air units and terran didn't know layout their base so its units wouldn't get stuck, also it's possible that Alphastar's APM limitation prevented it from properly using infestors, which were zerg's op unit at the time. Alphastar's protoss didn't seem to have any big derps like those. Given those issues I don't think Alphastar's level was representative of the game balance.

[@iglidor]

@@MattSipka That is exactly what zerg player would say to push his agenda for nerfing Terrans

[@fragrantwinter8233]

Incredible... It’s like watching a form of computational evolution in action. The ‘arms race’ Tommy described is so easily paralleled with the struggle for survival that life faces: Obsolete or otherwise ineffective agents are deleted or overridden, while the more effective agents grow in use and each instance builds upon the experience of the last to improve itself. The hills that are fairly niche or differing strategies are simply various ways to accomplish the same ultimate goal - Win/Survive and pass on the knowledge to the next set of agents to improve upon... Perhaps I’m looking too far into this whole process for biological parallels; but the clarity of it those parallels makes it all difficult to ignore.

[@dgam4211]

it definitely works exactly like that, even if it is very inefficient in a calculations standpoint

[@ryangarrison8572]

Brilliant analysis

[@willyolio9590]

The APM thing was far more significant than you realize. TLO's APM was high because he tends to do some useless clicks, like, ordering units to move to an area, he could click 5 times when 1 would do, or hold down a hotkey button (the APM then jumps to the keyboard repeat speed). It's a habit lots of SC2 players have, often just to keep the fingers warmed up. However, when it really matters most (intense battles with lots of precision), these pro players are actually only issuing maybe 300 APM worth of unique and intentional commands. Whereas Alphastar uses these moments to spike up to inhuman levels of 700+ precise commands. Then it issues few commands during the lull between battles, and the average ends up looking "normal" or even low. This was obvious to starcraft veterans watching Alphastar bounce its stalkers around. It's a pro move people have done before, but not with that many, not that consistently, and even the best pros can never do that while ALSO managing production and other things at home.

[@Treblady]

appreciate the communication style and positivity around ai. I want to learn in this field.

[@beryvice1598]

Please talk to BeastyQT (here on KZbin) about AlphaStars strategies and weaknesses. He is a grandmaster player who is very experienced and knowledgable about the game. He played some games vs AlphaStar and he would certainly enjoy sharing his thoughts with you on the matter!

[@jl6723]

I kind of want to see an inhumane version of alpha star where there are no limitations attached. I really wonder have inhumanely good it could get and what the top end of Starcraft would be like.

[@CG-eh6oe]

Would be super boring. The problem is that the game is not balanced for perfect micro, so certain units would become more and more OP the better the micro gets. My guess is it would be blinkstalkers, so with your suggestion you would just see sick microbattles of blinkstalker armies without any real strategy.

[@meandyours]

@@CG-eh6oe Are you sure it wouldn't be bunker or medivac/droplord/warp prism micro?

[@CG-eh6oe]

@@meandyours Im not sure, i wrote its a pure guess. The things you mention are possible exploits of infinite micro as well, another option would be mass marines (imagine marines who always dodge any AoE...)

[@meandyours]

@@CG-eh6oe have you seen microbot yet? that chanel have a great demonstration of it

[@manuel2824]

Probably efficient multitasking all over the map

[@Catatonic419]

200 YEARS OF EXPERIENCE!!

[@TheLibertyFreedom]

I've been looking forward to this! Thank You!!

[@emceha]

Wow, I just had a huge wave of nostalgia when I saw TLO. I hope life treats you well.

[@Nighthawk20000]

Thank you for this. I talked a bit about this on my podcast but I was mostly restating things I've read online. This video was a good way of educating myself

[@richy3454]

What was found by watching replays is alphastar kept its apm low most of the time and then would jump to almost 1000 apm during a fight. This kept the overall average low but its peak was way too high

[@o4ugDF54PLqU]

Humans also spam to 1000 apm as well. In the first version of the AI this was a problem since it could do actions precisely and had a high effective APM. This was fixed in the second version so that it had to essentially use a mouse and KB, meaning it made quite a lot of micro mistakes and had a lower effective APM.

