AlphaStar: The inside story

[@wdai03]

The alpha go documentary was pretty good, this would make a nice sequel

[@robotec0007]

there is already a documentation about alphastar

[@xv179]

@@robotec0007 whats the link?

[@alexhp26]

@@robotec0007 Where indeed?

[@Demon-vq3pr]

That was a AI vs whole country thing

[@pmcg534]

Is there? What’s it on?

[@russellthorburn9297]

I was gratified to see how gracious the players were in the face of such a defeat.

[@edwardyoung8241]

That really impressed me too. Although, maybe they edited out all the swearing!!! Hahah

[@benjaminlavigne2272]

"professionnal players"

[@Hazlitt123]

Idra would have rage quit game 1

[@SnOrfus]

These guys got to be some of the best in the world on the backs of hundreds of losses. If course they would be pros about it.

[@clickpwn]

in a game of starcraft, you learn more in games that you lose.

[@userNo31909580]

I'm not yet completely satisfied in Alphastar. 1) The APM cap has to be changed. Alphastar had lower mean APM than the human pros but in engagements its APM shot up to 1500's. That equals 25 actions in a second. No human player is capable of doing this with great precision. The APM should be capped in two ways. Firstly, the mean averge per minute as they did in this demonstration but more importantly, Alphastar should be restricted in such a way that the huge spikes are not possible. Perhaps a limit on how many actions it is allowed to do within a second? 2) Especially in the games against Mana, Alphastar demonstrated inhumane ability to control its units. It won game 4 because of how clean its micro was, not because it outplayed Mana. Good army control is impressive and clearly requires a certain level of intelligence but this kind of perfect blink stalker micro doesn't really exist within the game. Deepmind should differentiate between the agents intent and execution. Perfect micro is more than just speed. Precision is much more important. Human pros missclick all the time. There needs to be a random element in how Alphastar controls its units. 3) The agents seemed to be stuck on the strategies that they had trained in. I'm not entirely sure if this is the case but I think there was zero tech switches from Alphastar. It seemed to be stuck on the army compositions it initially chose. In a few games the composition was clearly sub-optimal considering what the human player had. Reacting to the scouted information and actually being able to flexibly adjust strategy on-the-fly is exactly the kind of intelligence that Deepmind should strive towards. While Alphastar is impressive and Deepmind limited its mechanical ability a little it was still very clear to someone familiar with Starcraft that it won mostly because of its superhuman mechanical ability. It could perceive all of its own units at the same time, it had superhuman mouse precision and macro abilities. TL;DR Cool but not quite there yet. Alphastar won because of superhuman mechanical abilities. I want to see a bot that plays as slow and as sloppy as the average pro player and still be able to achieve >50% winrate against the best in the World.

[@ChristianSpliid]

great observations. thanks!

[@neoqueto]

I think that still the #1 thing that makes it not as impressive to me is the fact that the AI didn't have a visual input, or seeing what's displayed on the monitor, interacting with the UI - they basically modded the game so that the agent controls some triggers and reads a bunch of variables.

[@userNo31909580]

They will most likely never train an agent using only the unmodified visual input because the game simply can not run fast enough. These agents needed the equivalent of 200 years of gameplay. Playing SC2 in realtime just takes too long.

[@diradosveta]

I wonder if Alphastar would put its tl dr to the end of its comment

[@MusicPerusal]

@@neoqueto Well yeah... It's a computer.

[@TheAIEpiphany]

We need a game called Alpha so that DeepMind can create AlphaAlpha in a truly alphaic way.

[@xordux7]

or a game called BetaGamma

[@marneninagavenkat7149]

You're there now, expecting that from you 😃😃

[@dr_flunks]

Congrats DeepMind. You're an impressive team.

[@ClipCourt]

True that really impressive indeed.

[@elimurphy-trotzky1231]

Being able to se the whole map at once is an incredibly massive advantage.

[@loremipsum7513]

They do cap it when it play vs pros

[@dannygjk]

Yes but it had fog of war iirc.

[@ZLO_FAF]

@@dannygjk yes, but it can still see cloaked units. at least in the early versions. (there was review or a replay that showed AlphaStart instantly queuing an observer after it has seen a DT for like only a 1 second in it vision range. and APM for early versions was very hight. as much as 1500 apm with stalker micro is deadly, and it can select units any amount of units in whatever way it wants in very non-human-like way. in games vs TLO it selected only parts of army to target-fire TLO's units (carriers?) and it selected absolutely random units (can't do that with box selection) and selected perfectly enought units to one-shot TLO's carriers.

[@dontgetmadgetwise4271]

@@ZLO_FAF Yes. Imagine how good it would be if the human constraints were lifted?! One second of visibility is a lifetime for a CPU.

