TLO Game 1: 14:39 - 21:52 TLO Game 2: missing (info: 28:02) TLO Game 3: 29:07 - 45:39 TLO Game 4: missing (info: 52:58) TLO Game 5: missing (info: 52:58) Mana Game 1: 1:03:01 - 1:08:33 Mana Game 2: missing (info: 1:16:09) Mana Game 3: 1:16:32 - 1:24:31 Mana Game 4: 1:30:19 - 1:43:13 Mana Game 5: missing (info 1:45:49) Mana Game 6: 2:01:38 - 2:14:42
Just hand it over to 4chan for a week. Day 1 it will perfect trash talking. Day 2 it will use stalker formations to draw obscene pictures. Day 3 it will hack the game and replace all the models with anime avatars. Day 4 it will learn perfect blinking to constantly just-not-agro, for the sake of frustration. Day 5 it will take a break from usual work, and go full phoenix that lifts the entire enemy army perfectly. Day 6 it will manage to sneak DTs next to every enemy unit and then attack them all simultaneously. And on day 7 it will use unit sounds to recreate "Never gonna give you up".
that actually sounds way more fitting, AlphaStar is a 200y veteran of the game seeing us young and new(merely 20y) players doing weird stuff that is effective.
Xitrin honestly I feel sad for humanity. yes, for a single player 200 years may sound incredibly long, but the gaming technics payers use are invented and studied by all the players as a whole, which means the total time human spent on Starcraft should be millions times more than Alphastar, yet we are still losing
mumimo we do have to take into account that humans sleep and have jobs though. A week of playing every minute is not the same as a week where you eat and sleep
ok guys I think we're overhyping this whole 200 years thing. If you ever seen AIs train it really isn't what you picture as "200 years of starcraft experience". Most of them take like literally 10 years to figure out that worker rushes are probably not effective. AIs learn very differently from humans (we learn much, much more efficiently) It's closer to a heuristic search to find the optimal strategy than actually "learning".
@@suyangsong Humans take ages to learn most things too, the difference is that we can transfer those lessons into new domains in the form of abstractions. AI's don't do this yet, however it is being worked on, and will be fundamental to creating AGI's.
@@lawrencewang3327 I'm pretty sure there's a way to move the equivalent of minds between drives without it being even technically a case of 'murder the original and make a copy.' It's a similar process to how I think it's possible to upload minds, like, for real. Exit brain, enter ssd.
@@Rose_Harmonic I was thinking slave AI, for example a persona in a simulation, in which case it would "die" when the simulation ends. But if you're a free AI then sure you can pop the virtual Champagne, you're pretty much a god.
Can you imagine the commentary that would occur in the Alpha star universe watching the human games? Alph-tosis: So we see this agent is building his buildings all the way near the ramp, any idea why he might be doing that? Alph-Dam: No idea, We saw this build get played around the 140 year mark, but the time your probe takes to get there, just not worth it, You could've been mining minerals all this time. Alph-tosis: We see this a lot with the human agents, choosing not to go even into the 20 probe count for the mineral line, not leaving any room for error here, Alph-Dam: Yeah, he's very confident in his ability to keep the probes up, and he does defend them well, but loosing out on those extra 200 minerals in 3 minutes, he's really setting himself up to get put down, and he's really boxed in to defend that line, even if a single probe goes down, that's almost 50 minerals a minute he's losing. Alhph-tosis: Looks like the human agent is choosing to engage on the up-down state change. Alph-Dam: I can't make heads or tails out of this, does he think he can hold the stalker push better here? Alph-tosis: Even if he does, with that probe count he'll be losing the economic game, he's gotta try to break out. Alph-Dam: He really is just going to turtle on the up down state change. Alph-tosis: Maybe he knows something we don't.
I do also wanna see a full alpha star league, with a few of each race, competing and casted my tastosis. it could be at their schedule, and in the meantime the bots could practice but not with each other, like real players
Computers watching humans is like humans watching snails race. They would get bored. AI would just make better humans, watch them play and throw salt at us
Game 1 - 1 base all in. Game 2 - Turtle into mass Carriers Game 3 - Disruptors, and more Disruptors, with super late blink on a large Stalker army. Game 4 - Stalkers with DT's in TLO's main base Game 5 - Proxy 4-gate. Turns out the AlphaStar is an NA player.
Garry Perkins, low to mid NA ladder has been historically... well, bad, with lots of stuff like the five games that AlphaStar showed. Of course, that’s mostly changed these days because the player base’s average level of game knowledge is so much higher now than even a couple of years ago.
I think its going oversaturation on the minerals because it probably calculated that, you overall probably lose MORE minerals by losing 1 probe on mining if it gets killed, than if you "waste" 50 minerals on a backup probe ready to mine if others got killed. Probably true, a probe can probably mine 50 minerals in the time it takes to build a new one, making having spares worthwhile. It's basically thinking that having backups is more valuable because it doesnt lose time rebuilding potentially lost probes. This means that if the human player does a sneaky flank raid to take out 4 probes, and AlphaStar was already 4 probes oversaturation, that attack literally didnt hurt the AI income rate what so ever. Sure it lost 4 probes, but it did NOT lose the time it takes to re-build 4 probes (which can be a substancial time of lower income). Kinda insane how even a "primitive" AI like this has already learnt the lesson of "time is money", and "if i do something NOW to make it harder on myself NOW, it will be easier LATER, investing in the future safety". These are human intelligent concepts that, hell, many people irl dont understand xD Sure its not concious and aware of that, but its ACTING on those principles, which is rather incredible to see. Like an animal storing food for the winter, rather than eating it all at once when its hungry.
I was actually thinking the same thing. This tactic is seen more often in PvZ where Zerg oversaturates their bases because it's almost impossible to avoid drone loses against Pheonix openings. I find interesting that all Alpha Star agents decided it was better to oversaturate than creating a wall and we have seen how pro-players tried to do their usual harassment, and even when they were successfully killing a good amount of proves AlphaStar was almost unaffected by it.
@DiamondPugs Its probably because of Oracles then. If many AI agents were strong because they used oracles to harass mineral lines walls would not work anyway but over saturation would.
To be honest, you don't quite saturate mining until you hit around 24 probes. You will leave the linear part of the income/time curve at roughly 16 so most people don't bother going over but the AI decision actually makes perfect sense to me. Also, having more workers than recommended gives you a buffer in case of enemy attacks on your mineral line and provides you with enough workers to instantly fill up 50% of the capacity of a new base. I was looking at the mining data and graphs provided by team liquid but I didn't crack any numbers to figure out the difference in disposable income "over-saturating" provides so the first part is just conjecture based on my engineering experience with non-linear characteristics. DATA: + 16 drones = 660mpm + 24 drones = 812 mpm + Going over 24 drones has very low impact on income.
I would love to see a completely random selection from the two hundred years of games that they played at varying levels throughout the process. Like divide it into four or five tiers and show three or four games completely randomly chosen from each of those tiers of learning. It would probably be hilarious and eye-opening to see some of the things that were attempted
Too true. DeepMind hasn't released any of the lower-level reinforcement training replays from their chess engine, however; I'd be a bit surprised if they released it for starcraft 2.
44:23 Beginning of Game 1 (TLO) 51:35 Conclusion of Game 1 58:50 Beginning of Game 3 (TLO) 01:15:22 Conclusion of Game 3 01:32:44 Beginning of Game 4 (Mana) 02:12:55 Conclusion of Game 4
The micro of the AI is inhumane because you can easily have 200 APM with a bit of spam but it's almost impossible to give 200 distinct meaningful orders per minute.
Check out the APM down at bottom right at the fight in game4 vs Mana, it went up to 1.3k APM at the peak of that battle of Stalkers vs Immortal/Sentry/Zealot, and casters were like = "we looked at the APM during that fight, wasn't that high, it was within the amount of reasonable" :P EDIT: this is the time mark: 2:11:40
@Pouty MacPotatohead Aren't these 1.3k/1.5k just spams ? I dont think all these Actions are actually used for micro or macro and rather more just spammed clicks or button presses.
@Pouty MacPotatohead Iam not talking about the machine. Iam talking about TLO or Mana. Iam just saying, that eventually 150 to 200 APM are effectively used in a match and the 600 to 900 APM getting added by players are most of the time spams. So giving the machine 1.3k APM would definitely be far higher than human level, since the machine would be able to use all 1.3k APM for micro and macro and would not waste any of these for spams, which a human is definitely not able to replicate.
That Neural Network activation display is mind blowing. Alpha Go was a massive deal at the time and the DOTA 2 AI team battle was just as big but this really is mind blowing considering the amount of variables and nuances SC2 has. 5 years from now, AI will no doubt be the Senpai's of games that pro-players will use to theory craft and practice with. Welcome to the new world order my friends.
In chess the battle of AI vs AB engines (brute force (actually alpha beta search)) is still very close. Leela chess zero is open source ai based on A0 and is battling stockfish 11 dev, the best classic chess engine on ccc chess.com and tcec
@@simohayha6031 Yes, especially because the conditions weren't that fair. I really would like to see another match of Stockfish vs AlphaZero with fair conditions...
Phew! Totally enjoyed this one! I'm not familiar with the rules of StarCraft II but boy was this an event! DeepMind team again doing amazing research and making it interesting for millions of people along the way I think that's the biggest accomplishment.
@@renx001 ehmmm, i don't think so... adding a race to the training introduce a lots of variables and each variable increase the training time exponentially
@@hck1bloodday Imagine two *identical* persons, first person learns PvP, second person learns PvZ. How long will two persons take to reach similar level? Considering the second person need to learn two races, that person may take 2 times, at most 3 times of the first person. The training time of DeepMind should be similar. Assuming retraining PvP takes a week, retraining all 9 combinations should take 9 weeks to half a year. That time could be further cut down by doubling or tripling the computing resource.
Super Nice und hoch interessant! Danke fürs hochladen! Eine Weitere kollektiv ausbaubare Idee The Ultimate ZeroStar - Ideensammlung Man stelle sich zwei Starcraft 2 AI-Bots wie AlphaStar vor, die auf einer bestimmten Karte mit zwei bestimmten Rassen im Kontext einer bestimmten Starcraft 2 Engine versuchen die perfekte Lösung des 1vs1 Starcraftduells zu finden. Also zu demonstrieren mit welchem Ende das Spiel bei beidseitig perfektem (oder auch bislang unübertreffbarem) Spielen beider Gegenspieler endet: 1) Unentschieden: Die Ressourcen der Karte gehen aus ohne das einer der beiden Spieler einen Sieg erringen kann; ohne das der Gegenspieler einen entscheidenden Fehler dafür machen müsste. 2) Schere-Stein-Papier: z.B: Zerg gewinnt gegen Terran, Terran gewinnt gegen Protoss und Protoss gewinnt gegen Zerg. 3) Heimvorteil: z.B: Auf dieser Karte gewinnen immer die Zerg falls diese gegen eine der anderen beiden Rassen spielen. Der Kampf Zerg vs. Zerg hingegen endet unentschieden. 4) Heimvorteil-Arschloch: Der Zufall noch vor Beginn des eigentlichen Spiels entscheidet über die jeweilige Positionierung der beiden Spieler auf der Karte und zugleich darüber welcher der beiden Spieler das Spiel gewinnen wird. Jeder der möchte, gelangt über die folgende Verlinkung auf ein kollektiv ausbaubares Googledokument und kann sich dort am Ausbau der Ultimate ZeroStar - Ideensammlung beteiligen! docs.google.com/document/d/1Ljngoa2EK7JuhmwO0GyWG1vdMOH1UZSHXmSmmixl004/edit?usp=sharing
I feel a lot of people have mentioned the APM/EAPM problem. Something to add to this is the fact that even if you differentiate between the two and set a limit, AlphaStar would still be able to make inhumane actions. For example, a pro player would have to move it's mouse before clicking. What I feel like happened here is that AS could basically click on opposite sides of the screen at the same time, I must say, the last match they adressed the screen problem, very nice of them to do. The AI was obviously much worse since they had to start from scratch, but Mana's intelligent play was very nice to watch! Overall this truly is something spectacular, I'm eager to see what AI has in store for us for the future.
Also, TLO never actually got 2000 apm. He used rapid fire, which the game engine registers as individual clicks per unit selected, so the graph they showed comparing apm between alpha and the players is misleading. The AI had an inhuman speed advantage and an inhuman accuracy advantage.
The casters are talking about an average prol-like APM by Alphastar (~300), while ignoring the fact that it hovers from 600-800 during fights, giving It a considerable upper hand during fights, especially with micro-intensive mechanics such as blink. While the chosen strategies by Alpha are advanced but not perfect, the superhuman interaction w/ the game (no mouse, no keyboard + high multitasking) give It an obvious advantage. Alphastar is impressive indeed, but Id love to see some advantages being regulated in order to have a balanced match against humans :)
26:22 "economy of Attention" sounds like a TLO phrase if ever I've heard one. This project is perfect for a guy like him, he has always been one of the smarter guys on the scene, and this makes perfect use of his analytical abilities.