[@CallsignYukiMizuki]

@@o4ugDF54PLqU Obviously exaggeration, but for a player to be doing 1000 APM is like 16 and a half actions per second. And idk man, most I've seen is ~500 APM in the most intense fights and the usual is ~300 APM

[@o4ugDF54PLqU]

Callsign-YukiMizuki people have managed 1000 before, I remember seeing it in a casted game. IIRC it’s due to how the game calculates APM, it isn’t actually how many times you click a button, but how many commands are sent, so zerg players holding down a button on larvae can spike the APM.

[@paratus04]

As amazing as Alphastars progress was it’s interesting that it still loses unlike in Go or Chess. It seems the massive solution space coupled with the lack of perfect knowledge means it’s not possible to always have the perfect strategy. Case in point, while there is no video I’m aware of epic cheeser Florencio took a shot at Alphastar and apparently crushed it. I believe he did a variation of his drop a nexus in your opponent base and recall your units into their base. Not a playstyle Alphastar likely saw before.

[@rowanunderwood]

The difference in APM could purely be down to the fact that the AI didn't need to move the camera to see :)

[@HansPeter-xy7sl]

moving ur cammer dosnt count as apm in sc2

[@tuhaggis]

@@HansPeter-xy7sl It does when using camera location keys which is how high level players do it.

[@zaxtonhong3958]

TheITS No, camera locations are not actions according to the game. An action is selecting or commanding a unit. Camera locations do neither of those.

[@tuhaggis]

@@zaxtonhong3958 Just tried it in game and you're right, it doesn't count towards APM.

[@MaisieDaisyUpsadaisy]

TheITS Camera locations does not count towards APM. Source: masters player.

[@BrokenSymetry]

I always felt Alphastar won most of it's matched through superior unit control, while generally failing at long term strategy, planning and improvisation.

[@Sekir80]

Yeah. The lack of ability to even build a wall and still win was very confusing.

[@RoyalFusilier]

That's kind of AI in general at the moment. We can build an algorithm capable of executing decision-making, but we can't actually make one that understands why it's doing it. People aren't wrong to be a little wary of such a decision-executing entity.

[@WylliamJudd]

This is the video I've been looking for!

[@xPuhx7]

Holy moly, that's incredible I even understood something, you made a great job considering the complexity of the AI structure.

[@AIandGames]

That's the goal! If you can get something out of the video I'm doing my (self appointed) job. 😀

[@xPuhx7]

@@AIandGames I'll try to come back and rewatch the video for the better understanding some points. As a journalist I highly appreciate your channel because each video helps to come closer to truth.

[@wukong6853]

I need to go back and watch your videos, This shits really cool

[@ConnoisseurOfExistence]

Great video!

[@nagi603]

16:55 "This places alphastar in the top 99.85% of all ranked play" you mean top 0.15%. Or above 99.85%. Top 99% is the majority of all players.

[@TorNis7]

wtf, wasn't expecting to find you in comments here XD

[@ethanlocke3604]

nagi603 I died when I heard that 😂

[@Speed001]

Or above 99.85%, still we understood.

[@tri99er_]

yeah, top 99.85% means all of playerbase except lowest ranking bronze players.

[@Poo0fi]

Imagine this AI used in the campaign. xD

[@SpeakChinglish]

99% player failed to complete first tutorial level and raged quit the game

[@TehPwnographer]

@@SpeakChinglish Sounds like Driver lolol

[@theatheistbear3117]

@@SpeakChinglish Lore Accurate Zerg.

[@sushestvobezvolnoe]

Alphastar version one was just a legit cheat. One of the thing StarCraft is known for is microcontrol and multitasking. Alphastar defeated TLO and Mana by simply using the most basic unit and controlling pathfinding no way a human could do. Yes apm of alphastar was lower, but it was 100% effective. While humans mostly click just to check their status, alphastar knew it without a click. Alphastar just knew when and what by measuring the time in itself. Another interesting thing is that alphastar reacted to one pixel fluctuation from invisible units. So it knows where an invisible unit is and assumes it can't hurt it. The next version looked more human, but had other troubles. First of all it couldn't adapt fast, so several games in a row against one agent and a human would probably defeat it to the end of times. The next thing is that it couldn't react to dumb. The agent that defeated the world champion had lost to the kid that did a photon rush, which is like the most basic strategy in low leagues, but is rarely seen in proplay

[@WardenNFG]

Awesome! Was watching some stuff about this the other week, can’t wait until aliens use games to practice the perfect game vs humans and then skynet us

[@Marinealver]

Of course this isn't the AI Activision Blizzard is looking for. They are more interested in AI that makes you buy microtransactions. They want AI salesmen, not AI players.