[@jessiejanson1528]

@@ZLO_FAF If their goal is to be the best StarCraft player, they failed miserably because they essentially are cheating. if their goal is to use StarCraft to make a better AI... only they could know if its happening. But increasing the difficulty to be more human like would certainly push the systems limits and force it to become better at playing. Hard to be impressed with it when its playing on easy mode.

[@AG-dj6bc]

For the sake of us, i only hope that your ia safety team is the strongest of all.

[@rodrigoserafim8834]

I am glad our AI overlords are learning so quickly to defeat our puny human armies. I also hope they learn to read youtube comments so they know how much us loyal subjects love them. bº*r#i+n$g h)e]l&p sorry for that, must have pressed my elbow against the keyboard.

[@OsvaldoBayerista]

X2.

[@t-.-t.]

XD

[@zissler1]

I can’t wait for them to take over I’m sure they’re intelligent and will let me be an ambassador.

[@marianpazdzioch6632]

I like how Alphastar worked around APM limitation basically by saving up his available APMs for when they are needed most and then use them in bursts :)

[@ManjulaBandara]

Came here after the AlphaGo Full Documentary. Impressive stuff!

[@polesetter_01]

Please make this a full docu as well. Alphago was one of the best docus of the past years.

[@raidiniom22]

Alphastar vs OpenAI

[@martiddy]

That would be great to watch

[@AdityaSingh-lp5rp]

With unlimited power

[@tonglu3699]

Don't you give them ideas! If they end up learning from each other, that would be the end of us.

[@ivoriankoua3916]

@DarkGrisen if you really think so you're funny , you do think serously Google is going to invest $500Million on a company just to play game ....think a little , I could give you clue if you want but try to think deeply about it.

[@dsm091]

@DarkGrisen I am thinking the machines they are using is their quantum beast letting it learn all it can. It needs to make them money. Micro targeted ads to a single person is their goal.

[@bigchunk1]

This work inspires me to get into computer science.

[@Oscar-if6lq]

Getting into mathematics first is probably more important than Computer Science if you want to do something like this.

[@jeremywright9511]

@@Oscar-if6lq Very late comment but yes. Math is foundational but computer science is a great extension upon already established knowledge. Given the option, I would feel computer science is more approachable to the everyday learner than an emphasis on mathematics.

[@tomchrysostomou4895]

Brilliant, brilliant, brilliant, Deep Mind has done it again! Well done, very impressive!

[@kwanarchive]

The thing about AI that people underestimate is that the volume of automated training can quickly help an AI to overcome the mistakes in its learning. With humans, it's very hard to convince someone they're wrong, let alone change someone's mind, and often you just have to breed new humans and hope the old ones die off fast enough so that their mistakes don't get meaningfully passed on to the next generation.

[@ClipCourt]

Major Fan of AlphaStar overcoming programming complexities for a better future.

[@anilkumarsharma8901]

Are you able to differentiate any equation🔔 upto infinity???

[@krallis123]

I think it's extremely cool what you're doing but I'd be interested in knowing exactly what reaction time you used for AlphaStar and also I think you said it was performing at 500APM, I think it would be interesting to see how it changes and adapt when you turn down the APM. How does it behaves at 450? 400? 300? I think there are a lot of extremely cool observation that can be done by looking at AlphaStar and especially kind of see where the victories come from, of course it will be a mix of strategy and mechanics, I'm curious to see what heppens when gruadually limiting more and more the "mechanical" skills (APM) of the machine. Anyway really solid work, keep it up!

[@synchronium24]

@1Alino"APM is not a configurable parameter for this AI. " They said in the video multiple times that there is a cap on AlphaStar's APM.

[@solidpain9098]

They have limited (or trained) it to have 300 APM overall (it can peak if it needs to, but it has to get an overall APM of about 300). Without any APM limitation it would beat ass a lot more.

[@benjaminlavigne2272]

300 average apm

[@Loppy2345]

Vote up if you want to watch AlphaStar vs Serral

[@o0NeoTDN0o]

The AI doesn't play against Terran or Zerg yet.

[@soulclean1983]

February 15 2019 is the day.. for this great battle

[@AdityaSingh-lp5rp]

There is one

[@AdityaSingh-lp5rp]

kzbin.info/www/bejne/lXOyoWRmq52eq68

[@dosomething6975]

Keep up the good work! You have no idea the smile you bring to the faces of those who love to see technology advance! Thank you for all the hard work! Love it!

[@triplea657aaa]

I know it had superhuman micro, but let's talk about how amazing it is that it plays human strategies so well, and plays like a human.