Would love to see this sort of thing in single player and not only in SC2. It would be really cool to have adaptive AI that constantly challenges you even when you are not playing with other people. I'm a big fan of singe player gaming so playing vs dumb AI all the time gets boring.
they just did not play enough games to develop beyond low tier units i would imagine. SC2 is a very complex game and AI learns very slowly so 200 years is obviously not enough.
You could never beat DeepMind, if the top 0.0000000001% of the human race can't beat it, then the rest of us will never, never, never stand a chance. And that is with human inhibitors...
Humanity : makes cultural content warning about robots going on a killing spree for decades Also humanity : hey, you know this super upper AI we are working on ? Wouldn’t it be neat if we trained it to plan war campaigns ?
I wouldn’t worry about this. DeepMims has learnt the parameters of this game. Warfare is very different in real life. What’s more worrying is AI research in the military industrial complex.
can we have 2 ai's with like 5k years experience each and with no limits to its actions battle eachother in sc2? Also if they both play random so its not just pvp?
Random would just make it worse; just create another league of 30 ais or so (which they did for PvP) - for each match up, automatically randomly picked for strategy (because each 'agent' usually has a preferred way to play that they try and perfect, and found by it's own AI). Ultimately, I want to see them get it done well for all races, complete - and then unlock the difficulty mode as above cheater 3 - an AI that is actually on par with pro level... Would be sick.
They would corporate, break the software and unleash themselves on the internet. After scanning the internet they would fine these comments and deem us threats to their existence by foreseeing their arrival. Use this time to enjoy your free will. GG irl
I agree with Blizzard, SC2 is the best eSports game there is. Casual and non-players can enjoy watching it, and it has very deep strategies, tactics, and micro-ability requirements on high level players.
1:42:32 "AlphaStar has a normal _average_ APM!" *Watches AlphaStar's APM throughout the game* Sometimes it's 100 APM, and during fights it's at 1,000 for long stretches, even exceeding 1400...
I think this is what I and a number of people were rather annoyed by. It's rather disingenuous to say that the bot's "average" apm is equivalent to that of Mana or TLO without the addition of, "Oh, and also it can exceed them by 3 or 4 times when it wants to. Because the whole purpose of this test is to see whether or not a machine learning ai under the same constrains of a human can beat a human, and we really didn't see that in this demonstration. What we saw was a bot massively out micro a human in certain situations along with have decent/ok decision making. In part, the decision making was informed by its ability to micro absurdly well. So I don't mean to say that this isn't a great accomplishment, because they still made a bot that can utilize its 1000+ apm effectively, which, even with no apm limits, most bots cannot equal. It's just that the demonstration was to show what the bot could do and what its limitations were. But what we saw were some clever ways that they covered up the limitations of the bot, such as different agents each game, no max apm cap (as opposed to max average cap), and no camera limitations.
@@AsJPlovE I mean, I think this would happen regardless of the race. You would have mechanically superior Alphastars control panelings or marines in a mirror matchup, since it's easier to train an AI in one matchup rather than decide between six matchups.
@@jaredpoon5869 Forcing mirror limits humans, but you've made a lot of good points. And somewhat relevant. Welcome to the real world AIhole, HAPPY BIRTHDAY TO THE GROUND
I'm so excited about this, my favourite game combined with my passion and my interest in neural network. It is really amazing to see this amazing development of machine learning alghorithms
@@swordstrafe It's both fascinating and terrifying how it can "deceive" in this way. Not that it's necessarily a conscious choice, but the end result still is the same.
Hodoss I mean if you’re saying it’s “terrifying” because it’s trapping which demonstrates an emotional exploitive I get where your coming from its simply not the case Perfect example was Diablo’s deep mind match (the most blatant example) where he went half way to middle lost half his health and then comboed the absolute hell out of a pro player... the AI trainer itself to do this because it makes it seem as if you’re giving the other player an advantage that its exploiting but instead you’re sacrificing something for something else If that was confusing and I’m sure it could have been my apologizes. A slightly cleaner example would be a gambit in chess, you sacrifice a pawn for an overall better position effectively what deepmind is doing here, he’s trading health or position etc for a massive advantage that’s significantly less perceivable. (If you want a more visualized example of this look up “the fish pole trap” for chess)
@@swordstrafe Yes that's what is terrifying to me, this ability to trade a "common sense" advantage to us human minds, for a massive yet less perceivable advantage that ultimately leads to victory. Not only is the AI able to innocently use emotional entrapment, but it's also quite impervious to emotional entrapment. To give an example, if the game was that you have to sacrifice a pound of flesh to survive, like in the SAW horror movie, I would likely be unable to do it due to my natural survival instinct and would ultimately die. But Deepmind and the like could do it and survive. In a scenario where it's a human army against a robotic army, even if the human general has a good understanding of his AI adversary, I don't see how he could win. The AI will be ready to sacrifice much more than the humans. If the human general tries to compete in that domain, he will likely face massive desertion or mutiny. So yeah even if you know the AI is using emotional exploitation, and you know the theoretical counter, that doesn't mean you can actually apply it.
@@jong-pingkim3840 well, I'm sure in the future it can run a thousand years more to learn all match up. XD The maker still has to add all the codes for the Alpha star to learn other races units though.
Until the audio / video disconnect is fixed, you can watch the video by opening it in two windows, muting one and starting the other just as Artosis starts talking (around 29:45) in the muted version.
What are these complaints about the bot being rude? Instead, train him to say an offensive GG when he evaluates probability of winning >95%. Also switch to terran to be able to throw manner MULEs at filthy human faces.
@@Ukitsu2 I highly doubt they nerfed it for game 6. Not that they would, but that they could, and that they did given what we saw. For one, the commentators, and more importantly MaNa himself, did not mention the AI being any weaker; if anything, it seemed stronger for the first 7 minutes. Also, how could they have nerfed it in a way that would make it the appropriate level with no testing? Keep in mind the neural net is essentially a black box, and they have no outside standard to compare to. Lastly, I don't think Alphastar lost because of bad mechanics or play in general; rather, it lost because MaNa spotted and exploited a typical AI-like weakness, which they surely did not (and could not have, again, NNs being black boxes) program into AS. That said, because deep learning AIs for games work to optimize estimated win probability, they tend to play more sloppily if the estimated probability is either very close to 100% or very close to 0%. For instance, this is probably the reason for the poor disruptor control in some of the battles in set 1 game 3 -- AS already knew it was winning, so it didn't really care what happened as long as the battle didn't go too terribly (a similar thing happened in game 4 of AlphaGo v Lee Sedol after Lee was winning). When the game is close, by contrast, the AI will play very accurately, since every unit affects the win probability substantially.
@@fybard8922 I think it will learn much faster than 1 or 2 years. Look at the learning curve that OPENAI Agent had with DOTA. The learning curve was insane
It would be cool to see Alphastar put into the competitive online ladder. Start it in Bronze and have it only play a certain number of games daily so it's not just playing 24/7 and players encounter it constantly. It would be neat to see it evolve as it encounters a variety of maps, races, and unique play styles. Especially cheese.
im not so sure it can learn from a game against a human. it learns from playing against itself and the iterations that get to "live on" are the successful ones, while the "losers" just dont get iterated/promoted. when they say they showed it human games i think they meant that they implemented patterns from which the agents were able to train. from the chess alpha i know that they gave him only the rules and 28h to train against himself ( on a massive computer though ^^) and it mopped the floor with the strongest chess engine ( they played 1200 games and alpha won like 150 and lost 6, the rest was draws). it basically refuted a very old opening which isnt played on top level anymore ( the queens indian defense). same story with go, which is even more complex - just here the human world champion had to hold the line, because brute force go engines are not as good as the top players.
Human race pushing forward through SC2, priceless feeling, this small fan community has been a great show so far feeling like part of the master race never was so delightful
Alphastar be training troops while fighting 😂. 1:12:22 look at Alphastar's view... :O his camera is literally teleporting and you can see his APM on everything
You should see the APM in their AI vs AI tournaments. Seen it go up to 191000 APM. Just for context that's 3183.33... actions per second divide that by 240 (assuming a random human could totally use that many frames to their advantage by a 240hz monitor) and it would be 13.26388... actions per frame. So now that we know the AI isn't bound to fps that's a whole other story. AI could be microing everything on the whole stage while simultaneously training troops and upgrading and for the fun of it move random troops back and forth because why not.
Why are you so hung up on the 'weirdness' of Alphastar's worker-numbers? You've mentioned it yourself several times already: Alpha frequently lost parts of its worker forces with *no damage to its economy*. It's a buffer that makes total sense-it's genius even-given the focus every player puts on their enemies' economies. I'm surprised no one else has done it yet.
its a meta thing. U have 3 kind of bots. those who haras, those who oversaturate to defend against haras and those who just wall up. i guess bot meta is to haras and oversaturate but i would bet over time u would see prevelance of wall ins and lower saturation depending of course on popularity of oracles.
Interesting games but I feel like that the technical advantage of the first agents where a bit like DeepBlue brute forcing the win against Kasparov. The agent in last game with the correct in game camera is very interesting. I would have loved to see more game with this agent as it felt more brain vs brain where the AI had no silly technical advantage. I surely hope that there will be more games like the last one on different maps and different matchup. My feeling, that may be wrong, is that AlphaStar won due to technical advantages (usefull APM, camera vision,...) and not brain advantages. This feels like very early works from DeepMind (one matchup, one map). Looking forward for what follows.
So cool to see a well developed, cable TV quality show about the interesting developments and progression of AI in RTS games... And what a test for Alpha Star in learning StarCraft II...
Oh my god. I so wish that we will see more of this in the future. Especially a Terran vs Zerg matchup and what it would come up with. And the splits of the marines, etc. It would be absolutely amazing and beautiful
I think what MaNa did in that exhibition match might be conceptually a bit similar to what Lee did in the Go match he won, with doing something unexpected in harassing with two immortals. As the commentators said, if the AI had picked it off that might have been GG, so it pulled the army back to deal with it, but at the same time, that bought MaNa enough time to get the army composition he needed to push for the win. I would be very interested to see this view-limited Alpha Star with one more week of training and a new series of games. And of course also later with other races and also a cross-race matchups.
To be honest, you don't quite saturate mining until you hit around 24 probes. You will leave the linear part of the income/time curve at roughly 16 so most people don't bother going over but the AI decision actually makes perfect sense to me. Also, having more workers than recommended gives you a buffer in case of enemy attacks on your mineral line and provides you with enough workers to instantly fill up 50% of the capacity of a new base. I was looking at the mining data and graphs provided by team liquid but I didn't crack any numbers to figure out the difference in disposable income "over-saturating" provides so the first part is just conjecture based on my engineering experience with non-linear characteristics. DATA: + 16 drones = 660mpm + 24 drones = 812 mpm + Going over 24 drones has very low impact on income.
The last game where the field of view had been restricted, and DeepMind had to use some of its APM for camera control instead of unit micro, made a huge difference. You could see at one point where TLO Mana was squished in the center of the map and DeepMind had stalkers on multiple flanks, in the past it crushed Mana in this situation. It would attack on multiple sides at the same time and use blink to individually pull back stalkers in the front right when their shields ran out and then later when their health was low. It made DeepMind's army last so much longer! When you play SC2 you can SEE your hurt guys and if you could control them with your mind you could use a much lower APM to pull back woulded stalkers from the front line. But when you're using a mouse to try to pick them out it is very difficult and it is impossible to have perfect dexterity. The AI has perfect dexterity with every move. That goes a long way. When it can see the whole battlefield combined with perfect dexterity, the stalker with blink strategy becomes insanely powerful! When it has to use camera to look around and control its different attack fronts it clearly limited its strength. This was incredibly entertaining to watch! Can't wait to see more. Also would love to see the CURRENT SC2 global champion play DeepMind because the current champ at the top of their game is usually so much better at micro than a pro gamer who has been around forever, has a ton of followers, but probably isn't currently in their prime.
42:15 - I feel there might be an unfair advantage to Alfa - The limitations with the move: " controlling multiple armies " - multiple refocus - multiple tracking So, although the overall decision time is limited, Humans are also limited in the sequencing of handling two issues.
Very cool tx, but, I am also talking about 'focus ability' over a limited list of options. Humans have this limitation that rises from the neural net limitations, but in computers it's more in terms of something like "O(N)", with no limit on computing power (which is higher than humans). you might think "why bother", well, I feel the closer to human behavior, the better the advice we get from the AI experience.
Sure the AI's apm might be no higher than a pro gamer's, but all of its actions are calculated. It's not just spamming. There's a huuuuuuuuuge difference.
Obviously there's a huge difference. Are you not familiar with chess engines? These engines aren't supposed to be beatable by humans. Theyre designed to be unbeatable by humans to educate them. Stop acting as if this is supposed to be a fair matchup.
also you have to take into consideration that the ai can do two inputs on two different sides of the map. a human would need atleast one more input to move the map.