[@IRMentat]

I remain unconvinced by the current definition of AI learning systems. It could be a mental block but it feels more like a brute force approach to mirror human experience at a faster rate but with an identical method of decision making, that they need a LOT of human time and input to deal with even fixed state systems is a cutting edge research reality (you may have 100000 lines of code acting but maybe 1000 is doing the bulk of the work, good luck parsing that with a small team). Impressive in its own way but still very much an expensive starting block rather than an actual runner. Then again, it’s a field where all you need it to get it right once (and recognise so) and the methodology will branch out exponentially from there point. Assuming of course that the current methods are not in and of themselves a form of scientific dead end.

[@omegaRST]

Learning systems will always require improving through trial and error (that is what LEARNING means), but I agree that current learning systems require a lot of computational power and waste a lot of time on mistakes. It is for this reason that most modern research is focused on finding the best way to teach these systems (look at the AlphaStar league with the exploiters) with two objectives: finding a way to learn better (get better results) and finding a way to learn cheaper (avoid wasting time on mistakes)

[@TheSheaim]

The problem with current AI learning is that an AI is still, while complex, based on a deterministic, digital system with a comparatively low (compared to a brain, and not just a human brain) complexity. Which means it cannot abstract what it hasn't simulated. It can extrapolate, but it's a limited functionality. Humans are capable of running "simulations" much faster, due to the pattern-seeking capability that machines are yet to match. That's why we can't have research computers that can get a new theorem that we can't even think of - only improve on the results we already have, decreasing the error margin etc. Creative thought is still beyond AI's capabilities and thus they will always be inferior in solving problems with no deterministic outcome. It may be possible to create a computer that beats every single SC player, even in any game imaginable, because most of those are deterministic or at least have known probability scores or good statistical data. You can't create an AI that will always win a war. Because humans are infinitely better in thinking outside the box, of which the AI is incapable of - it will always work within the boundaries of any system you put it in. A human can break the system - disobey orders, think of an unorthodox solution, initiate unexpected action etc... Until AIs get to this level they will never beat human ingenuity.

[@vir042]

@@TheSheaim Very good example, thats also why Florencio, a Master league player (not highest level), known for his VERY creative cheeses crushed alphastar while Serral (top3 in the world but very macro oriented) had a harder time with alphastar. Serral also didn't play with his own setup and hotkeys so that might have changed it as well, but the point is still relevant because I'm very certain Serral would crushed Florencia no matter hotkeys etc. When Alphastar comes up against someone that start the game with placing a nexus right in your base it instantly falls apart. :P

[@CardboardArm]

After watching many AlphaStar matches I noticed that the AI never really 'outsmarted' it's opponent. I would argue it's main strenght as a 'player' was that is was emotionless and stayed cool under pressure and relentless in it's agression. In a match against Serral it tried to contain Serral but was in a weaker position, however Serral had moments of doubt, indecision and pauzes to think, each one giving AS a small advantage untill it got the dominent position. There is not a single replay where AS outmaneuvered it's opponent, or tricked it's opponent, had a better build order or got a superior unit composition. It was never smarter, it just made fewer non-strategic mistakes. This is a big difference with chess and go and AS did not 'solve' Starcraft or surpass humans the same way.

[@DavenH]

This was a very well produced video. However, I'm not sure how all popular analysis of A* misses the insights on its super-human EPM in the first iterations that beat TLO/Mana. Putting the "top 0.15%" in context is also necessary. AlphaZero got well above the top professional ELO rating for Chess and Go, but nowhere near that in Starcraft, despite DM throwing zillions of resources at the problem. Ignoring that distinction misses interesting features of the hardness of the game and the true challenges AI still faces even on fully computable environments.

[@lrmcatspaw1]

It would have been a good idea to contact a starcraft 2 expert to talk about the part that are considered "unfair/inhuman" in the matches between the AI and the human players (one of them that they had to not use specifict strategies because the bots were not able to react properly, including certain units like dark templars if I recal correctly).