[@calgar42k]

Let s just hope AI never do moral judgment like humans are doing! Because your value could very well be considered 0.

[@PereCatou]

@@calgar42k lovely...

[@crypto_div]

This is like watching the scientists in Jurassic Park hold a baby T-Rex and not realizing that one day it’s going to grow up and eat them.

[@ReasonMakes]

Alphastar appears to have won because of its superhuman mechanical ability to see and control its units at all times rather than a superhuman mind. Mouse movement speed, visible units on-screen, Actions per SECOND, precision, reaction time - all of these mechanical abilities need to be taken into consideration much more seriously/restrictively if we want to develop a Deep MIND rather than a technique impossible for humans to execute. For me, the excitement around AI is the intelligence, not the aimbot. Lovely project and clearly a lot of excellent work so far but personally it looked very unfair to the pros.

[@bobedwards8896]

Agreed. And I think it's pointless to compete an ai v human. If course the ai will win, the game itself only has so many viable styles and ways to win. It's not hard to find out those ways. Especially if you can do hundreds of thousands of games to learn

[@Moonz97]

AlphaGO achieved the deep MIND part, where mechanical skill isn't part of the equation, but I agree its not fun playing against an aimbot

[@ReasonMakes]

@@bobedwards8896 I think playing against AI is an amazing opportunity. The fastest way to learn is to play against people better than you, but for the pros there is no one better than them. Could you imagine how beneficial it would be if you got used to playing against AI that is to pros as pros are to mid-tier human players? You would become superhuman in your strategy. But only if the AI is playing by the same rules of our world. In real-life you cannot move your hands and fingers fast enough to achieve the 1000+ APM that Alphastar exercised at some points, so the AI can use that unfairly and it can also enable the AI to be lazy with its builds. It would be very educational for human pros to see what an AI would do with the same limits that humans have in terms of strategy rather than relying on things like flawless micro

[@ReasonMakes]

@@bobedwards8896 I would also like to add that SC2 is similar to Chess and Go in that although the rules are finite and relatively easy to understand, actual strategies are endless. Perhaps there is an OP strategy that we simply have yet to realize. AI could pioneer those for us since it is able to play hundreds of games simultaneously and gain more experience - and insight - than a human ever could

[@AlfalfaQc]

@@user-yd6qk6nk2i It sees the complete map but only what's actually visible to its units and building. It doesn't see through the Fog of War. I agree it's still a slight advantage to be able to see the whole map but it's still incomplete information as in it doesn't know what the opponent is doing.

[@TheNightmoose]

I can't wait to see the progression! Sooner or later it should be able to play any race, against any race. It would be interesting to see if it prefers playing a certain race, or if it perhaps chooses dependent on certain maps or opponents etc

[@clementfouquet2483]

Congratulation on that achievement DeepMind team. Can't wait to read more about your model when you'll release a paper later. I wonder if you developped a macro strategy policy that gives short-term objectives to a micro policy ? or if you have only one policy network coupled to a LSTM for memory ?

[@thejuiceweasel]

Imagine a grand tournament held each year where the best players in each e-sport try to beat such an AI. I'm sure that would sell some tickets.

[@caulijutsu1575]

They shouldn’t allow the computer to see more than what a player can. Make the computers only input the output of the screen.

[@dannygjk]

Done.

[@jonathanwalther]

Model human decision making. Model an army. Model a map. Use enough time and computing power. The military is already knocking.

[@davidwuhrer6704]

The military is already obsolete.

[@jaidenrichard99]

I can see overfitting problem of undetermined (hidden) situation on every replay game. How can solve this.

[@akamezo]

Now it is the time to see the clash of A.I : OpenAI Five VS DeepMind AlphaStar .

[@xxxx85]

1:48 - Do they mean that AlphaStar can see the entire map? Or do they mean it sees its own entire map? Because if it's essentially maphacking, it's not as impressive. I really hope it's not, that it's scouting and needs to draw conclusions from what it sees. Also, shouldn't Brood War be even more of a challenge? PLEASE DO BROOD WAR!

[@marty34534]

EVERY "AI" in every game ever sees the entire map, Deepmind or not

[@MsHojat]

@@marty34534 He's talking about fog of war. In addition, Alphastar can access the entire play area all at once; players cannot do that (they have the minimap, but minimap doesn't distinguish units, nor allows unit selection, nor shows health or other important things; they would have to use a zoom hack, and even then things would be too small for them to see and click. Alphastar doesn't have that problem because it doesn't have to click nor "see" (visually decipher/recognize) anything.