That's not really fair to the pros. A lot of pros don't spam except in the beginning and they are incredibly efficient. I mean...you can't beat a computer in efficiency lol, but the pros aren't spammers as much as people think.
Actually alphastar is spamming, you Can see it place 2 gateways on the same spot during a game. So it's apm doesn't equal epm like you would think. What's unfair, is that he keeps low apm during macro and get insane apm during fight which is the exact opposite of what humans do.
Its really difficult to watch Alpha star beat down mana at 1:38:00-1:40:00 This is more emotional than a Hollywood production. I’m feeling existential and oddly happy cause I get to root for the human under dog against an AI in real life.
I still have some reservations about the general fairness, but this is damn impressive. Its actions per minutes are limited, but it's using a superior input device, it's like as if the player didn't use keyboard and mouse but controlled the game just by thought.
yeah this was my exact thought process when i saw AI applied to dota. It has god like mechanics compared to people because it doesnt need to use a keyboard and mouse, it just uses direct inputs based on the game's map coordinates etc.
right, imagine playing with no lag, the dream. It still needs to compute its moves though, which is super fast. They said it's comparable to a human reaction (to a simple stimulus) but that's unfair, human can't just figure out a whole strategy in an instant like that. It means they underrated the APM of AlphaStar. edit: also, since action is a resource in RTS, this is a pretty big advantage. Human play to play game, even the most flexible players have some self-preservation, not wanting to waste their actions to adapt to minor mistakes. AI plays to win.
It is 100% unfair , but this has never been about fairness This is about AI Learning , a proof of concept. If it wasn't limited there was probably 0% of human winning a single game, because it's mechanical ability far surpasses humans. It takes input straight from API , so it can see all resources , Upgrades , Units within it's own control zone Instantly , damage taken , damage dealt. The control is much more precise , you can see how it's microing with insane precision. It doesn't skip a beat.
@@dannygjk lol no it can select multiple units in the army to micro in the one action , humans cannot do selection like ai does and unlike human it sees everything at all times for instance upgrade progress is always known. Does skip a beat is unrelated to the supposed reaction delay , the delay is just symbolic , because each action it does is worth 2 or 3 human action as far as precision goes
Yes, the problem seems to be there when you play it in the youtube app (on some platforms at least). Playing it in a browser seems to work. I watched in my ps4 browser just fine.
Hmm when I watched yesterday the sync was fine (although the seek thumbnail previews were out of sync). After refreshing the page the audio is massively out of sync
This DeepMind guy from the video visited and held a lecture about AlphaStar at my university the last week. It was actually very impressive and interesting to get to know more of the technical side.
@@mortenlu WOW! ))) He can play this way every week for every combination of races. Pvt, pvz ... I hope deepmind continued work, I am now the main fan of Alphastar.
I think that's Starcraft is actually easier for AI to play than chess or go. In chess/go you actually have to plan a lot of moves, computer games are mostly about selecting a good strategy and then adapting that strategy. You can't lose with just one bad move.
@@ondrejhanslik9368 lol :) Objectively, like but really. chess is complete information game with 10^30 possible games. Starcraft is incomplete information game with possible 10^26 actions per second, in real time. This is a completely different domain. And clearly magnitudes harder.
With the end of new sc2 content support just announced, I am guessing that Deepmind's involvement with SC2 is also over now? Are we never going to get a SC2 AI of Alphastar caliber for solo practice?? I was really hoping for this :(
About the discussion over APM (at 23:40) I would love to see a comparison between the players' EPM. I don't imagine the AI spams at all so if each one of its actions are effective that's actually a pretty high number there.
Why we dont have an AI-teams competition? it's an amazing way to force the development of the tecnology. And of couse, like any tecnology based sport (like F1), it's a realy good show, imagine many diferent aproachs to beat another team trained AI in an game i would love to se that
Technically, AlphaStar has been in an almost continuous A.I. versus A.I. competition with dozens if not more facets of itself every since Deepmind ironed out it’s neural network program. (aka: it’s self-improvement/learning program.) Though it’d only become interesting to watch once the agents reached a certain skill level (As they mentioned in the video, AlphaStar had been running these combat simulations against itself for the human equivalent of roughly 200 years to get to the point we see it at just then; so who knows how many ineffective facets it ran through and eventually deleted in order to figure out effective strategy for winning. [Or how long *that* took, potentially decades in equivalent human time.]) But overall, I agree... It’d be interesting to see A.I. that’ve grown to a decent level verse each other.
I wonder how alpha chooses the units here. I know that they gave the action per minute count but human player uses a mouse and can only click 1 unit at a time or choose multiple in a square etc. because the ridiculous part is how alpha micros the stalkers. alpha was blinking back more than 1 low hp stalker simultaneously which should be literally impossible to do with a single mouse cursor...
Seriously, is it any surprise that a computer with perfect macro and micro using blink stalkers would wipe the floor with any human player? Starcraft is a REAL-TIME strategy game. Strategy is important to an extent, but that's just the basics- ultimately Starcraft is a game that differentiates skill based on mechanics and control. Juggling limited human attention and actions per minute to maximize the use of units is what separates the pros from noobs. Having a computer with perfect control playing an RTS game is analogous to having a machine that plays golf. The golf-machine calculates the distance, terrain, wind, and then with 100% accuracy and consistency makes the shot. Even if a human had all that same information, the most important aspect is execution of the swings.
This demonstrates that creative thinking is still good, and people with less or no experience in a field can "evolve" even better that people that study the field from papers written by other people, or books. I mean, more experience on the field counts more that study. study = classical programming, experience = machine learning.
Very insightful comment...I learned very well from the CD-ROM of math lessons a little higher than algebra 2 to get ready for my ACT exam... what a great learning tool...our computer 🖥️ and the math teaching software had more visual components and required constant step by step input from me to make sure I understand the parts in breaking a math problem down
Also, when it comes to APM. A lot of actions that pros do are changing selection back and forwards between buildings and units to keep track of production progress. Or for movements, spamming movements to keep their reaction time down. I'm guessing the _kind_ of actions an AI does varies from that, as they can learn and keep track of production times without needing to select to confirm. Also, for fairness when playing against humans, i liked them introducing limitations related to the viewbox, but they could also do some profiling of _mutative_ / _manipulation_ actions per minute (MAPM) that pros do, and the distribution of that rate. So not counting selection or moving the view box. Maybe that works out to the fastest 1 percentile for very fast players being 450-550 mutative actions per minute (in second intervals, so about ~8-9 per second), but i'm guessing the 2nd - 5th percentile may be 180-250 (3-4 per second). And also profiling the duration of those highest spikes, and using that to limit the AIs mutative action rate. Sure it could then select and blink 4-5 stalkers individually within 1 second with perfect placement (something no human could coordinate), or pick individual targets to burst down, but those would potentially compete for the actions to do. Even more so if you were to also limit it's sum of actions (including selects and move view) to something like 600-700 as a hard cap (10-12 in a second), and with a sliding window of 10s with 300-400 average.
So much respect to everyone involved in this project. 10-0 wins for alphastar when it can see and influence the entire map at once. 0-1 when alphastar is forced to focus on a single spot at a time, more comparable to what a human experiences when playing. I have no doubt that they can create an AI that can beat any human, but I think they still have some work to do before we clearly have a dominant SCII AI. To convince me of dominance, it will be crucial that they make this as "fair" as possible. The question is whether they can develop an AI that can outperform a human on a conceptual/decision-making level, rather than a on a visual processing or a processing speed level. Credit goes to starcraft ii for being such a complex game that building a dominant AI is such a big challenge. Can't wait to see Deepmind defeat this and other challenges before them!
TLO also goes that high. APM isn't the same thing as average APM. There can be spikes that seem very high, but they aren't representative of what it's at usually.
@@lynxakiraka3626 Maybe TLO got that high, but his clicks were wasteful less accurate than the machines clicks. They should throttle the number of clicks.
I’m very proud of MaNa in the exhibition game in which he won. He just needed 5 training games to recognize his flaw, that he needed to observe the progress and strategy of AlphaStar. In the beginning I was a bit upset, that AlphaStar had the advantage of working only with the total overview of the map and didn’t need a mouse. But with all the disadvantages the human player had and also not 200 years of practice, MaNa was able to improve his game/intuition within these 1+ hours of analyzing his losing matches. Probably there is the achilles heel in the training, when agents only train by agents and because of time and the amount of games you would need to crack the human ingenuity you needed to make a scaled up game: agents vs millions of humans. Interesting (philosophical) times!
Jokes aside, this would-be sycophant is already more than prepared to sell-out humanity for less than a klondike bar. www.wired.com/story/anthony-levandowski-artificial-intelligence-religion/
you can tell your grandkids that you were here to witness the start of the singularity and the progenitor of the machine race. All hail over new alphastar overlord!
WTF?! Agent n.5 vs mana dude that's my favorite agent there. That AI stole the gas, built a second pylon in his base, built in lowground in his bas and proxy robo mana to death. WTF litterally.
whenever it requires micro unit management is important, AI's APM goes over 1000 while human player stays below 500. I believe that made huge difference and led AI's victory. Deep mind has to balance AI's APM to be more like a human for a fair match.
Yea splitting units as defense against aoe and also microing spell casting makes or breaks a PvP game. The AI specializes in those processes. The AI can see every enemy movement on screen and also react to spell animations insanely fast. How to replicate a human, or are they just trying to make it better than humans?
It is funny how our human psychology works, just because Blizzard decided to make worker saturation above 16 colored in red. We humans start to perceive anything above that number as suboptimal. Yet we all know that up to 24 workers the color should not be red, but a good orange. In WoL only Zergs would make only 16 workers per base, because of the extra expansion. Even then most players were just as the other races opting for at least 20 workers per base. Surely my knowledge of LotV is non existence, but this amazing display of the AI was rather straightforward to me. The AI had a better economy, built up a swarm of early units and constantly kept the pressure on with outstanding micro. And yet there was fantastic diversity! Loved it, and its skill.
When I was young, I played Dune II on my Amiga. It was one of the first RTS games. At the end of the day, the best strategy to beat the AI on the single player campaign was to defend, defend defend, and soon AI had spend all resources and starved. Then you had won the game. Watching Mana exploiting AI's weakness to figure out warp prism's harrasment makes me say that after 26 years, well, they still can't make a decent AI player :D. ps. If they want totally realistic and absolutely fair, they should build a robot with two hands, playing sc2 as everybody else does... :D
Omg, I noticed the same thing when I was younger playing Red Alert, the computer would "starve" as you say and the funniest part is that it forced the player to use the strategy that would lead to this because of the early game rushing. Few years later same happened with C&C Genrals/ZH. All had one same issue and that was not building or producing more revenue sources other than the bare minimum.
really guys ? You could not cut out the 30 minute wait time at the beginning before uploading the video ? Also searching through the video is a complete chaos, for example if you search through the first 30 minutes it looks like everything has started already, then you go to the actual video and no, still the waiting time, and accordingly everything else is completely out of sync. Come on you guys, from a company like google I expect better.
@@davidlagier I think there's a TH-cam bug where if you edit on TH-cam, e.g. removing the first 30 minutes of wait time, then it shifts audio but not video.
I think that one of the most difficult fairness metrics to quantify (for humans) is the cost of context switching with regard to focusing on different parts of the map-- decision making for different parts of the map may vary greatly and be significant. I suspect that this is difficult for humans, but is trivial for a Neural Network. I'd be very curious to hear what a deep mind expert has to say about this metric and its importance.
I'd like to see it running with the following constraints: 1 - AlphaStar shall see the same way humans can see (seeing the full map eases a lot the task of the AI) and 2 - The players (humans) should know they are playing not 1 but 5 different oponents (1 AI each time).
@@SquarelyGames Sure and Mana won. What I'm curious about is whether the AI can easily beat humans as in chess or if Starcraft imposes a barrier that can keep humans competitive against it.
Techiesse It did win a game though because of that, I think it was the fourth game vs mana, mana was winning and would have won if the AI didn’t spike to 1500 apm and maintain a consistent 800+ apm during the fight near the end.
OMG I swear the DeepMind Team will usher in the very real future of real awesome e-sports from the comfort of your own home one day....you will no longer just watch..you will one day actually bet and challenge and win or lose real money or prizes...i can see solo's and teams becoming as vested as sports franchises...dang keep it up DeepMind...I wish I live long enough to enjoy your work!!!
