AlphaStar's Zerg vs Serral's Zerg, Game 5 AlphaStar vs Serral - Game 4: • AlphaStar vs Serral - ... Also check out more of my content: Twitch.tv/Artosis Artosis
It is not a single "agent" as they call it. I would guess each game was played by a different AI agent, each having different tendencies they build up over their training.
Uh, no? They didn't let AlphaStar use any mouse or keyboard. Can you imagine how hard it is to play StarCraft 2 just using the power of your mind to move units? That's a huge handicap and AlphaStar still wins.
@@michaelbuckers Imagine how fucken scary it would be if Alpha star played with two robot arms holding onto a mouse and keyboard. You'd look over to an empty chair and suddenly lock eyes with Nintendo's R.O.B
kek 1tap I been playing sc2 for 6 years, never knew there such a thing, until today...... game std all the way,I guess that’s different between noob and pro, lol=)
I've just stumbled on your analysis videos and watched them all in one go because I am generally interested in everything AI. Big thanks for explaining it in a noob-friendly manner, I don't even play Starcraft.
Great content, especially for people who don't have the slightest idea of what starcraft is.:) i'd suggest to not repeat the concepts that you've already tackled in one video to lessen the redundancy, but then again, it's very intuitive especially for new players.:) thanks for this serral game review
@@Singularitarian good to hear that.:) it's a great thing the video can reach out to spectator that aren't playing the game . hope you'll be able to enjoy and learn more of StarCraft :)
Dan I want to say I have played StarCraft since StarCraft one and have truly appreciated your casting and now you’re content the entire time thank you for sticking it out and your love for StarCraft
Another strength of alpha star which hasn’t been mentioned is that unlike a human player which can have a bad game, you’re always playing the best version of alpha star, it’s never having a bad game and it’s always playing at it’s best
This doen't mean anything. Terran is the hardest race to master without end units. And AI still has ways to go. One AI can beat humans with all the races then we can see which race is superior based on how aphastar performs versus itself
@@colemantanner9360 You realize how long it took to refine the three races into what they are right now? If you'd do a complete rework of terran you'd be able to play it in the year 2040 mate. Terran is great the way it is. All that's missing is a bit of very fine buffing and finetuning.
just what?! I'm actually in awe but in the back of my mind it feels like a meta change... hilariously the AI is showing what is now possible and what human players will do with it. i LOVE this game and I really like how the AI is approaching it. It makes US better. gogogo re
When Serral saw that it was an all in due to the queens marching across the map, is there a reason that he didn't build 2-3 spines to supplement his roaches? I think they'd make the defense a lot more tenable, but since Serral is infinity better than me, I'm curious if that would have been a clearly losing play in a way I don't see.
@@友-16 - Thanks for the simple explanation. Makes sense - I haven't actively played since WoL, so I always forget about the units added in the expansions.
Christopher Gallagher plus, the spines would have to be defending the natural, which would pretty much mean just giving up the third. And if that happens, then Alphastar can just contain and choke him out. That’s why Serral chose to fight between the natural and third initialy, because simply giving up the third would mean he just dies slower.
Two reasons. The first is you'd have to give up your third as well as your Evo and Roach Warren wall at your natural, and it would be very difficult for you to move out. This alone would give Alphastar a lot of damage and the attack would not longer be as "all in" The second reason is Ravagers are great against static defense anyways. They aren't great for their cost thanks to Corrosive Bile.
@@veidro no, alphastar was all in so even if serral sacrificed his 3rd he would still be ahead on economy. He would then be forced to take favorable engagements because he would be low on army. The only reason serral chose to make units over spines is because he is a pro gamer which means he would rather micro the few extra roaches. And in roach vs roach battles it's 90% numbers 10% upgrades.
And this is what AlphaStar was afraid in the previous game, where it played as terran. It also probably used low tech units cause it never made it so far with terran in a game against zerg.