[@cwjalexx]

imo BW would be a lot less of a challenge for an AI because of the high mechanical demands that would be trivial for an AI. An AI would essentially be using auto-mine and infinite-unit select. This alone would lessen the burden on strategic superiority allowing an AI to crush human players simply by having more 'stuff' than the opponent.

[@nicoladube4175]

the best training coach for the pros out there in my honest opinion. Give them few weeks of fighting the A.i and let them loose against humans after; i predict a win for those who trained and fought with A.i

[@lucaslopesf]

I love this videos, keep updating us of the backstage

[@yu.h.c2051]

0:04 What language?

[@DarkDog09]

I'm pretty sure that's dart! Edit: I was wrong it's 100% lua!

[@abcdxx1059]

Shouldn't they be using sonnet or tensorflow ?

[@ZettaZone]

@Lucieon You define functions in python by "def name:" not "function". Also you don't end functions with "end" in Python. How some user with a weird name already said it is Lua. @Primo xx: Tensorflow and Sonnet are machine learning libraries/frameworks. Not programming languages.

​@@srt-fw8nh ​ no, it's Lua

[@swdev245]

@ Maybe Lua because of the Torch framework which is, IIRC, popular in reinforcement learning research

[@webcrawler9782]

maybe worth mentioning: Mana beat AlphaStar on the live stream match afterwards

[@zhenzhu8248]

@deepmind, can we see a game without any handicaps on the AI? Just for the curiosity. Clearly if you have to handicap the AI, we didn't get to see the Alphastars full potential.

[@subject8332]

I agree, we don't make self-driving cars distracted by talking on the phone or by limiting their sensors to simulate human behaviour and perception. That would defeat the point. Biological systems don't have machine accuracy, but machines don't have billions of years of evolution driven by natural selection behind them either. I am interested in the battle between the two systems!

[@namfromvietnam]

They are restricting its mechanical ability, not intelligence. Without that restriction, it will rely on mechanics and not develop intelligence/strategies. A more obvious example is shooting game, if you just make an 100% accurate aim bot (mechanical ability that is impossible to human), then it'll easily win every game without caring about strategy or any other aspect of the game, that'll be boring

[@zhenzhu8248]

When a grandmaster league player play against a silver league SC2 player, they should put an APM cap of 30 to 50 APM to make it fair? I understand your point of view, make it a strategy match not apm match. But there's no such APM limitation in SC2 matches in real world, so alphastar's effective APM is fair game in my opinion.

[@namfromvietnam]

@@zhenzhu8248 It'll be very boring, and nobody will care. Are you gonna be impressed if an aim bot beat you in a shooting games? In fact, if that is the case, there is no need for alphastar, some random bots out there probably can beat human. Do researchers care about some silly games that nerds play, or "fairness" or "try to win"? No they don't, it's not the point. The whole point is measurement of AI progress. If I were them, I'd go so far as limit its mechanic bellow human to push it develop new strategy to make up for the disadvantage, now that will be interesting and get press attention

[@zhenzhu8248]

I just want to see what an unrestricted alphastar can do Lol. You want to see a restricted one under your terms. Let's just put it that way. We want to see different things for our own reasons. Pointless for you and I to add any more after this comment

[@williamgoo7621]

If only TotalBiscuit is still alive today to cast this match.

[@antoniolewis1016]

:(

[@momololo3223]

He's casting in Heaven!

[@VramsGamingChannel]

It's amazing how calm these pros are when they lose. I get flustered and upset and if I lost 5-0 I'd feel humiliated. They let it slide off like water to a duck's back.

[@dannygjk]

It's exciting to them to compete. You cannot reach your full potential if the most important thing is winning.

[@nahblue]

@@dannygjk I love that they don't give up when they are down either. They have a strong mental game, and they really want to win (just look at Mana going at it in the showmatch), but handle the losses too.

[@B1gLupu]

It's almost like they do it for a living and are used to losing to top players

[@deadwingdomain]

Have you ever played the AI in star craft? It always been brutal.

[@rodney7150]

Oh my god, this video gave me the shivers. Awesome

[@flytalkers]

What's the background music?

[@sibyjoseplathottam4828]

Great work. I think this will have applications in power system optimization.

[@christopherarmstrong2710]

1:40 “Alcohol is not just to defeat these players. Alcohol is to do it in the right way.” Wait, there was drinking involved?

[@jessiejanson1528]

"our goal"

[@Godishus]

Dagnit.. now I can't un-hear alcohol. ^_^

[@adubizon2243]

But can it defend a cannon rush in 2v2 with a teammate unwilling to do anything but get BCs?