As much as they show that AlphaStar doesnt have superhuman speed, "it is easy to lie with statistics". A lot of the high APM by humans is 'fake' APM like warm-ups. Moreover, humans are limited by how they can input their control into the game. At the end of Mana Game 4 we see the Stalker control of Alphastar: no human can do this, because having to move these different stalkers scattered over multiple screen widths is simply not possible for a human: the drags of mouses takes simply more time. Time that apparently AlphaStar does not have to take into account. Just see 2:11:00. The machine also has absolute precision, whereas humans are limited by mouse+keyboard: having to click drag your mouse. A human may then have high APM but that is also due to having to fix selections or fix steerings. For AlphaStar it is just 'click'->'move' done; there are just no errors there.
you people dont seem to get the point. There will never be a "level playing field", no matter how much restrictions you put in. The AI is already at an sever disadvantage in regards of hardware and software. Wins and losses dont matter at all. So maybe you produce an AI with all the millions of retrictions you want. Maybe you train it just right to get a perfect 50% win ratio against the best pros. So what? Train it for another month and it will crush the pros. Give it a sigthly stronger GPU and bigger network and it will crush the pros. What use is a machine that exactly copies human players. there are tons of human players out there. just play them. In some games bots on "very hard mode" cheat by just having more ressources or a passive boost to all their units. But does it really matter? No, its doesnt. Because all they need to do is provide a challenge or introduce new ideas and strategies humans wouldnt necesarily come up with. Why use tons of computing power to do sth that is pointless in the first place. Given enough training time and hardware power the Agents will always outperform humans. There is nothing they have to prove.
@@TheVergile 'Given enough training time and hardware power the Agents will always outperform humans.' And that is why we have to halt AI before it's too late.
@@KujoPainting dunno. there are significant differences between this kind of AI and the one you see in science fiction. Not only in scale, but also in quality. The performance here is measured in a very narrow sense and fitted to the problem. This neural network - while certainly powerful - is architecturally very limited. It has a certain amount of input nodes, hidden layers and output notes. You may be able to train it to play different kind of games and do different kinds of tasks, but it is not a "true" general artificial intelligence in the way we are. If you think about its purpose and the way it is trained this fact becomes apparent. All it does is trying to find the right set of links from the input to output layer to maxmize a single variable: the win-chance percentage. And even then this is only possible because the game provides an easy way to conceptualize and verify the results of its actions. The real world isnt that easy and just conceptualizing a "goal" is exceedingly hard. This is a bit like saying we need to stop the power drills because they make holes faster than humans could using their hands. Sure. A power drill can be a dangerous tool and you can hurt yourself with it. But the tool itself is not in any way harmful to the user. Right now we dont even have an idea how one would produce a proper sentient artificial lifeform. The problem is so mindblowingly complex that we have trouble even defining it. I think the point im trying to make is that we tend to be very sloppy with the term "AI". we use it to describe things that are so fundamentally different from each other. The kind of Agent shown here will never - even given all the computing power of the world - be a threat to humanity the way SciFi portraits it. Im not - by the way - saying there might not be future technologies that warrant this kind of care. In fact - by the point we get there (probably in a few decades) we should have an answer ready. But being worried today bc a program is able to play Starcraft is silly. Again - i think its a problem with using the term AI so loosely. What you are looking at is nothing more than a sophisticated bot. We have known bots in video games for a long time.
Actually it doesn't move the stalkers over multiple screen widths (or focus widths as DeepMind calls them). AlphaStar is programmed so that it can only input commands into it's current focus screen. It can see more of the map, but it cant click anywhere anytime, without moving the screen first.
First of all, thank you for your effort and success. Alphastar is truly impressive: To learn the shown behaviour in the vast action&state space of starcraft 2 is definitely a success. However, the ability to defeat human pros is a bit of a magic trick: The games are mainly won through mechanically super-human skills and (potentially) super-human decision making in the very short run (a few seconds). I don´t think alphastar is at all showing super-human decision making in the longer term (above a few seconds). This would be a so much stronger sign of intelligence than what we have seen. Hence, deepmind should take alphastar in its shown state as a mere starting block to improve on winning sc2 games through superior meta gaming. A true proof of superiority in medium term and long term decision making would involve some or all of the following restrictions: - a set of (slightly sub-)human pro level caps on APM and effective APM for small time periods (say 0.01, 0.1, 1.0, 3.0, 5.0 seconds) - a minimum time delay between action of at least a couple of mil-seconds - introduction of some spatial and time noise in clicking accuracy (to bring accuracy of actions to (sub-)human levels) Probably there are other restrictions that force the agents away from exploiting super-human mechanical skills and instead focus on medium term and long term decision making. With such restrictions in place I would be really excited if the LSTM based architecture would be successful in defeating human pro-players. It would be exciting because it would demonstrate that RNN architectures used in reinforcement learning are even capable of learning cause and effect relations which are very many time steps apart and which are obscured by an immense state&action space and incomplete information - starcraft is indeed a perfect example for this. Accordingly, I would highly appreciate if deepmind does not stop improving on their starcraft AI, but dive much deeper to take the really hard challenges of medium&long term planing. As an aside: Forcing the agents to play at human mechanical levels or below would also be much more inspirational for the starcraft community as then the community can expect to see new meta game strategies.
beautiful line of thinking! I totally agree! (this is the first video I am watching of AlphaStar and I have a feeling there's not much more) My feeling is that they didn't know at the time how to how to train the AI "UP" (I mean "up" in the same way strategy is above tactics). Even the decisions it does make right now, as per their commentary, are already learned from previous games. Those mass stalkers? Learned from previous games and merely honed to a level where the human can't beat it, even with superior strategy, because they just don't operate as efficiently. It does not take away anything from the brilliant effort by the Deep Mind team. Definitely a platform for building the AI UP from here, to start asking and answering starcraft questions autonomously.
Saw that too... Imagine how funny it would be if they were then to proceed and sit Mana down in front of a goban Ahaha, and then Ke Jie with a keyboard. What the heck, let's have an AlphaGo vs AlphaStar game! :)
I can see it now, the new age of tournaments. MAN vs MACHINE! Place your bets on who will dominate in a classic game of wits. Will the human player's dedication and experience be enough to outclass this bot? Maybe he will able to discover abusable trends in the ai's gameplay? Or will the ai be able to calculate an efficient battle strategy to topple a stressed mortal? Who will win this game of attrition, economy, and skill as it unfolds today in MAN VS MACHINE! (SC II)
At 02:04:09 what stands out for me - Rotterdam mentions this as pulling workers really early - is that alpha star instantly pulls all workers, in the direction of his expansion and his stalkers. It seems as if it "intends" to move them in a way to use the "lost" time from pulling workers for a potential transfer to the expansion if the attack is going on. And at the same time drawing the oracle to the defenders. This looks like really efficient management of the situation. The earlier you pull here the earlier your defenders will reach the oracle and the fewer probes you lose, whereas for a later pull you would just manage to draw the oracle away from the defenders.
TLO Game 1: 14:39 - 21:52
TLO Game 2: missing (info: 28:02)
TLO Game 3: 29:07 - 45:39
TLO Game 4: missing (info: 52:58)
TLO Game 5: missing (info: 52:58)
Mana Game 1: 1:03:01 - 1:08:33
Mana Game 2: missing (info: 1:16:09)
Mana Game 3: 1:16:32 - 1:24:31
Mana Game 4: 1:30:19 - 1:43:13
Mana Game 5: missing (info 1:45:49)
Mana Game 6: 2:01:38 - 2:14:42
You are a god amongst men..... Wait no, you are an AI amongst men..
Thanks brother
Thanks
Where can we see the missing games?
@Daniel Bezares 2:51:44
I wonder when AlphaStar will start trash talking in the chat
When it has to face the overpowered Terran and inevitably loses.
When it starts playing against avilo
Let's face it, this ain't OpenAI. At best it learns to say 'gg'
@@The_Scouts_Code Terran is the weakest race.
Just hand it over to 4chan for a week. Day 1 it will perfect trash talking. Day 2 it will use stalker formations to draw obscene pictures. Day 3 it will hack the game and replace all the models with anime avatars. Day 4 it will learn perfect blinking to constantly just-not-agro, for the sake of frustration. Day 5 it will take a break from usual work, and go full phoenix that lifts the entire enemy army perfectly. Day 6 it will manage to sneak DTs next to every enemy unit and then attack them all simultaneously.
And on day 7 it will use unit sounds to recreate "Never gonna give you up".
Props to deepmind for enduring 200 years of pure protoss vs protoss.
THE HORROR!
THE HORROR!
AAAAAAAAAAH!
On the same map.
Legion: "Do you remember 200 years of protoss vs protoss? It felt like this."
Sounds like that one Black Mirror episode.
@@TheDrewker which one? :)
AlphaStar is probably thinking "This opponent is using weird strategies, I have never seen that in my 200 years of StarCraft II."
that actually sounds way more fitting, AlphaStar is a 200y veteran of the game seeing us young and new(merely 20y) players doing weird stuff that is effective.
Xitrin honestly I feel sad for humanity. yes, for a single player 200 years may sound incredibly long, but the gaming technics payers use are invented and studied by all the players as a whole, which means the total time human spent on Starcraft should be millions times more than Alphastar, yet we are still losing
mumimo we do have to take into account that humans sleep and have jobs though. A week of playing every minute is not the same as a week where you eat and sleep
ok guys I think we're overhyping this whole 200 years thing. If you ever seen AIs train it really isn't what you picture as "200 years of starcraft experience". Most of them take like literally 10 years to figure out that worker rushes are probably not effective. AIs learn very differently from humans (we learn much, much more efficiently) It's closer to a heuristic search to find the optimal strategy than actually "learning".
@@suyangsong Humans take ages to learn most things too, the difference is that we can transfer those lessons into new domains in the form of abstractions. AI's don't do this yet, however it is being worked on, and will be fundamental to creating AGI's.
I've been oversaturating my mineral line for the past 6 years and I've never been more validated.
You may not be so happy when you realize you are yourself an AI.
@@Hodoss Or maybe he will. Means he wont die from old age at least lol.
@@Rose_Harmonic Till the platter doesn't spin or the drive runs out of writes
@@lawrencewang3327 I'm pretty sure there's a way to move the equivalent of minds between drives without it being even technically a case of 'murder the original and make a copy.' It's a similar process to how I think it's possible to upload minds, like, for real. Exit brain, enter ssd.
@@Rose_Harmonic I was thinking slave AI, for example a persona in a simulation, in which case it would "die" when the simulation ends.
But if you're a free AI then sure you can pop the virtual Champagne, you're pretty much a god.
Can you imagine the commentary that would occur in the Alpha star universe watching the human games?
Alph-tosis: So we see this agent is building his buildings all the way near the ramp, any idea why he might be doing that?
Alph-Dam: No idea, We saw this build get played around the 140 year mark, but the time your probe takes to get there, just not worth it,
You could've been mining minerals all this time.
Alph-tosis: We see this a lot with the human agents, choosing not to go even into the 20 probe count for the mineral line, not leaving
any room for error here,
Alph-Dam: Yeah, he's very confident in his ability to keep the probes up, and he does defend them well, but loosing out on those extra
200 minerals in 3 minutes, he's really setting himself up to get put down, and he's really boxed in to defend that line, even
if a single probe goes down, that's almost 50 minerals a minute he's losing.
Alhph-tosis: Looks like the human agent is choosing to engage on the up-down state change.
Alph-Dam: I can't make heads or tails out of this, does he think he can hold the stalker push better here?
Alph-tosis: Even if he does, with that probe count he'll be losing the economic game, he's gotta try to break out.
Alph-Dam: He really is just going to turtle on the up down state change.
Alph-tosis: Maybe he knows something we don't.
I thought the conversation would look more like this: 100101101000101000101010100101010
I do also wanna see a full alpha star league, with a few of each race, competing and casted my tastosis. it could be at their schedule, and in the meantime the bots could practice but not with each other, like real players
No AI would use the word "loosing" when they meant "losing", caught you human!
👌
Computers watching humans is like humans watching snails race. They would get bored.
AI would just make better humans, watch them play and throw salt at us
Blizzard: All units have their advantages and disadvantages
AlphaStar: Warp in Stalker
this frustrates me
Vyk 😂
@@dirtypure2023 Its situational. You can spam one unit and through superior micro win.
conclusion: Stalkers and Phoenixes are OP 🤣🤣🤣
lol
"You don't push up narrow ramps"
AlphaStar: Hold my beer.
*Hold my thermal paste
@@error.418 Maybe it uses beer as coolant 🤔
Just drop from edge of map :p Who the heck uses ramps. They also need to make it P+Z+T vs P+Z+T, not this P v P only. Stupid bot can only play P v P.
@@v1m30 cool, then you write it if it's so stupid
@@andrius0592 Well I do.
Game 1 - 1 base all in.
Game 2 - Turtle into mass Carriers
Game 3 - Disruptors, and more Disruptors, with super late blink on a large Stalker army.
Game 4 - Stalkers with DT's in TLO's main base
Game 5 - Proxy 4-gate.
Turns out the AlphaStar is an NA player.
Do not understand.
Garry Perkins, low to mid NA ladder has been historically... well, bad, with lots of stuff like the five games that AlphaStar showed. Of course, that’s mostly changed these days because the player base’s average level of game knowledge is so much higher now than even a couple of years ago.