[@shatteredknight1129]

First professional test was against professional. Yes was playing Protoss, but he wasn't a professional Protoss. He played Zerg and Terran. Everything, but Protoss.

[@tryhardfailer]

alphastar was capped at a certain average apm, wich is why its so much lower than the pro players not because it was learning from players with lower apm

[@mimszanadunstedt441]

AlphaStar got a different AI per map or something afaik, and still takes an eternity to learn in real time.

[@Murtaskegg]

I think Alphastar has the advantage of being able to macro very effectively. Having watched quite a few of these matches, it seems to be human like in micro management, albeit with some incredible moves sometimes, but simultaneously is able to macro, and does not fatigue like a human does.

[@mrkiiv]

Gr8 vid

[@AlexanderDraconis]

AlphaStar Final TZP: Overmind 2.0 The AlphaStar League: Cerebrates . . . Us: food

[@LovusEternius]

I understood all of this.

[@ignamax08]

Pretty sure u didn't at 1:11

[@DavenH]

@@ignamax08 there's nothing obscure about state spaces

[@pedrenriquegg]

It would be nice to mention that at BlizzCon 2019 everyone there could try playing against AlphaStar Final. And the best player in the world, the most consistent Zerg on the planet, Serral, played 5 games against the AI and lost 4 of them. That is a big accomplishment for AlphaStar, bigger than Gary Kasparov vs DeepBlue.

[@Denestra74412]

I was there and I was able to beat the AI in a ~40 minute macro game and there were a lot of players that were platinum or diamond that beat it by cheesing the AI. The AI makes a lot of dumb mistakes ie when I played against it I had a lot of air units and it would some times land vikings in the middle of the fight.

[@chardonnay5767]

I have to commend you on the way you’ve set up your ad timings. It’s rare to see a video where an ad popping up doesn’t cause an immediate desire to close the whole video. I watch on a mobile app so there’s no need for anyone to recommend ad blockers.

[@AIandGames]

I like to think of ad breaks as mutually beneficial. I get a fraction of a cent and your brain gets a breather.

[@theqsanity6053]

What are those Brood War bots called? Looked cool. If like to look them up

[@InfoRanker]

Are they still working on the project? Will we see more of Alphastar in SC2?

[@vir042]

It's over, the final version was named "final" for a reason. It might be some people are still working on it but the large group have moved on, or at least that was the impression I got from their last few interviews they did around October/november 2019 or so?

[@7grims]

how many gigas or teras does this AI occupy, and how exponentially it grows in disk space after learning ?

[@ccrenshaw6184]

Alphazero didn't just reach Grandmaster status in Chess. It is the world's best chess player by a large margin.

[@dgam4211]

you cant be the world best in a board game by a large margin...

[@prophecynewt2250]

You can have a higher skill level by a large margin.

[@Annatar0]

I'd like to see AI development like this in FPS games as well but there are none that I know of. I know from back in the day when Quake 3 had custom bots that would learn and adapt by storing data of events in the game. There hasn't been any big leaps in FPS AI since Quake 3 and Unreal Tournament 1999.

[@josedelapinio]

Awesome voice ! (And video ;-) )

[@ricardoariel1239]

Last month, the world champion Serral tried a build invented by AlphaS, he lost, but that pro players learn moves that the AI ​​develops is incredible.

[@CG-eh6oe]

How the hell can Deepmind claim that the unlimited vision of the first version was insignificant? Mana immediatly won vs AS once its vision was limited...

[@AIandGames]

Notice how DeepMind won't announce the findings of any of the work unless they have some big talking points to swing around? It's scientific research meets big buzz PR.

[@CG-eh6oe]

@@AIandGames I aggree on the scientific research meets big buzz PR part, but thats exactly what is so confusing: How did they not come up with a better lie? Everyone who saw the first ten matches immediatly saw how important the vision limitation was.

[@Vanddark]

The limitation on vision introduced almost of a bug to the AI, they were reacting with all their army to a single drop, vision limited reduced drasticaly the AI's calculation of the risk of pulling all your army and losing positioning that is one of the fundamentals of the game. And it was easely exploited by Mana.

[@Blitzkit]

Just think if you can have a very slow efficient APM and can spike APM up fast and efficiently and return back to your normal APM without any backlash to your mechnics.