Which is interesting because (especially with the latest balance pack) protoss is not as popular as terran or zerg due to the fact the higher unit investment doesn't compare to the efficiency of the other races. The ai managed to master protoss competatively
@@zvxcvxcz Efficiency having to do with being able to produce more units faster for less resources/supply. I am not saying protoss can't beat terran and zerg, only that they are popular because you can field an impressive amount of units as zerg very quickly. Even if they are weak and un-upgraded their function as scouts and sappers is undeniable.
I wonder how alpha star choses to go for a certain strategy like an all-in vs. a defensive play from the get go. Does it model the oponents mindset and make assumptions about what the oponent expects? Or is it just random? Or does it follow past experiences?
Its an AI, so essentially it takes what it's learned from previous matches. Pretty much it knows when it sees X it can expect Y, so Z is the strategy its had most success with in that scenario. Hard to explain, but that's how it thinks in a nutshell.
the ai played millions of games so it knows what happens in most of the scenarios but if it didnt it just make its best guess and the guesses it makes are very good, this AI first learned from pros replays then started playing vs itself and learned from that, its the best way to train an AI currently and everyone is doing it in complicated games like dota and sc2 less complicated games like chess and go that has a number of states (sc2 and dota have infinite numbers of states) can be learned by only self learning and doesnt require replays.
@@hamzanasab1713 It still learns from human replays via supervised learning. After learning from human replays it refines it's technique via reinforcement learning which is A* playing against itself.
Alphastar would've won the ZvT game had splitting been added to Alphastar's abilities. Fights would've looked much different if Alphastar didn't clump the siege tanks together, and didn't A-move the marines into banelings/ultras. I could see Alphastar beating Serral in that match with proper splitting and tank formation due to its macro.
You’re right, but keep in mind that AlphaStar doesn’t really have “abilities” that it can be intentionally taught. Everything you see it doing (macro, micro, harass, drops, all-ins) are emergent properties of a model that is defined at a more fundamental level of inputs and outputs.
Do you know if Serral prepared for AlphaStar at all? Did he watch replays and know it's strategies and tendencies, like he would for another top level opponent?
@Cédric Actually... Alphastar has been training on pro game replays (including serral's games) since it started. Alphastar would still be a worthless AI without the pro game training.
Its initial knowledge of the game was trained based on pro-level replays. However, since then, it has learned and evolved exclusively from self play. It wasn't made specifically to prefer strategies that counter Serral's play style for this match.
@@letao12 Incorrect. Everytime they pull it down and tweak it (which has happened multiple times now), they feed it pro gamer replays. I never said it was made to prefer strategies that counter Serral. However by watching his replays it made the exact correct call based on almost no information (nothing but knowing he made an 2nd base). Sorry but that is quite fishy.
There's a difference between "learned from pro games" and "prepared for match against Serral". Everyone else's talking about the latter here. Also, source for your claim that "everytime they pull it down and tweak it"?
@@letao12 Look im not gonna go searching for it when you can google it on your own. But yes they feed the AI pro games between versions, they have said as much a few times.
It amazes me really that it's just SO bad at terran compared to the other races. But as others have pointed out, it played millions of games and just "chose" those few iterations of itself that had the best chance of survival.
Yeah, it is sorta random (but not really). So it happened to be terran that sucks. Note: there are thousands other reasons why this might be. (eg. terrain has some strats that require more exact timing or terrain might be harder to learn, or many other reasons)
It's also possible that the AlphaStar Zerg bots are weak against Terran because they haven't learned to properly use Vipers and Infestors. This would also end up making the Terrain bots weak against human Zerg players.
Because he wasn't sure how all in the upcoming attack was going to be. He did cancel his Lair at 6:57 when he saw roaches moving out. Since roaches have a build time of 19 seconds that gives him about enough time to have them ready when the attack hits 30 seconds later.
Someone explained it above, it would mean to give up the third, and Alphastar would be able to contain Serral and choke him out. Basically it's just the slower death. So it's better to fight immediatly.
I mean if you havent noticed this already then you're probably a lil blind but in 2019, half the shit Alphastar was doing was uncommon/weird/not what a pro would do. Go play sc2 in 2022, and every single opening build is a direct copy of alphastar and now the Meta, and has been the meta for the years since 2019.. Basically the pro's watched every single replay, studied them, copied them, turned them into the meta. Only problem is since 2019 we havent seen alphastar... I wonder why? Because if they release it again, the meta will change for the next 5 years and everything will just be copied again.