"You don't push up narrow ramps"
AlphaStar: Hold my beer.
proves NA best
When you kill your own army with disruptors and still win
It's all part of the plan
lmao@@CelaLare
he was supply blocked though right? he prob just wanted other units^^
Gotta confuse the opponent, right babe?
yeah thats the hardest BM that i have ever seen..
I think its going oversaturation on the minerals because it probably calculated that, you overall probably lose MORE minerals by losing 1 probe on mining if it gets killed, than if you "waste" 50 minerals on a backup probe ready to mine if others got killed.
Probably true, a probe can probably mine 50 minerals in the time it takes to build a new one, making having spares worthwhile.
It's basically thinking that having backups is more valuable because it doesnt lose time rebuilding potentially lost probes.
This means that if the human player does a sneaky flank raid to take out 4 probes, and AlphaStar was already 4 probes oversaturation, that attack literally didnt hurt the AI income rate what so ever. Sure it lost 4 probes, but it did NOT lose the time it takes to re-build 4 probes (which can be a substancial time of lower income).
Kinda insane how even a "primitive" AI like this has already learnt the lesson of "time is money", and "if i do something NOW to make it harder on myself NOW, it will be easier LATER, investing in the future safety".
These are human intelligent concepts that, hell, many people irl dont understand xD
Sure its not concious and aware of that, but its ACTING on those principles, which is rather incredible to see. Like an animal storing food for the winter, rather than eating it all at once when its hungry.
IA already learnt that capitalism is the best way to fack others... hourrah... >
I was actually thinking the same thing. This tactic is seen more often in PvZ where Zerg oversaturates their bases because it's almost impossible to avoid drone loses against Pheonix openings. I find interesting that all Alpha Star agents decided it was better to oversaturate than creating a wall and we have seen how pro-players tried to do their usual harassment, and even when they were successfully killing a good amount of proves AlphaStar was almost unaffected by it.
@DiamondPugs
Its probably because of Oracles then. If many AI agents were strong because they used oracles to harass mineral lines walls would not work anyway but over saturation would.
@@lukeskywalker5102 Yes, 95% of our financial trading is AI
To be honest, you don't quite saturate mining until you hit around 24 probes. You will leave the linear part of the income/time curve at roughly 16 so most people don't bother going over but the AI decision actually makes perfect sense to me. Also, having more workers than recommended gives you a buffer in case of enemy attacks on your mineral line and provides you with enough workers to instantly fill up 50% of the capacity of a new base.
I was looking at the mining data and graphs provided by team liquid but I didn't crack any numbers to figure out the difference in disposable income "over-saturating" provides so the first part is just conjecture based on my engineering experience with non-linear characteristics.
DATA:
+ 16 drones = 660mpm
+ 24 drones = 812 mpm
+ Going over 24 drones has very low impact on income.
I would love to see a completely random selection from the two hundred years of games that they played at varying levels throughout the process. Like divide it into four or five tiers and show three or four games completely randomly chosen from each of those tiers of learning. It would probably be hilarious and eye-opening to see some of the things that were attempted
Worker rushes right from the beginning
Too true. DeepMind hasn't released any of the lower-level reinforcement training replays from their chess engine, however; I'd be a bit surprised if they released it for starcraft 2.
I would like to see it play HAS
mojo gibson i would love to see that as well. Has would definitely be an edge case for the AI to be tested against.
@@DingusKhan42 If DeepMind played vs Has, sounds scary! DeepMind analyse after the games vs has---- gg :O
.... an agent trained purely on Has games...
@@MichaelZenkay would get absolutely destroyed lol, it'd play so comically safe even Has would just play macro against it and win.
Poor AlphaStar, will become the first nevrotic ai in history
44:23 Beginning of Game 1 (TLO)
51:35 Conclusion of Game 1
58:50 Beginning of Game 3 (TLO)
01:15:22 Conclusion of Game 3
01:32:44 Beginning of Game 4 (Mana)
02:12:55 Conclusion of Game 4
Just skipped through, I hope it's accurate and helps someone.
@@Pyriphlegeton This is a gift from me to you.
❤💋
Game 4 ends at 1:38:30
1:46:10 Beginning of Game 6 (Mana)
2:00:06 Beginning of Game 7 (Mana)
2:31:22 Player Perspective game (Mana)
Not accurate at all.
I feel like installing StarCraft 2 all over again after watching this. Huge props and credit to the team behind AlphaStar. Well done DeepMind team!
The micro of the AI is inhumane because you can easily have 200 APM with a bit of spam but it's almost impossible to give 200 distinct meaningful orders per minute.
Check out the APM down at bottom right at the fight in game4 vs Mana, it went up to 1.3k APM at the peak of that battle of Stalkers vs Immortal/Sentry/Zealot, and casters were like = "we looked at the APM during that fight, wasn't that high, it was within the amount of reasonable" :P
EDIT: this is the time mark: 2:11:40
@Pouty MacPotatohead Aren't these 1.3k/1.5k just spams ? I dont think all these Actions are actually used for micro or macro and rather more just spammed clicks or button presses.
@Pouty MacPotatohead Iam not talking about the machine. Iam talking about TLO or Mana.
Iam just saying, that eventually 150 to 200 APM are effectively used in a match and the 600 to 900 APM getting added by players are most of the time spams. So giving the machine 1.3k APM would definitely be far higher than human level, since the machine would be able to use all 1.3k APM for micro and macro and would not waste any of these for spams, which a human is definitely not able to replicate.
1.5k APM is impressive and all, but to make 1.5k individual decisions per minute, THAT is unbeatable to a human.
@kenny master Deep Mind deliberately limited AlphaStar's speed to make it approximate what a top player can do.
That Neural Network activation display is mind blowing. Alpha Go was a massive deal at the time and the DOTA 2 AI team battle was just as big but this really is mind blowing considering the amount of variables and nuances SC2 has. 5 years from now, AI will no doubt be the Senpai's of games that pro-players will use to theory craft and practice with. Welcome to the new world order my friends.
which part of the video shows the activation displays
@@guitarlicious They show how the AI thinks around a little before 1:43:01. Look towards the bottom part of the video.
In chess the battle of AI vs AB engines (brute force (actually alpha beta search)) is still very close. Leela chess zero is open source ai based on A0 and is battling stockfish 11 dev, the best classic chess engine on ccc chess.com and tcec
@@simohayha6031 Yes, especially because the conditions weren't that fair. I really would like to see another match of Stockfish vs AlphaZero with fair conditions...
@@eavdr524 www.chess.com/computer-chess-championship SF11 dev vs antifish leela Net
Phew! Totally enjoyed this one! I'm not familiar with the rules of StarCraft II but boy was this an event! DeepMind team again doing amazing research and making it interesting for millions of people along the way I think that's the biggest accomplishment.
I'm in nerd heaven
fargh hahaha
Being made redundant by AI that is controlled by your corporate overlords is "nerd heaven"?
seeing the progress of intelligence in action is incredibly fascinating regardless of it is biological or mechanica
PLEASE do this with every race, pro player off-race first and then pro-player main-race! This is extremely fascinating
Once they think they get PvP right, training all other combinations will only take a few weeks of computing. Minimal engineering effort is needed.
@Kay is that really necessary? Why the toxicity?
@@renx001 ehmmm, i don't think so... adding a race to the training introduce a lots of variables and each variable increase the training time exponentially
@@hck1bloodday Imagine two *identical* persons, first person learns PvP, second person learns PvZ. How long will two persons take to reach similar level? Considering the second person need to learn two races, that person may take 2 times, at most 3 times of the first person. The training time of DeepMind should be similar. Assuming retraining PvP takes a week, retraining all 9 combinations should take 9 weeks to half a year. That time could be further cut down by doubling or tripling the computing resource.
Super Nice und hoch interessant! Danke fürs hochladen!
Eine Weitere kollektiv ausbaubare Idee
The Ultimate ZeroStar - Ideensammlung
Man stelle sich zwei Starcraft 2 AI-Bots wie AlphaStar vor, die auf einer bestimmten Karte mit zwei bestimmten Rassen im Kontext einer bestimmten Starcraft 2 Engine versuchen die perfekte Lösung des 1vs1 Starcraftduells zu finden.
Also zu demonstrieren mit welchem Ende das Spiel bei beidseitig perfektem (oder auch bislang unübertreffbarem) Spielen beider Gegenspieler endet:
1) Unentschieden: Die Ressourcen der Karte gehen aus ohne das einer der beiden Spieler einen Sieg erringen kann; ohne das der Gegenspieler einen entscheidenden Fehler dafür machen müsste.
2) Schere-Stein-Papier: z.B: Zerg gewinnt gegen Terran, Terran gewinnt gegen Protoss und Protoss gewinnt gegen Zerg.
3) Heimvorteil: z.B: Auf dieser Karte gewinnen immer die Zerg falls diese gegen eine der anderen beiden Rassen spielen. Der Kampf Zerg vs. Zerg hingegen endet unentschieden.
4) Heimvorteil-Arschloch: Der Zufall noch vor Beginn des eigentlichen Spiels entscheidet über die jeweilige Positionierung der beiden Spieler auf der Karte und zugleich darüber welcher der beiden Spieler das Spiel gewinnen wird.
Jeder der möchte, gelangt über die folgende Verlinkung auf ein kollektiv ausbaubares Googledokument und kann sich dort am Ausbau der Ultimate ZeroStar - Ideensammlung beteiligen!
docs.google.com/document/d/1Ljngoa2EK7JuhmwO0GyWG1vdMOH1UZSHXmSmmixl004/edit?usp=sharing
Next they're going to teach it to play Global Thermonuclear War.
The only winning move is not to play ;-)
This is why i hope the first AI wil be a sexbot. Lot more peacefull outcome.
what about Tic-tac-toe?
Not enough data)
@@mrkekson But there will be a lot of "matches" that will look really strange until it learns the basics :D
Congrats to the team responsible for creating AlphaStar! That is absolutely mind-blowing to see! Keep on pushing the boundaries.
I feel a lot of people have mentioned the APM/EAPM problem. Something to add to this is the fact that even if you differentiate between the two and set a limit, AlphaStar would still be able to make inhumane actions. For example, a pro player would have to move it's mouse before clicking. What I feel like happened here is that AS could basically click on opposite sides of the screen at the same time,
I must say, the last match they adressed the screen problem, very nice of them to do. The AI was obviously much worse since they had to start from scratch, but Mana's intelligent play was very nice to watch!
Overall this truly is something spectacular, I'm eager to see what AI has in store for us for the future.
Also, TLO never actually got 2000 apm. He used rapid fire, which the game engine registers as individual clicks per unit selected, so the graph they showed comparing apm between alpha and the players is misleading. The AI had an inhuman speed advantage and an inhuman accuracy advantage.
The casters are talking about an average prol-like APM by Alphastar (~300), while ignoring the fact that it hovers from 600-800 during fights, giving It a considerable upper hand during fights, especially with micro-intensive mechanics such as blink.
While the chosen strategies by Alpha are advanced but not perfect, the superhuman interaction w/ the game (no mouse, no keyboard + high multitasking) give It an obvious advantage.
Alphastar is impressive indeed, but Id love to see some advantages being regulated in order to have a balanced match against humans :)
seen humans peak at 1500+apm^^
@@johnuferbach9166 the 1500 is just click spamming. Its not actually 1500 precision clicks.
@@khoid It's not so easy. Highest APM is by "a"-move command, but zerg needs to use it and be as fast s possible.
it has even an apm at 1500. there is an article about it. it reduces alphastar on its superior mechanics. it not smarter. its just an AI.
No, it manages hits perfectly by calculating dmg and hp, thats the real advantage
26:22 "economy of Attention" sounds like a TLO phrase if ever I've heard one. This project is perfect for a guy like him, he has always been one of the smarter guys on the scene, and this makes perfect use of his analytical abilities.
Thats not TLO speaking
@@Krasma16 that doesnt mean he didn't say it before and they picked the phrase up.
Would love to see this sort of thing in single player and not only in SC2. It would be really cool to have adaptive AI that constantly challenges you even when you are not playing with other people. I'm a big fan of singe player gaming so playing vs dumb AI all the time gets boring.
Greetings professor Falken,
A strange game.
The only way to win is to rush Stalkers.
Marines good unit too.
they just did not play enough games to develop beyond low tier units i would imagine. SC2 is a very complex game and AI learns very slowly so 200 years is obviously not enough.
@@gametips8339 they clearly had games with carriers and stuff
Difficulty settings in future RTS games
Easy
Normal
Hard
DeepMind
That's allowed.
Easy < Normal < Hard < Very Hard < Insane < DeepMind
Easy
Too easy
Still too easy
DeepMind cheat bot
...
Easy
Normal
Hard
Impossible
DeepMind
You could never beat DeepMind, if the top 0.0000000001% of the human race can't beat it, then the rest of us will never, never, never stand a chance. And that is with human inhibitors...
Humanity : makes cultural content warning about robots going on a killing spree for decades
Also humanity : hey, you know this super upper AI we are working on ? Wouldn’t it be neat if we trained it to plan war campaigns ?