[@aw4955]

Impressive. Now let's see Google build a Supreme Commander or Total Annihilation grand master. Then let's see it win a 3v1. I'd love to see that

[@VB-92]

Hope I get to see an AI play an RPG one day.

[@CubemasterXD]

Amazing video as allways! If theres more info now: Could you maybe do a video on Elon Musks Dota AI? in your 2018 video on MOBAs in general you said "there isnt much public information on how it works yet", did that change? Is there another MOBA AI im unaware of?

[@mohamedyasser2068]

since 14 days of training is equivalent to over 200 years of human training, and it still can be defeated in rare times, something is really missing, the human brain way of learning is much more efficient so and if we somehow are able to memic this mystery learning algorithm the human brain uses, I'm sure within hours of alpha zer training we can get a bot that can beat the 44 days trained bot in every single match

[@0x0404]

Alphastar needs some value for committing to something. Just another value thrown in to the blender but it should help stop those situations where it is using energy to speed something along then it just cancels it when it was 95% done.

[@chonkyduck]

so they gonna add a new difficulty into the game

[@hohhoch3617]

They actually just recreated ladder. You can get to a competent level if you only focus on basics. But to rise higher you need to form coherent strategies. Then those strategies need to hold up against cheese or you need too be able to spot the cheese quick enough that you can change stratagems. This leads to thinking on the fly and changing builds according to the game state so you don't lose simply because you lost the build coin flip. Couple all of this together with the ability to optimize your strategy, excellent Micro/Macro (read: Basics), and the ability to adapt to a patch and you've got a grand master. Everything they've got going on already naturally happens, leading to the best rising to the top. Which is really cool. It's sort of a form of validation. These AI are meant to learn in a way similar to a human, so putting them in competitions similar to ladder, makes sense.

[@dgam4211]

if only it was that easy

[@googleyoutubechannel8554]

I can't help but feel AlphaStar was overhyped by DeepMind, if you watch the replays, it becomes clear that AlphaStar's playstyle was quite brittle, wins came mostly due to 'brute force', and it had not developed a sophisticated understanding of the 'metagame', which, tellingly is called 'IQ' by starcraft players.

[@dannyf1168]

unfortunately, macro will always bet the ultimate strategy.

[@nicoenerlan724]

Imagine if AplphaStar can micro every single unite. GG

[@Music-ym3ft]

AI and Games

[@Phagocytosis]

I would be very interested in a video on MuZero, its community implementations and what the future may bring, if and when it will become used and usable for any arbitrary game, that sort of thing... but that might be too vague a topic for an actual video. Still, thought I would put it out there. EDIT: Some of this is addressed in the video after this one (84VVFxHAdSY), which I just also watched. But not the specifics of what I was talking about. More on topic to the video at hand, this made me wonder... did they fully remove AlphaStar from ladder, or is it still out there? It would be cool if they left it on there (still playing only against those who opted in, naturally), and if they would also activate its ability to continue to learn from ladder games (perhaps occasionally aggregating the gathered data if real-time updating is not feasible), rather than just from games played in its own league.

[@mattmagliocca956]

I wonder if DOTA2 would be considered easier or harder to build an AI for than Starcraft? Less micromanagement but much more communication and coordination between different players. Different builds and selecting appropriate heroes could also add complexity. In starcraft you can pick basically any race and have a reasonable chance of victory, however DOTA requires you and other intelligent agents to select characters to balance against your opinions. Combat also requires a lot more coordination and timing

[@PureMatteo]

DOTA was already broken by a grouop of 5 AI. All they need to do is outplay the opponent in creeping which is basic micromanagement, and then snowball. It looks like the coordination and timing is not hard to calculate for the computer either, when all your teammates do better 1v1 on their lanes at some point you just have too many levels over the opponent and are overstuffed and it calculates when to push. It was done by another team than deepmind you can google it. It was first try as well. As complex as it can be for human players it doesn't have the inherent specifics of an RTS game, 1V1, with an empty canvas on a new map. Also remember MOBAs play on a single map removing a layer of complexity.

[@coot33]

A lot of Alphastar advantages comes form it's better interface with the game world compared to humans players. Same with the dota Ai.

[@AshValen]

In all fairness, in real world uses of this the AI will likely have better real world interfaces considering they will be able to assume whatever physical body design is needed for the task.