[@s.czerniawski789]

The comments about unrealistic APM/micro are missing the big picture. Once there's a system that can reliably beat pro human players (by "cheating" or otherwise) they can use it as a BENCHMARK to compare to. A benchmark which is much more convenient to use than inviting individual players. You can't get 100 pro players to show up on 3 second's notice and have them play consistently 24/7 for a week... Once they have that they can work on adding more human like constraints, if they find it useful. Also the idea that the AI has to be as constrained as humans are is just as silly as objecting that the first planes were allowed to be much bigger than birds and ran on gasoline instead of seeds. We're not trying to replicate birds or human players. We're trying to build something that does not exist yet.

[@ericnathanstucky246]

S A Y I T L O U D E R F O R U S I N T H E B A C K

[@illarionbykov7401]

Nope. It's cheating to grab headlines by big corporate/university teams. The same BS happened with chess. It took another decade for small, independent AI people to get fair wins against top humans in chess (compared to IBM Deep Blue's rigged so-called "win" over Kasparov) In the Japanese chess (shogi) world, the professional shogi association made strict rules to prevent this kind of BS, barring pro players from accepting rigged matches vs. AI, and forcing fair matches only, so it took AI an extra 2 decades to declare "victory" over top pros in shogi compared to chess. The difference of a fair contest can be that big.

[@dannygjk]

@@illarionbykov7401 Incorrect, I have been following computer chess since 1981 and your claims are false.

[@bebop417]

1:58 thats obviously not true from gameplay footage. It's average apm is lower for good reason but it still spikes absurdly high during crucial moments like combat micro. Also players use actions differently than a perfect ai, like spamming keys for example, so a lower apm for the AI would seem to happen naturally.

[@cjfletcher325]

It can see the whole map? That has to be the most important and difficult Challenge to over come IMO.

[@SanctusAcappellaSATB]

Did they set up reaction time, map visibility and clicks per minute as human?

[@jonnyhirse5949]

For this old version map visibility and apm peaks were unfair, reation time was ok.

[@pmcg534]

The alpha go movie was one of the best documentaries I have seen.

[@AzOuLs]

To many nerds together.I am afraid of what they can create next. Keep doin a good job

[@rosscof3660]

Amazing and scary at the same time...

[@pythonmini7054]

what programming language did they used ?

[@montanac223344]

You may not like it, but after 200 AI years of training this is the most advanced BM of all time 0:35

[@as1df1gh]

lol

[@xiangyizhang1849]

After watching the games, I think the limitation should be imposed on EPM not APM. As a fan who has watched starcraft games since 2009, I would say no one ever could do the micro-control like AlphaStar did to Phenix and Stalkers. For a human player, select more than three target units instantly is very tricky. When human players tried to do control like this, they might be failing to select any unit, where the action is counted into "APM" not "EPM". Numerous useless actions always happen during a game for human players. My conclusion is DeepMind considers the fairness in terms of APM's quantity, but they should have introduced another aspect which is the quality of "APM". The solution is simple, either switching to EPM or gives a probability (close to human players) of mis-operation.

[@adammbq]

You see confidence at 4:27

[@zengamer321]

The two big things are APM spikes that look normal but are completely impossible because of the precision and camera.

[@ryuzaki_ray]

They need to make AlphaMayweather.

[@RepsUp100]

They are coming!

[@fg786]

Can this AI be used for developers to balance their game before relase and for future expansions, or would the players and AI ability be so different, that the balance for the AI would in the end lead to unfair plays players would use?

[@cavalyn]

I think a lot of people are being ridiculous on this - and it's showed in Dota 2, Chess and Go for respective AI tournaments. They see it as a threat, insult it and make excuses for why it won and how unfair it is. Some even go so far as to call the people behind it cheaters. This is such a short sighted way of looking at things. This technology isn't going to be used to beat people at video games. It's doing that now because video games are a good way to test its ability to grow and lets them develop the AI based on something that has a clear outcome (win or lose) and is competitive. The real technology is going to be used for medical purposes, research, industry and so much more. It may well have the potential to change our lives and for the better. We should be encouraging that and the people who work on it. Not getting worked up because it beat people at a video game. Though, to be fair, I'd say the SC2 community has taken it better than the Dota 2 one did. Sheesh. Still, even seeing some comments here, it's just "It only won because..." and "these programmers are lying..." and "it cheated" or "it's pointless to make AI play against humans". Even if we JUST assumed it was only ever designed for playing video games, I'd still encourage it because wouldn't we all like games - strategy, rpg or otherwise - to have better AI? Instead of what you get in many games where the AI just gets a stat or economy advantage, etc...