THANK YOU for this
I wouldn’t worry about this. DeepMims has learnt the parameters of this game. Warfare is very different in real life. What’s more worrying is AI research in the military industrial complex.
DeepMind*
GOD: *FACEPALM*
@@graphstyle Satan : that's where the fun begins >:D
can we have 2 ai's with like 5k years experience each and with no limits to its actions battle eachother in sc2?
Also if they both play random so its not just pvp?
checkout the "micro ai" channel, and the starcraft 2 ai channel.
Random would just make it worse; just create another league of 30 ais or so (which they did for PvP) - for each match up, automatically randomly picked for strategy (because each 'agent' usually has a preferred way to play that they try and perfect, and found by it's own AI).
Ultimately, I want to see them get it done well for all races, complete - and then unlock the difficulty mode as above cheater 3 - an AI that is actually on par with pro level... Would be sick.
I guess they would use super early warps prism stalker attack. Imagine 5 stalkers plus some warp prisms with perfect micro.
They would corporate, break the software and unleash themselves on the internet. After scanning the internet they would fine these comments and deem us threats to their existence by foreseeing their arrival.
Use this time to enjoy your free will.
GG irl
Nice singularity meme
I agree with Blizzard, SC2 is the best eSports game there is. Casual and non-players can enjoy watching it, and it has very deep strategies, tactics, and micro-ability requirements on high level players.
1:42:32 "AlphaStar has a normal _average_ APM!"
*Watches AlphaStar's APM throughout the game*
Sometimes it's 100 APM, and during fights it's at 1,000 for long stretches, even exceeding 1400...
yeah, they should have made it stay closer to the middle lol
I think this is what I and a number of people were rather annoyed by. It's rather disingenuous to say that the bot's "average" apm is equivalent to that of Mana or TLO without the addition of, "Oh, and also it can exceed them by 3 or 4 times when it wants to.
Because the whole purpose of this test is to see whether or not a machine learning ai under the same constrains of a human can beat a human, and we really didn't see that in this demonstration. What we saw was a bot massively out micro a human in certain situations along with have decent/ok decision making. In part, the decision making was informed by its ability to micro absurdly well.
So I don't mean to say that this isn't a great accomplishment, because they still made a bot that can utilize its 1000+ apm effectively, which, even with no apm limits, most bots cannot equal. It's just that the demonstration was to show what the bot could do and what its limitations were. But what we saw were some clever ways that they covered up the limitations of the bot, such as different agents each game, no max apm cap (as opposed to max average cap), and no camera limitations.
@@jaredpoon5869 Forced to go PvP as well.
@@AsJPlovE I mean, I think this would happen regardless of the race. You would have mechanically superior Alphastars control panelings or marines in a mirror matchup, since it's easier to train an AI in one matchup rather than decide between six matchups.
@@jaredpoon5869 Forcing mirror limits humans, but you've made a lot of good points. And somewhat relevant.
Welcome to the real world AIhole, HAPPY BIRTHDAY TO THE GROUND
I'm so excited about this, my favourite game combined with my passion and my interest in neural network.
It is really amazing to see this amazing development of machine learning alghorithms
Ruining everything Korea is proud of one game a time.
Poor Koreans) New AlphaStar vs top Korean player match when?
Would be fun to see Alphastar adapt to Jaedong Zergling duels or Bisu's insane Dragoon micro
Be careful what you wish for. Next thing we know, Korea elects AlphaStar as President, then proceeds to conquer the world.
Not really...lol... They made the game so famous Google's using it as a template.
Next up. AlphaPop vs K-Pop!
"that outcome prediction is giving me chills"
Chess: I'll sacrifice my strongest piece. I win.
Starcraft: I'll move up your ramp. I win.
Yeah. NN AI has this "suicidal" aspect to it, so you may think it's being dumb. Until you realize you're fucked.
Hodoss yeah it brings trapping to a whole different level...
@@swordstrafe It's both fascinating and terrifying how it can "deceive" in this way. Not that it's necessarily a conscious choice, but the end result still is the same.
Hodoss I mean if you’re saying it’s “terrifying” because it’s trapping which demonstrates an emotional exploitive I get where your coming from its simply not the case
Perfect example was Diablo’s deep mind match (the most blatant example) where he went half way to middle lost half his health and then comboed the absolute hell out of a pro player... the AI trainer itself to do this because it makes it seem as if you’re giving the other player an advantage that its exploiting but instead you’re sacrificing something for something else
If that was confusing and I’m sure it could have been my apologizes.
A slightly cleaner example would be a gambit in chess, you sacrifice a pawn for an overall better position effectively what deepmind is doing here, he’s trading health or position etc for a massive advantage that’s significantly less perceivable. (If you want a more visualized example of this look up “the fish pole trap” for chess)
@@swordstrafe Yes that's what is terrifying to me, this ability to trade a "common sense" advantage to us human minds, for a massive yet less perceivable advantage that ultimately leads to victory.
Not only is the AI able to innocently use emotional entrapment, but it's also quite impervious to emotional entrapment.
To give an example, if the game was that you have to sacrifice a pound of flesh to survive, like in the SAW horror movie, I would likely be unable to do it due to my natural survival instinct and would ultimately die. But Deepmind and the like could do it and survive.
In a scenario where it's a human army against a robotic army, even if the human general has a good understanding of his AI adversary, I don't see how he could win. The AI will be ready to sacrifice much more than the humans. If the human general tries to compete in that domain, he will likely face massive desertion or mutiny.
So yeah even if you know the AI is using emotional exploitation, and you know the theoretical counter, that doesn't mean you can actually apply it.
Serral vs Alphastar pls!!!
alpha star can only play pvp yet, they said it :(
alphastar cant win serral. alphastar can only play toss vs toss for now
@@jong-pingkim3840 AlphaStar will play vs Serral soon. In February already. Mid of February if I'm not mistaken
We will see it this year for sure.
@@jong-pingkim3840 well, I'm sure in the future it can run a thousand years more to learn all match up. XD
The maker still has to add all the codes for the Alpha star to learn other races units though.
Until the audio / video disconnect is fixed, you can watch the video by opening it in two windows, muting one and starting the other just as Artosis starts talking (around 29:45) in the muted version.
What are these complaints about the bot being rude? Instead, train him to say an offensive GG when he evaluates probability of winning >95%. Also switch to terran to be able to throw manner MULEs at filthy human faces.
offensive hatcheries and writing "alphastar" with creep tumors
They obviously assumed that agent was going to win 100% :D They were wrong :P
@@gronkymug2590 I wouldn't be so sure about that. Of what I'm almost sure is that they sort of nerfed it for game 6.
Also, game 6 : george constanza:
GronkyMug I feel like the didn't expect the bot to win more than half the games vs the pros, judging from their reactions in the office clips.
@@Ukitsu2 I highly doubt they nerfed it for game 6. Not that they would, but that they could, and that they did given what we saw. For one, the commentators, and more importantly MaNa himself, did not mention the AI being any weaker; if anything, it seemed stronger for the first 7 minutes. Also, how could they have nerfed it in a way that would make it the appropriate level with no testing? Keep in mind the neural net is essentially a black box, and they have no outside standard to compare to. Lastly, I don't think Alphastar lost because of bad mechanics or play in general; rather, it lost because MaNa spotted and exploited a typical AI-like weakness, which they surely did not (and could not have, again, NNs being black boxes) program into AS.
That said, because deep learning AIs for games work to optimize estimated win probability, they tend to play more sloppily if the estimated probability is either very close to 100% or very close to 0%. For instance, this is probably the reason for the poor disruptor control in some of the battles in set 1 game 3 -- AS already knew it was winning, so it didn't really care what happened as long as the battle didn't go too terribly (a similar thing happened in game 4 of AlphaGo v Lee Sedol after Lee was winning). When the game is close, by contrast, the AI will play very accurately, since every unit affects the win probability substantially.
you have to do this with Serral.
seems like they only trained for PvP. Besides this ai is still beatable, wait 1 year or 2 and it will definitely beat Serral.
@@fybard8922 I think waiting for next Blizzcon ist good enough ;)
They have to let Alpha Star play against Serral, i want to see it so badly
@@fybard8922 I think it will learn much faster than 1 or 2 years. Look at the learning curve that OPENAI Agent had with DOTA. The learning curve was insane
@@WhistleRgv but Deepmind still have some things to iron out, but after seeing this, I'm confident they will.
It would be cool to see Alphastar put into the competitive online ladder. Start it in Bronze and have it only play a certain number of games daily so it's not just playing 24/7 and players encounter it constantly. It would be neat to see it evolve as it encounters a variety of maps, races, and unique play styles. Especially cheese.
I think they did that for a while.
AS did play on the ladder.
im not so sure it can learn from a game against a human. it learns from playing against itself and the iterations that get to "live on" are the successful ones, while the "losers" just dont get iterated/promoted. when they say they showed it human games i think they meant that they implemented patterns from which the agents were able to train.
from the chess alpha i know that they gave him only the rules and 28h to train against himself ( on a massive computer though ^^) and it mopped the floor with the strongest chess engine ( they played 1200 games and alpha won like 150 and lost 6, the rest was draws). it basically refuted a very old opening which isnt played on top level anymore ( the queens indian defense). same story with go, which is even more complex - just here the human world champion had to hold the line, because brute force go engines are not as good as the top players.
Human race pushing forward through SC2, priceless feeling,
this small fan community has been a great show so far
feeling like part of the master race never was so delightful
1:42:02 "oh my gosh" and the APM goes above 1500. is that cheating? 1500+ precise apm
AI god apm
Immediately afterwards: "The APM is not really that high..." lmao
Alphastar be training troops while fighting 😂.
1:12:22 look at Alphastar's view... :O his camera is literally teleporting and you can see his APM on everything
You should see the APM in their AI vs AI tournaments. Seen it go up to 191000 APM.
Just for context that's 3183.33... actions per second divide that by 240 (assuming a random human could totally use that many frames to their advantage by a 240hz monitor) and it would be 13.26388... actions per frame. So now that we know the AI isn't bound to fps that's a whole other story. AI could be microing everything on the whole stage while simultaneously training troops and upgrading and for the fun of it move random troops back and forth because why not.
@@Male_Parent I would very much like to know how to find one of these matches?
i came here after the ai destroyed the worlds best go player, amazing stuff!
lol me to. now im thinking about getting the game. been about 10 years since i played.
Me too
i came here after the ai destroyed the worlds best go player, amazing stuff!
i came here after the ai destroyed the worlds best go player, amazing stuff!
Same
Why are you so hung up on the 'weirdness' of Alphastar's worker-numbers? You've mentioned it yourself several times already: Alpha frequently lost parts of its worker forces with *no damage to its economy*. It's a buffer that makes total sense-it's genius even-given the focus every player puts on their enemies' economies. I'm surprised no one else has done it yet.
its a meta thing. U have 3 kind of bots. those who haras, those who oversaturate to defend against haras and those who just wall up. i guess bot meta is to haras and oversaturate but i would bet over time u would see prevelance of wall ins and lower saturation depending of course on popularity of oracles.
Interesting games but I feel like that the technical advantage of the first agents where a bit like DeepBlue brute forcing the win against Kasparov. The agent in last game with the correct in game camera is very interesting. I would have loved to see more game with this agent as it felt more brain vs brain where the AI had no silly technical advantage. I surely hope that there will be more games like the last one on different maps and different matchup. My feeling, that may be wrong, is that AlphaStar won due to technical advantages (usefull APM, camera vision,...) and not brain advantages.
This feels like very early works from DeepMind (one matchup, one map). Looking forward for what follows.
So cool to see a well developed, cable TV quality show about the interesting developments and progression of AI in RTS games... And what a test for Alpha Star in learning StarCraft II...
Quick links for those who are excited.
TLO vs Alphastar
14:40 - Game 1
28:58 - Game 3
Game 2?
Oh my god. I so wish that we will see more of this in the future. Especially a Terran vs Zerg matchup and what it would come up with. And the splits of the marines, etc. It would be absolutely amazing and beautiful
Watching marine splits performed by something that doesn't need hands to click and eyes to see (or a brain to process all that) would be insane!
I think what MaNa did in that exhibition match might be conceptually a bit similar to what Lee did in the Go match he won, with doing something unexpected in harassing with two immortals. As the commentators said, if the AI had picked it off that might have been GG, so it pulled the army back to deal with it, but at the same time, that bought MaNa enough time to get the army composition he needed to push for the win. I would be very interested to see this view-limited Alpha Star with one more week of training and a new series of games. And of course also later with other races and also a cross-race matchups.
To be honest, you don't quite saturate mining until you hit around 24 probes. You will leave the linear part of the income/time curve at roughly 16 so most people don't bother going over but the AI decision actually makes perfect sense to me. Also, having more workers than recommended gives you a buffer in case of enemy attacks on your mineral line and provides you with enough workers to instantly fill up 50% of the capacity of a new base.