[@coot33]

Lots of hard work are currently going to have Ai control mechanical objects. This sorts of mechanical intelligence is hard things like hands, feet, walking, etc. They are hard to train as well as the transfer from simulated computer training to real world has a lot of noise.

[@brofenix]

@13:22 uhhh..........what is up with that curve of TLO's APM. I doubt TLO actually maintained those high APM numbers for a significant amount of time and that average of 678 APM seems way higher than his actual average. Maybe it's accurate but it doesn't seem like it to me.

[@trestenpope286]

ALPHASTAR POGCHAMP

[@dinopacitto2986]

we are A.I., it has mastered us and we are A.I. and A.I. is us.

[@LuoSon312_G8]

Scientists creates AI Gamers: the most revolution has begun Scientists: can we combine Halo wars 2 Ai with AlphaStar? Gamers: pls no. Make it stop.

[@razvandragomir7403]

All in all this is an impressive work but for what future purpose?

[@AIandGames]

Like I say early on, the idea is that these techniques can later be applied in other problems that are not games. Games are just easily accessible platforms to solve complex problems. In fact since AlphaStar DeepMind has published new work where they use similar tech for projects that predict protein folding and detect breast cancer from mammograms.

[@redeamed19]

When alpha star gets uploaded at the pentagon and takes over all the Terran military drones...do you want to play a game?

[@VictorF0326]

I wonder if AlphaStar would take GM#1 over Serral if it went online on EU ladder.

[@FiguraCinque]

Alphastar already showed superhuman unit micromanagment vs Mana. I bet the latest agents that run against the grandmasters are intentionally kept on a human level for some not quite clear reason

[@rompevuevitos222]

Only the APM is being limited atm

[@doberman7]

10:40 ....200 years of human experience, I don't understand that analogy, could somebody explain?

[@Ugnaski]

The AI has 200 years of playtime. That's achieved by "playing" multiple games at once. So imagine a player plays a 30 minute game. He gains 30 minutes of experience. The AI plays 5 games simultaneously each of which is 30 minutes. So it gains 150 minutes of experience in the same time span.

[@evilseedsgrownaturally1588]

Ivan Rubio ... Imagine you play SC2 for 200 years, non-stop. No toilet-breaks, no sleep, no eating, just SC2, 24/7, 365 days a year, for 200 years straight. Now you have 200 years worth of game related data. Now imagine AlphaStar downloading all of your data in 1 day. Now AlphaStar will play Starcraft 2 on a experience level that took you 200 years to obtain, but took AlphaStar 1 day to obtain.

[@bannanabean8792]

He does know that out of the top 3 players in the world only 1 of them is Korean right, like the considered best player is finish the second is Italian and the third best is Korean

[@Fine_Mouche]

so zerg is the better race asumle to A.I ?

[@supersnizelz]

AI was not apm capped when he first played against mana that's why he lost it was stalker spam with unreasonable micro

[@charleouel9012]

Well it depend alot on what they want the AI to do (AKA if they want it to win or if they want it to play humain like) Doing the 2 thing would be hard to do since humain are ultra complex haveing choice made simply by how cool you fell of what you eat or stuff like that. And alot of Mistake being made not on purpuse but to learn things out. Like AI do I guess?...

[@erikm9768]

Pretty lame by Deepmind how they never updated this subject again after all the goodwilll of the community and Blizzard.

[@danielalorbi]

not their goal, they moved after go as well

[@Speed001]

18:36 lol

[@arconsWoWLPs]

At 13:15 you say each Action is useful. Thats not true. Especially in the beginning of a match you spam quite a bit just to get yourself going and especially when it comes to TLO he spikes at over 1000 APM due to a bug/feature in the game. Therefore the effective APM(EAPM) is quite a bit lower than the APM. For the AI I'd assume that APM=EAPM

[@prophecynewt2250]

I recently watched a professional sc2 caster estimate Serral, the number one sc2 player in the world, to have about 240 EAPM. Its probably not accurate enough for weighty discussions, but gets the point across.

[@javierguardiola7122]

APPLY THE AI TO ZERO-K it has an epic micro to me oppinion, the macro is nice too, works with lua scripts (easy to write new balances), and the terrain is DESTRUCTIBLE, ALL the terrain uwu

[@Hohoho701]

The real question is "can it beat a korean"

[@ktktktktktktkt]

16:47 "Of the approximately 90,000 players who play starcraft 2 on the European servers, that places Alphastar within the top 99.85% of all ranked play." Hmm... sounds like the bot wasn't very good after all...