[@jamiefishers895]

"The real technology is going to be used for medical purposes, research, industry and so much more. It may well have the potential to change our lives and for the better. We should be encouraging that and the people who work on it." True. But it could also be used as a weapon or tool for terrible purposes such as to damage, control, manipulate, or exploit. Also, the technology you see here looks quite harmless, but there are some who see it as the building blocks to something far worse and more chaotic. Unfortunately, man's desire to innovate something so powerful can only be delayed - never avoided (unless the world ends first of course).

[@vignesh3by4]

Hi, If we think they are doing all of these work only to be used for medical, research and industry then we have not even touched the base case, that is it's actual use in military intelligence. Why do u think they are starting AI testing with strategic games like Chess, GO, DOTA. They are testing how fast it can learn the enemy's attack pattern and come up with a counter attack and the most import use of all is to monitor the human internet usage. What they like, what they dont like and can alter even result of common searches accordingly. When AI test is successful, all of the companies will give admin access to these AI for accessing and changing their server contents. Think about all the threat that can come with it.

[@bawzzzz]

@Coeur Do not talk if you don't know what you are talking about. People pointing out obvious flaws in the implementation of mechanical restrictions are PRECISELY the people who want to see AlphaStar get even better at strategical thinking. You are portraying those people as the exact opposite of what they are. You are being the embodiment of "short-sighted". Being a fan of a technology is not being a "yes-man", but speaking out when there is something disingenuous being displayed.

[@cavalyn]

@@bawzzzz Thanks for a pointless comment. Here are some of the comments I was referring to when mentioning how pedantic people are being: "God it bothers me so much how they claim to have limited the AI's reaction time and speed to resemble a human. They are lying and they know it. " - Instantly suggesting the programmers are lying because they have nothing better to do than want to win at a computer game. "The solution is simple, either switching to EPM or gives a probability (close to human players) of mis-operation." - Trying to find ways to worsen Alphastar because it being more fair at a computer game is more important than it being improved to be a valuable tool. "They shouldn’t allow the computer to see more than what a player can. Make the computers only input the output of the screen." - More on how it should be limited for no other benefit than to make us feel better. I.e. no improvement to DeepMind's mission, etc. "Why did you feel the need to lie about AlphaStar's mechanical limitations? Everybody in the Starcraft community can see that the level of micromanagement that it uses is unattainable by human players. Everything about this is really cool, but this sours it up for me." - Again, just more focus on perceived unfairness. I never once suggested - as you seem to think - that people pointing out mechanical flaws was the problem. As I said in my post, my issue was with people attacking AS and the team behind it for their work.

[@lawrencepatrick2486]

When AI takes over people's job eventually it will be different story.

[@rodafy9110]

absolutely incredible work. would love to see this taken to a team-based game like dota 2, and see if a machine learning bot can climb to the highest rank in the solo queue of that game too. presents a new challenge with the addition of having to cooperate with unpredictable and unreliable human teammates to achieve victory, in five different possible roles with hundreds of heroes and millions of unique team configurations and items and so on.

[@leematthews6812]

I've been running Fold@Home on my PC for years, and AlphaFold has leapfrogged over any previous efforts by a wide margin. As with any technology, the applications have good and bad prospects. Let's hope the good outweighs the bad in future.

[@abcdxx1059]

0:04 ?? Not tensorflow or sonnet what are you guys up to ?

[@parskatt2971]

What language is that? I don't recognize it.

[@icedanilzation]

What game would be next level up for ai?

[@rangv733]

Can someone tell me the practical benefit of developing such AI which plays games well?

[@Hamsters831]

when is the serral vs alphastar match?

[@eddiegohwj]

So many ways the game could be rigged to the AI's favour. This is not Go or Chess, whereby reaction time and visual perceptions weights so much in real-time strategy game. Use BostonDynamics and build some arms & eyes, and play on the fair ground & same inputs (mouse and keyboard) as human players. Dont even have to limit the robot's mechanical capabilities (APMs and response time), and i will be convinced.

[@davidwuhrer6704]

They certainly would have to limit the robot's mechanical capabilities. Robot actuators are much faster, stronger, and very much more precise than human motor control. Plus, reaction time: Human reaction time is on the order of 200 ms. Robot reaction time is below 1 ms.

[@ZLO_FAF]

@@davidwuhrer6704 Robot will still have to check what are those dots on minimap. and will not be able to spot cloaked units while his screen is elsewhere

[@yorkshire_tea_innit8097]

Alphastar definitely plays to his strengths. Which are micro and multitasking.

[@trumpsama3510]

I'd like to see the AI try to play a team of pros at the same time

[@o0NeoTDN0o]

AI vs 4+ pro players?