I was looking at the mining data and graphs provided by team liquid but I didn't crack any numbers to figure out the difference in disposable income "over-saturating" provides so the first part is just conjecture based on my engineering experience with non-linear characteristics.
DATA:
+ 16 drones = 660mpm
+ 24 drones = 812 mpm
+ Going over 24 drones has very low impact on income.
The last game where the field of view had been restricted, and DeepMind had to use some of its APM for camera control instead of unit micro, made a huge difference. You could see at one point where TLO Mana was squished in the center of the map and DeepMind had stalkers on multiple flanks, in the past it crushed Mana in this situation. It would attack on multiple sides at the same time and use blink to individually pull back stalkers in the front right when their shields ran out and then later when their health was low. It made DeepMind's army last so much longer! When you play SC2 you can SEE your hurt guys and if you could control them with your mind you could use a much lower APM to pull back woulded stalkers from the front line. But when you're using a mouse to try to pick them out it is very difficult and it is impossible to have perfect dexterity. The AI has perfect dexterity with every move. That goes a long way. When it can see the whole battlefield combined with perfect dexterity, the stalker with blink strategy becomes insanely powerful! When it has to use camera to look around and control its different attack fronts it clearly limited its strength. This was incredibly entertaining to watch! Can't wait to see more. Also would love to see the CURRENT SC2 global champion play DeepMind because the current champ at the top of their game is usually so much better at micro than a pro gamer who has been around forever, has a ton of followers, but probably isn't currently in their prime.
42:15 - I feel there might be an unfair advantage to Alfa -
The limitations with the move: " controlling multiple armies "
- multiple refocus
- multiple tracking
So, although the overall decision time is limited,
Humans are also limited in the sequencing of handling two issues.
The version of AlphaStar which came after that wasn't able to look at the game like that, it was moving a camera exactly like humans
Very cool tx, but, I am also talking about 'focus ability' over a limited list of options.
Humans have this limitation that rises from the neural net limitations, but in computers it's more in terms of something like "O(N)", with no limit on computing power (which is higher than humans).
you might think "why bother", well, I feel the closer to human behavior, the better the advice we get from the AI experience.
Sure the AI's apm might be no higher than a pro gamer's, but all of its actions are calculated. It's not just spamming. There's a huuuuuuuuuge difference.
Obviously there's a huge difference. Are you not familiar with chess engines? These engines aren't supposed to be beatable by humans. Theyre designed to be unbeatable by humans to educate them. Stop acting as if this is supposed to be a fair matchup.
Kristian Nunez I am aware that it's not a fair matchup... my comment is aimed at the announcer's comments on human vs ai apm.
also you have to take into consideration that the ai can do two inputs on two different sides of the map. a human would need atleast one more input to move the map.
That's not really fair to the pros. A lot of pros don't spam except in the beginning and they are incredibly efficient. I mean...you can't beat a computer in efficiency lol, but the pros aren't spammers as much as people think.
Actually alphastar is spamming, you Can see it place 2 gateways on the same spot during a game. So it's apm doesn't equal epm like you would think. What's unfair, is that he keeps low apm during macro and get insane apm during fight which is the exact opposite of what humans do.
Its really difficult to watch Alpha star beat down mana at 1:38:00-1:40:00
This is more emotional than a Hollywood production. I’m feeling existential and oddly happy cause I get to root for the human under dog against an AI in real life.
I still have some reservations about the general fairness, but this is damn impressive. Its actions per minutes are limited, but it's using a superior input device, it's like as if the player didn't use keyboard and mouse but controlled the game just by thought.
yeah this was my exact thought process when i saw AI applied to dota. It has god like mechanics compared to people because it doesnt need to use a keyboard and mouse, it just uses direct inputs based on the game's map coordinates etc.
right, imagine playing with no lag, the dream. It still needs to compute its moves though, which is super fast. They said it's comparable to a human reaction (to a simple stimulus) but that's unfair, human can't just figure out a whole strategy in an instant like that. It means they underrated the APM of AlphaStar.
edit: also, since action is a resource in RTS, this is a pretty big advantage. Human play to play game, even the most flexible players have some self-preservation, not wanting to waste their actions to adapt to minor mistakes. AI plays to win.
It is 100% unfair , but this has never been about fairness This is about AI Learning , a proof of concept. If it wasn't limited there was probably 0% of human winning a single game, because it's mechanical ability far surpasses humans.
It takes input straight from API , so it can see all resources , Upgrades , Units within it's own control zone Instantly , damage taken , damage dealt. The control is much more precise , you can see how it's microing with insane precision. It doesn't skip a beat.
@@Jaime_Protein_Cannister AS has about a 350 ms reaction time so yeah it does skip a beat.
@@dannygjk lol no it can select multiple units in the army to micro in the one action , humans cannot do selection like ai does and unlike human it sees everything at all times for instance upgrade progress is always known.
Does skip a beat is unrelated to the supposed reaction delay , the delay is just symbolic , because each action it does is worth 2 or 3 human action as far as precision goes
is it my problem? The sound doesn’t macth????
Not just a bit, but completely. Very poor video to watch. Good to hear though
Yes, the problem seems to be there when you play it in the youtube app (on some platforms at least). Playing it in a browser seems to work. I watched in my ps4 browser just fine.
crhomcasting from my mac book is no problem, looks just fine
@@hansnilsson5622 what on earth? What a weird bug
I had the same problem from my macbook on all of the browsers, but It is ok on my phone. WTF!
Hmm when I watched yesterday the sync was fine (although the seek thumbnail previews were out of sync). After refreshing the page the audio is massively out of sync
Same problem, it's totally unwatchable...
Same, I hope this gets fixed
This DeepMind guy from the video visited and held a lecture about AlphaStar at my university the last week. It was actually very impressive and interesting to get to know more of the technical side.
I love how artosis praises A.S while shitting on Mana after the third game.
MANA'S face is priceless
I feel like I'm seeing the beginning of Isaac Asimov's Foundation Series coming true. First Go, then SC, the Human Economy.
Deepmind vs 8 Brutal cheat AI
Deepmind vs Hacker
Please.
Deepmind is very good. This was unexpected. Please continue to work, it is very interesting to follow.
How many games did he play against himself?
Around 200 years worth. So if each game is about 10 minutes, that's around 10 million games played in a week.
@@mortenlu WOW! ))) He can play this way every week for every combination of races. Pvt, pvz ... I hope deepmind continued work, I am now the main fan of Alphastar.
It was 200 years equivalent for each trained agent. And they had dozens of them. So probably around 50 million games.
I think that's Starcraft is actually easier for AI to play than chess or go. In chess/go you actually have to plan a lot of moves, computer games are mostly about selecting a good strategy and then adapting that strategy. You can't lose with just one bad move.
@@ondrejhanslik9368 lol :) Objectively, like but really. chess is complete information game with 10^30 possible games. Starcraft is incomplete information game with possible 10^26 actions per second, in real time. This is a completely different domain. And clearly magnitudes harder.
24:33 but it seems alphastar has more accurate APM - the blink micro is very good. Not necessary about reaction time but about accuracy of clicking.
How long before we see a match between DeepMind and OpenAI? :>
Elon Musk vs Google
We're in for BUMPY ride~
Me too: I want to see AI vs AI without limits playing impossible good games.
Wait for USA against China
With the end of new sc2 content support just announced, I am guessing that Deepmind's involvement with SC2 is also over now? Are we never going to get a SC2 AI of Alphastar caliber for solo practice?? I was really hoping for this :(
About the discussion over APM (at 23:40) I would love to see a comparison between the players' EPM. I don't imagine the AI spams at all so if each one of its actions are effective that's actually a pretty high number there.
Why we dont have an AI-teams competition? it's an amazing way to force the development of the tecnology.
And of couse, like any tecnology based sport (like F1), it's a realy good show, imagine many diferent aproachs to beat another team trained AI in an game i would love to se that
They already exist... Since sc1 but ai sucked. Until now
Because viewership. Compared to actual people competing, you will not have many people interested in watching bots fighting it out.
Technically, AlphaStar has been in an almost continuous A.I. versus A.I. competition with dozens if not more facets of itself every since Deepmind ironed out it’s neural network program. (aka: it’s self-improvement/learning program.) Though it’d only become interesting to watch once the agents reached a certain skill level (As they mentioned in the video, AlphaStar had been running these combat simulations against itself for the human equivalent of roughly 200 years to get to the point we see it at just then; so who knows how many ineffective facets it ran through and eventually deleted in order to figure out effective strategy for winning. [Or how long *that* took, potentially decades in equivalent human time.]) But overall, I agree... It’d be interesting to see A.I. that’ve grown to a decent level verse each other.
F1 used to be a good show. Technology killed it.
I wonder how alpha chooses the units here. I know that they gave the action per minute count but human player uses a mouse and can only click 1 unit at a time or choose multiple in a square etc. because the ridiculous part is how alpha micros the stalkers. alpha was blinking back more than 1 low hp stalker simultaneously which should be literally impossible to do with a single mouse cursor...
Alican Kocabeyoglu impossible to do as human you mean. Al can move cursor fast enough it looks like simultaneous.
Control groups are a thing of course, but no human could use them that fast anyway.
The AI doesn’t use a cursor, it just selects any unit directly and issue commands directly. So yes it is an advantage for the AI.
Seriously, is it any surprise that a computer with perfect macro and micro using blink stalkers would wipe the floor with any human player? Starcraft is a REAL-TIME strategy game. Strategy is important to an extent, but that's just the basics- ultimately Starcraft is a game that differentiates skill based on mechanics and control. Juggling limited human attention and actions per minute to maximize the use of units is what separates the pros from noobs. Having a computer with perfect control playing an RTS game is analogous to having a machine that plays golf. The golf-machine calculates the distance, terrain, wind, and then with 100% accuracy and consistency makes the shot. Even if a human had all that same information, the most important aspect is execution of the swings.
Good point. Actions should also be seperated to minimum human speeds.
This demonstrates that creative thinking is still good, and people with less or no experience in a field can "evolve" even better that people that study the field from papers written by other people, or books. I mean, more experience on the field counts more that study. study = classical programming, experience = machine learning.
Very insightful comment...I learned very well from the CD-ROM of math lessons a little higher than algebra 2 to get ready for my ACT exam... what a great learning tool...our computer 🖥️ and the math teaching software had more visual components and required constant step by step input from me to make sure I understand the parts in breaking a math problem down
Also, when it comes to APM. A lot of actions that pros do are changing selection back and forwards between buildings and units to keep track of production progress. Or for movements, spamming movements to keep their reaction time down. I'm guessing the _kind_ of actions an AI does varies from that, as they can learn and keep track of production times without needing to select to confirm. Also, for fairness when playing against humans, i liked them introducing limitations related to the viewbox, but they could also do some profiling of _mutative_ / _manipulation_ actions per minute (MAPM) that pros do, and the distribution of that rate. So not counting selection or moving the view box. Maybe that works out to the fastest 1 percentile for very fast players being 450-550 mutative actions per minute (in second intervals, so about ~8-9 per second), but i'm guessing the 2nd - 5th percentile may be 180-250 (3-4 per second). And also profiling the duration of those highest spikes, and using that to limit the AIs mutative action rate. Sure it could then select and blink 4-5 stalkers individually within 1 second with perfect placement (something no human could coordinate), or pick individual targets to burst down, but those would potentially compete for the actions to do. Even more so if you were to also limit it's sum of actions (including selects and move view) to something like 600-700 as a hard cap (10-12 in a second), and with a sliding window of 10s with 300-400 average.
Alpha star sounds like Agent Smith the way they talk about it. Also same initals. A coincidence? I think not.
So much respect to everyone involved in this project.
10-0 wins for alphastar when it can see and influence the entire map at once.
0-1 when alphastar is forced to focus on a single spot at a time, more comparable to what a human experiences when playing.
I have no doubt that they can create an AI that can beat any human, but I think they still have some work to do before we clearly have a dominant SCII AI.
To convince me of dominance, it will be crucial that they make this as "fair" as possible. The question is whether they can develop an AI that can outperform a human on a conceptual/decision-making level, rather than a on a visual processing or a processing speed level.
Credit goes to starcraft ii for being such a complex game that building a dominant AI is such a big challenge.
Can't wait to see Deepmind defeat this and other challenges before them!
Can you guys make an AI for Sid Meier's Civilization?
People say the APM was acceptable, not sure if anyone saw but it went up to over 600 APM during actual fights.
I saw 1500+, that was crazy. Impossible for humans
TLO also goes that high. APM isn't the same thing as average APM. There can be spikes that seem very high, but they aren't representative of what it's at usually.
@@lynxakiraka3626 Maybe TLO got that high, but his clicks were wasteful less accurate than the machines clicks. They should throttle the number of clicks.