[@michaelchaney2336]

2^1650

[@fablesenpai8609]

O damn 3klik is a supporter. Hell ya

[@xD-lt1mq]

200 years aka 1 riot employee

[@bigboydancannon4325]

I don't trust Google one bit

[@AIandGames]

Shhhhhh.... they'll hear you.

[@furinick]

They probably just hired a korean and pretend it is an ai

[@Vit-Pokorny]

in few years from now google will builda droid factory and conquer the world.

[@aandredaandred3378]

10:56 It's as far from the correct pronounciation of his name as humanly possible. You could say "styrofoam cup" there and it wouldn't be much farther.

[@skinnyboy996]

How long will gtx1660 learn this stuff? That"s all I can afford

[@LightningbrotherG]

You could probably train AlphaStar 3 years give or take :3

[@cam-jm1fu]

The problem is that they don’t tell players when they are playing agains alpha star. So it kinda rigs it.

[@minhucovu6321]

Alpha Star is certainly impressive, but its decision making always seemed dull. It always go for the brute force approach, which makes sense, considering its best race is protoss;

[@tomaszurbanski3057]

Sorry, but it sounds funny to me when you've read Komincz like Kominktsch. You were so close, just drop second 'k' and will be way better ;)

[@AIandGames]

Y'know I knew I was gonna screw that one up. Need to research better!

[@Toxicflu]

Alpha star last year won in unfair matches vs tlo and mana. It had unfair map vision, and unfair micro. Those two things make alphastars basic strategy of going mass stalkers easy. Had it gone vs another race, mass stalkers is easier to counter. Deep mind still hasn't conquered starcraft. The fact that it's gm and not #1 Gm proves that it has weaknesses that humans will take advantage. Why aren't there more recent games being uploaded about this? Why are we still talking about it's first wins? This is all old news.

[@ComradeSkrald]

The only unfair thing about it going mass stalker was it could get so much value with them since it would blink away not that they dindt have an easier counter. It might not have conquered starcraft 2 but it did still beat a few still active top pro players, and their sample size is not too big, so I will be hard to say if it could have won against number 1 on ladder. They have uploaded the most recent games if you are talking about it playing on ladder? We are still talking about its first wins because its a hugh breakthrough

[@gabrielandy9272]

the newest versions of it fixed the vision and micro tho... and still won against other players.

[@hyouzanren1846]

😳😳😳...? Teaching an AI how to waging war games will NEVER end well!

[@Damaged7]

"APM, the number of valid and useful actions..." Most of the A in APM is just useless fluff designed to pad out the number.

[@seamon9732]

Is it still cheating with inhuman EAPM bursts during fights despite having an acceptable average ( which means nothing )? If so... meh.

[@zaxtonhong3958]

The later iteration used action per SECOND limits, rather than actions per MINUTE. So the inhuman bursts were drastically reduced

[@seamon9732]

@@zaxtonhong3958 Good

[@JohnTrustworthy]

The bots reached GM but nooooooo, Google couldn't possibly have released propaganda bots during the meme war.

[@bmlee9113]

You need Korean professional gamers for test.

[@anmolpatel793]

Is it necessary to teach AI to be a strategist commander before we teach them how to not walk and talk awkwardly and then they call me mad for prophecizing the end times

[@greenlemon9155]

So, fast calculator that copy human behavior ... cool and stuff

[@Luzarioth]

what i would like to see is an unchained alphastar. No APM restrictions (which where in place against TLO & MaNa) ! show us that perfect micro xD

[@1954BadCompany]

At 16:34 you say "the top 84% of all human players". You should have said "the top 16%". You repeated the error at 16:59.

[@AngelTyraelGM]

Man you kinda missed the point with apm, realisticly i expect bots to have much less apm but have near perfect timing and decisions when to do it, most of the pro league apm is trash apm it doesnt do anything it simply helps with focus it feels good to do, example in any micro battles between armys we as humans like to click alot like crazy even tho it doesnt matter we still go in the same direction, an ai can just click 1 time and then another if any actual change needs to be done