[@bobcat5844]

Id like to see them play top Korean Protoss players

[@HidingOverHere]

Id like to see a single AI with different strategies instead of 10 different AIs like they did in this demonstration

[@skrrpapa2936]

2falconpunch There is a huge difference between MaNa (Top 3 foreign protoss) and a player like Stats (best protoss in the world) (or sOs/Zest). I guess you arent into the sc2 pro scene but the difference in skill is actually massive especally between these 2 (MaNa - Stats)

[@skrrpapa2936]

2falconpunch What does this Change about Mana not being even Close to the Level of Stats??? Like i didnt even say it's Micro was human lmaooo

[@naskastark6730]

3:40 exactly what Lee seedol said . Lol

[@socrates_the_great6209]

Where can we see the replays?

[@XOPOIIIO]

Even if you can make the same amount of clicks, you're still have an advantage by being able to think 1000 times faster. It's not the amount of clicks that matters in Starcraft, it's how valuable and precise they are.

[@kvazau8444]

that just shows the intelligence is better, which is the goal.

[@bighands69]

@ХОРОШО The processors are working thousands of times faster but not the decision making which is at human levels.

[@XOPOIIIO]

​@@bighands69 No, it's not. It's still superhuman in speed. Neural Network doesn't learn to be fast, in fact from it's subjective point of view the game doesn't run at all, it just wait's for it's decision.

[@bighands69]

@@XOPOIIIO The ai works by using search algorithms and those were at human levels of search. DO not confuse the cycles of the hardware with the actual decision making of the ai. Humans because we have a visual cortex can search large numbers of visual patterns and add weight to various parts of the search the ai does not have that advantage.

[@XOPOIIIO]

​@@bighands69What do you mean by human levels of search? Is it human level to you to play 200 thousand ingame years in couple of hours? Because that is what they actually did. AI have visual cortext too, it's called Convolutional Neural Networks, just not this particular AI, because in this game it's useless to see units in all their animation instead of a single point on the map, so they simplified it.

[@desven123]

I would love to see some games AI vs AI, just to see what happens

[@johannes-7710]

Hello Deepmind Team, shouldn't now the next step be to create an AI that could learn any game in the world in 24h. So any Top 100 Player of a video game could come and challenge it for a 1v1 anytime?

[@lucaslopesf]

What will be the next DeepMind's move?

[@abliviax]

Impressive deep mind team, this is very cool and exciting! I think the improvement between TLO games and Mana games were improving the mechanics to beyond-human levels. Do you plan to scale back the "effective APM" to more human levels now that it is hitting like 1200-1500 apm (mostly effective, I assume?) in one reply I watched.

[@dannygjk]

Done.

[@AngelLestat2]

*Do not call victory yet* Alphastar needs to wind using the same information input and control output than players use, that should not be hard to achieve using different neural networks for visual recognition and interface to what alphastar knows and another neural network to link proper mouse movements to what alphastar expect to do. Without that, we never know how many action by minute is alphago saving or if still doing some kind of cheat by map focus knowledge or minimap. You can use all the processing power you need (it would not be much even with this addoms and you can keep training without these ANN interfaces). This will allow alphastar to play a normal version of starcraft 2. *Challenge Accepted?*

[@MsHojat]

I don't think that's going to happen, especially if it involves using a physical keyboard and mouse rather than simulating mouse and keyboard inputs. Although the visual recognition would probably be the worst/most-challenging thing to deal with.

[@AngelLestat2]

@@MsHojat it is easy to achieve, there are already thousands of neural networks that recognize image and even context in those images, and they add a description in text like "2 kids playing with his dog with a ball", and this was accomplish in 2015, here is way more easier, the neural networks just need an interface to link what it already knows, like units, buildings and terrain features. Trust me, that is super easy for them in real time, check in youtube image recognition in real time. The keyboard mechanics is just a link to the actions, and the mouse mechanics should use another neural network to link the action with proper movements. Of course that for learning and training the main strategic neural network should do it in their special software in which it can play thousands of games by day. But what I am suggesting are just neural networks that it would act as interface. The first old games that deep mind learned to play, they receive each pixel of information and the controls and figure out all from there, which is more complex of what I am suggesting. Check this channel: kzbin.infovideos In fact, the thing it would be a pain of the ass to design and build for them, is the mouse external mechanism,

[@juno1483]

They're going to be everywhere!

[@4c00h]

AlphaZero beat Stockfish 8.. could you guys post some games against version 11 running on a proper machine with >300 threads to make it fair? I think A0 has fallen behind since Lc0 took over last year

[@xabashdog]

I love Reinforcement Learning and its variants, hopefully would be working with you guys or somewhere else if not, creating agents that surpass humans by far! 😊

[@Kvk_007]

Wait did he say that the Alphastar can see the whole map? so there is no fog of war for the AI?