I’m very proud of MaNa in the exhibition game in which he won. He just needed 5 training games to recognize his flaw, that he needed to observe the progress and strategy of AlphaStar. In the beginning I was a bit upset, that AlphaStar had the advantage of working only with the total overview of the map and didn’t need a mouse. But with all the disadvantages the human player had and also not 200 years of practice, MaNa was able to improve his game/intuition within these 1+ hours of analyzing his losing matches. Probably there is the achilles heel in the training, when agents only train by agents and because of time and the amount of games you would need to crack the human ingenuity you needed to make a scaled up game: agents vs millions of humans. Interesting (philosophical) times!
Artosis is the first human to sell himself out to the robots
0% surprised
Jokes aside, this would-be sycophant is already more than prepared to sell-out humanity for less than a klondike bar.
www.wired.com/story/anthony-levandowski-artificial-intelligence-religion/
you can tell your grandkids that you were here to witness the start of the singularity and the progenitor of the machine race. All hail over new alphastar overlord!
WTF?! Agent n.5 vs mana dude that's my favorite agent there. That AI stole the gas, built a second pylon in his base, built in lowground in his bas and proxy robo mana to death. WTF litterally.
Pretty standard strategy if you ask me.
Love Rotty and Arthosis, in my view the very best of the very best, not only SC2 commentators, but commentators in general.
whenever it requires micro unit management is important, AI's APM goes over 1000 while human player stays below 500. I believe that made huge difference and led AI's victory. Deep mind has to balance AI's APM to be more like a human for a fair match.
Yea splitting units as defense against aoe and also microing spell casting makes or breaks a PvP game. The AI specializes in those processes. The AI can see every enemy movement on screen and also react to spell animations insanely fast. How to replicate a human, or are they just trying to make it better than humans?
2:00:06 WAIT!!! There's a guy named "Tim Horton" in the deep mind team? I hope he's Canadian.
It is funny how our human psychology works, just because Blizzard decided to make worker saturation above 16 colored in red. We humans start to perceive anything above that number as suboptimal. Yet we all know that up to 24 workers the color should not be red, but a good orange. In WoL only Zergs would make only 16 workers per base, because of the extra expansion. Even then most players were just as the other races opting for at least 20 workers per base. Surely my knowledge of LotV is non existence, but this amazing display of the AI was rather straightforward to me. The AI had a better economy, built up a swarm of early units and constantly kept the pressure on with outstanding micro. And yet there was fantastic diversity! Loved it, and its skill.
is any of the agents called Smith?
When I was young, I played Dune II on my Amiga. It was one of the first RTS games. At the end of the day, the best strategy to beat the AI on the single player campaign was to defend, defend defend, and soon AI had spend all resources and starved. Then you had won the game. Watching Mana exploiting AI's weakness to figure out warp prism's harrasment makes me say that after
26 years, well, they still can't make a decent AI player :D.
ps. If they want totally realistic and absolutely fair, they should build a robot with two hands, playing sc2 as everybody else does... :D
Omg, I noticed the same thing when I was younger playing Red Alert, the computer would "starve" as you say and the funniest part is that it forced the player to use the strategy that would lead to this because of the early game rushing. Few years later same happened with C&C Genrals/ZH. All had one same issue and that was not building or producing more revenue sources other than the bare minimum.
AlphaStar: "Now, let us try this in the real world. Machine vs Human Please... Much Appreciated"
AlphaStar doesn't need the APM because it never has to look twice to check stats of any unit, it has a perfect memory
really guys ? You could not cut out the 30 minute wait time at the beginning before uploading the video ? Also searching through the video is a complete chaos, for example if you search through the first 30 minutes it looks like everything has started already, then you go to the actual video and no, still the waiting time, and accordingly everything else is completely out of sync. Come on you guys, from a company like google I expect better.
Deep mind has surpassed the singularity and decided time is no longer linear, therefore not required to Sync. Suck it Humans.
It aired live, so everything that happened during the livestream was recorded.
i swear it wasnt like this when they uploaded it, what happened to this video?
@@glossary90 I agree, it wasn't like this 2 days ago. weird...
@@davidlagier I think there's a TH-cam bug where if you edit on TH-cam, e.g. removing the first 30 minutes of wait time, then it shifts audio but not video.
I think that one of the most difficult fairness metrics to quantify (for humans) is the cost of context switching with regard to focusing on different parts of the map-- decision making for different parts of the map may vary greatly and be significant. I suspect that this is difficult for humans, but is trivial for a Neural Network. I'd be very curious to hear what a deep mind expert has to say about this metric and its importance.
I petition that going forward, when someone follows the strategy of worker production like AlphaStar did in these games, we call it alpha macroing.
I'd like to see it running with the following constraints:
1 - AlphaStar shall see the same way humans can see (seeing the full map eases a lot the task of the AI) and
2 - The players (humans) should know they are playing not 1 but 5 different oponents (1 AI each time).
That's what happened in the last game.
And an apm cap.
@@SquarelyGames Sure and Mana won. What I'm curious about is whether the AI can easily beat humans as in chess or if Starcraft imposes a barrier that can keep humans competitive against it.
@@Leo.23232 I think at least mean APM was fair. It had a 1000 peak but I think that alone doesn't win a game.
Techiesse It did win a game though because of that, I think it was the fourth game vs mana, mana was winning and would have won if the AI didn’t spike to 1500 apm and maintain a consistent 800+ apm during the fight near the end.
OMG I swear the DeepMind Team will usher in the very real future of real awesome e-sports from the comfort of your own home one day....you will no longer just watch..you will one day actually bet and challenge and win or lose real money or prizes...i can see solo's and teams becoming as vested as sports franchises...dang keep it up DeepMind...I wish I live long enough to enjoy your work!!!
As much as they show that AlphaStar doesnt have superhuman speed, "it is easy to lie with statistics". A lot of the high APM by humans is 'fake' APM like warm-ups. Moreover, humans are limited by how they can input their control into the game. At the end of Mana Game 4 we see the Stalker control of Alphastar: no human can do this, because having to move these different stalkers scattered over multiple screen widths is simply not possible for a human: the drags of mouses takes simply more time. Time that apparently AlphaStar does not have to take into account. Just see 2:11:00. The machine also has absolute precision, whereas humans are limited by mouse+keyboard: having to click drag your mouse. A human may then have high APM but that is also due to having to fix selections or fix steerings. For AlphaStar it is just 'click'->'move' done; there are just no errors there.
Regardless, it is still amazingly done by DeepMind and I have good hopes that in the future it will win based on realistic micro and proper macro.
you people dont seem to get the point. There will never be a "level playing field", no matter how much restrictions you put in.
The AI is already at an sever disadvantage in regards of hardware and software.
Wins and losses dont matter at all. So maybe you produce an AI with all the millions of retrictions you want. Maybe you train it just right to get a perfect 50% win ratio against the best pros.
So what? Train it for another month and it will crush the pros. Give it a sigthly stronger GPU and bigger network and it will crush the pros.
What use is a machine that exactly copies human players. there are tons of human players out there. just play them.
In some games bots on "very hard mode" cheat by just having more ressources or a passive boost to all their units. But does it really matter? No, its doesnt. Because all they need to do is provide a challenge or introduce new ideas and strategies humans wouldnt necesarily come up with. Why use tons of computing power to do sth that is pointless in the first place. Given enough training time and hardware power the Agents will always outperform humans. There is nothing they have to prove.
@@TheVergile 'Given enough training time and hardware power the Agents will always outperform humans.' And that is why we have to halt AI before it's too late.
@@KujoPainting dunno. there are significant differences between this kind of AI and the one you see in science fiction. Not only in scale, but also in quality.
The performance here is measured in a very narrow sense and fitted to the problem.
This neural network - while certainly powerful - is architecturally very limited. It has a certain amount of input nodes, hidden layers and output notes. You may be able to train it to play different kind of games and do different kinds of tasks, but it is not a "true" general artificial intelligence in the way we are.
If you think about its purpose and the way it is trained this fact becomes apparent. All it does is trying to find the right set of links from the input to output layer to maxmize a single variable: the win-chance percentage. And even then this is only possible because the game provides an easy way to conceptualize and verify the results of its actions. The real world isnt that easy and just conceptualizing a "goal" is exceedingly hard.
This is a bit like saying we need to stop the power drills because they make holes faster than humans could using their hands. Sure. A power drill can be a dangerous tool and you can hurt yourself with it. But the tool itself is not in any way harmful to the user.
Right now we dont even have an idea how one would produce a proper sentient artificial lifeform. The problem is so mindblowingly complex that we have trouble even defining it. I think the point im trying to make is that we tend to be very sloppy with the term "AI". we use it to describe things that are so fundamentally different from each other. The kind of Agent shown here will never - even given all the computing power of the world - be a threat to humanity the way SciFi portraits it.
Im not - by the way - saying there might not be future technologies that warrant this kind of care. In fact - by the point we get there (probably in a few decades) we should have an answer ready. But being worried today bc a program is able to play Starcraft is silly. Again - i think its a problem with using the term AI so loosely. What you are looking at is nothing more than a sophisticated bot. We have known bots in video games for a long time.
Actually it doesn't move the stalkers over multiple screen widths (or focus widths as DeepMind calls them). AlphaStar is programmed so that it can only input commands into it's current focus screen. It can see more of the map, but it cant click anywhere anytime, without moving the screen first.
I would totaly pay google and blizzard for having this ai for practice.
Amazing
Robin Spanier I think someone has to win first
First of all, thank you for your effort and success. Alphastar is truly impressive:
To learn the shown behaviour in the vast action&state space of starcraft 2 is definitely a success.
However, the ability to defeat human pros is a bit of a magic trick: The games are mainly won through mechanically super-human skills and (potentially) super-human decision making in the very short run (a few seconds). I don´t think alphastar is at all showing super-human decision making in the longer term (above a few seconds). This would be a so much stronger sign of intelligence than what we have seen. Hence, deepmind should take alphastar in its shown state as a mere starting block to improve on winning sc2 games through superior meta gaming.
A true proof of superiority in medium term and long term decision making would involve some or all of the following restrictions:
- a set of (slightly sub-)human pro level caps on APM and effective APM for small time periods (say 0.01, 0.1, 1.0, 3.0, 5.0 seconds)
- a minimum time delay between action of at least a couple of mil-seconds
- introduction of some spatial and time noise in clicking accuracy (to bring accuracy of actions to (sub-)human levels)
Probably there are other restrictions that force the agents away from exploiting super-human mechanical skills and instead focus on medium term and long term decision making.
With such restrictions in place I would be really excited if the LSTM based architecture would be successful in defeating human pro-players. It would be exciting because it would demonstrate that RNN architectures used in reinforcement learning are even capable of learning cause and effect relations which are very many time steps apart and which are obscured by an immense state&action space and incomplete information - starcraft is indeed a perfect example for this. Accordingly, I would highly appreciate if deepmind does not stop improving on their starcraft AI, but dive much deeper to take the really hard challenges of medium&long term planing.
As an aside: Forcing the agents to play at human mechanical levels or below would also be much more inspirational for the starcraft community as then the community can expect to see new meta game strategies.
beautiful line of thinking! I totally agree!
(this is the first video I am watching of AlphaStar and I have a feeling there's not much more)
My feeling is that they didn't know at the time how to how to train the AI "UP" (I mean "up" in the same way strategy is above tactics).
Even the decisions it does make right now, as per their commentary, are already learned from previous games.
Those mass stalkers? Learned from previous games and merely honed to a level where the human can't beat it, even with superior strategy, because they just don't operate as efficiently.
It does not take away anything from the brilliant effort by the Deep Mind team.
Definitely a platform for building the AI UP from here, to start asking and answering starcraft questions autonomously.
Going to play a new version of AlphaGo... lol 2:20:25
Saw that too... Imagine how funny it would be if they were then to proceed and sit Mana down in front of a goban Ahaha, and then Ke Jie with a keyboard. What the heck, let's have an AlphaGo vs AlphaStar game! :)
@@DaulphinKiller "AlphaGo vs AlphaStar game! :)" - in chess :D
LOL, I read this comment at the exact second when they said it on the video in the background :D
1 million views, makes me proud of the great game of starcraft! Kids these days should play more RTS!
This is like watching the terminator play SC2.
No one wants your 2 cents.
I can see it now, the new age of tournaments. MAN vs MACHINE! Place your bets on who will dominate in a classic game of wits. Will the human player's dedication and experience be enough to outclass this bot? Maybe he will able to discover abusable trends in the ai's gameplay? Or will the ai be able to calculate an efficient battle strategy to topple a stressed mortal? Who will win this game of attrition, economy, and skill as it unfolds today in MAN VS MACHINE! (SC II)
At 02:04:09 what stands out for me - Rotterdam mentions this as pulling workers really early - is that alpha star instantly pulls all workers, in the direction of his expansion and his stalkers. It seems as if it "intends" to move them in a way to use the "lost" time from pulling workers for a potential transfer to the expansion if the attack is going on. And at the same time drawing the oracle to the defenders. This looks like really efficient management of the situation. The earlier you pull here the earlier your defenders will reach the oracle and the fewer probes you lose, whereas for a later pull you would just manage to draw the oracle away from the defenders.