[@typhoon320i]

I can't wait for it to be tasked with material science innovation. i.e. "We need a material that has 1000x the tensile strength of steel, but flexible like fabric...."

[@kirbs0001]

"Alphastar can't react quicker than a human" -displays millisecond-perfect timing and performs up to 1,500 APM

[@SPYTHandle]

I believe they did cap the APM to a realistic for human amount and also the amount of screens per second it could look at was also limited.

[@jonnyhirse5949]

@@SPYTHandle The average APM is on a realistic amount, but this version could use too much actions during a fight and use a low amount of actions before to compensate.

[@nkdms.2031]

Are you planning to do the same for Warcraft 3? There is defo a need for fresh strategies in the game so it might be a no brainer :) Great job!

[@penguinista]

It will be interesting to impose human-like limitations on an AI, like clicks per second limitation, innate imprecision, etc. We could find out if the AI 'brain' was beating the human brain or if it was winning based on speed, precision, multi tasking, etc. But we won't impose those restrictions on AIs we employ to address real world problems, so I am also interested in how unrestricted AI does against the best humans.

[@PurpleBraveGiraffe]

You guys have the best job ever...

[@TahaNasser]

A remarkable feat no doubt, but it would have been great to see AlphaGo face the top 5 in the world including Serral with their own mice and keyboards etc. It makes a big difference.

[@frederikspitra7832]

Well, you forgot that Mana won against the Agent who was not allowed to be everywhere on the map at the same time, although more to it glitching out than that. Still, good work. Lets see where this will go.

[@infinitejest9978]

Nice to see humans that outsmarted other spicies with their tools are now been outsmarted by that tools.

[@ronaldbaginski7889]

Anyone know how to contact AlphaGO ?

[@achyuthvishwamithra]

David Silver is an inspiration to many 🙌

[@ronaldronald8819]

This is Amazing! Whats next?

[@alexanderyell4693]

The only issue is yes the actions per minute are human like but it if go to effective actions per min during engagements the alphastar reached numbers of 600-800 per minute while mana was at about 400. Effective actions per minute ate the ones that matter.

[@HY-vz3ks]

When will AlphaMale be developed?

[@quentinlauterbach]

The Pros looked somewhat misstreated, since they kind of know that right now they got outplayed more by mechanical precicion and efficiency rather than straight up strategy. So if they can balance for example the APM in engagements to somewhat match the physical capabilities of a human being it would be more comparable. Nevertheless its needlessly to say that the AI already is very impressive and I´m looking forward on keeping an eye on Deeplearning for RTS and other types of games!!

[@pororo5249]

DeepMind had to consider Korean players with Starcraft 1 (remastered).

[@kwanarchive]

Korea has suffered enough at the hands of DeepMind already :)

[@97lovechina]

what's their chair?

[@bigdaddyrazor]

This Documentary is gonna be BIG! (Am I the only one hyped?)

[@cybervice1125]

Guys I watched the live stream and on the oversaturation on probes, maybe the tought process of the agent was that the enemy will attack me with a few units> I loose X numbers of probes> hence I make more probes. We saw that the agent was at 24 probes when TLO attacked with 2 adepts and after the attack the agent had 17 probes in main and another 3-4 going for the expansion...

[@danielbigham]

I wondered the same thing. Good thought. Another: Once you do expand, you can immediately jump to having 7 probes at it.

[@elmohead]

haha you can see MaNa oversaturating his minerals from game 2 onwards as well. Seems like oversaturation will be a thing in the future. It prevents your income from dipping when getting harassed.

[@LyubomirIko]

And it brings slightly more minerals too!

[@pliskenmovie]

Artosis did a full stream analyzing the games, and really zeroed in on the oversaturation of probes. It's about a 10% boost in income, and makes mineral line harassment less effective. It does however limit a few early options. Also, MaNa did a stream analyzing his gameplay against Alphastar, and was a little annoyed at the insane precision displayed by the AI. In his last game that he actually did win (when the agent wasn't allowed to see the entire map at once), him and TLO strategized that they could never match Alpha in micro, so they figured an army of brute strength would win. Not only did strength carry the game, but MaNa was definitely able to confuse Alphastar and exploit its lack of understanding.

[@elmohead]

@@pliskenmovie I watched MaNa's commentary, haven't watched Artosis'. But what you mentioned makes sense about the extra workers. It's good if you wanna play macro and have set builds (AlphaStar).

[@vadimbakaev2099]

When we can see Zerg?

[@Frumpbeard]

I like how even the people that developed the AI had no idea what the hell it was doing, and expected it to lose.

[@bighands69]

It will actually take machine learning to discover how the machines are coming to their conclusions. In other words it takes a car to catch a car.