I have not played or watched in many years so this is gonna be a noob question, but why didn't Serral defend that attack with sunken colonies? Wouldn't they have helped? Seems like he could have turtled a little bit until he had mutas and cleaned up the roaches and then used them to harras the hell out of alpha star. Obviously that didn't happen so I'm guessing I am misremembering a lot of things.
since ur calling them sunken colonies i assume ur a BW player *high five*. the ravagers, the unit morphed from a roach, has an ability that is long range and does a lot of dmg that is on cooldown that wrecks static defense.
*"Why does AlphaStar cancel building?"* This question is coming from someone who doesn't understand deep learning. Even after hundreds of years of playing against oneself, in some _human VS_ situations, DeepMind will know not to build the structure, but the only record/memory of doing so is to build it and then cancel 20 seconds later (because it's one of billions of possible timelines it has stored.) Remember, DeepMind can think of billions of possibilities for gameplay; in the meantime it can't calculate the _execution_ of each of these possibilities to perfect accuracy. What makes it the best player is mostly (a) it's comprehensiveness and then only (b) it's precision of KEY encounters, not its precision of building. Comprehensiveness means 'outside the box' thinking, not precision. The only time is will have perfect precision is when playing against itself.
TBF, real starcraft strategies can have similar forks. For example, in the eras I played, ZvZ was a very sharp game. One of the most critical strategic skills was to recognize your opponent is going a lot harder on economy than you, at which point you need to immediately switch gears to all-in attack to damage your opponent before your window of opportunity rapidly closes... but in that window of opportunity you can be extremely effective. Otherwise, you are at a major disadvantage and can only really hope to catch your opponent in a mistake.
Alphastar is ready for tournament play. I want to see Alphastar compete in real tournaments like WCS, Nation Wars, Blizzcon, etc. AI should be allowed to compete in tournaments.
@@Th3Sh1n1gam1 I don't think maps would be such a huge deal, it'd matter for TvX more, with siege tank and lib positioning, seeing how he did defeat a top tier player 4-1 And it gets better in days, not weeks, like humans. My capacity to advance through sc2 ladder is very slow. This beast doesn't get tilted, doesn't get ladder anxiety. It just keeps getting better every hour of every day.
Not really, while A* is very impressive in a straight up slugging match, Alphastar has been shown to totally suck against cloaked units (at least the protoss iteration). I suspect it'd get shredded by pros that takes advantage of that.
and when you are at it also let an motorcycle compete in tour de france. Look I like these matches and it could be a nice sideshow but it should not be allowed to compete in the actuall tournament. Do the same with chess and go and the humans may as well not bother.
Hmm the micro is pretty sick, would not be able to do it, maybe pros can quickly select individual units and pull back and morph to ravagers. The PC will always be better at micro.
AS does not have the same abilities it had in it's matches vs MaNa and TLO. AS's APM/APS, and reaction time in these agents were capped to equal GM level players.
@@Singularitarian Matching average and max APM doesn't make the micro equal. For instance a computer can essentially freeze time, assess the whole screen and issue the perfect order on the ideal unit/s. A human at max APM could never have the same precision and will lose concentration over time.
Wow! I did not expect this outcome. I wonder what Alphastars APM is though? In the matches against TLO I think they put a limit on it, but it was still able to output extreme APM in short bursts. Also, for an AI APM=EPM, which is the relevant number for human players.
Alpha star is limited in every match that it plays. They didn't want alpha star to be overpowering it's opponents just with more APM or Screens per minute. Also it's reaction time is pretty slow compared to a pro player. It's just very very accurate with everything it does and it makes it's limited actions count.
They fixed the APM thing like ages ago. How have so many people following this not heard of this yet. After the exhibition in Jan, they brought TLO into the Alphastar team, and they also consulted Mana and Kelazhur for restrictions that will make Alphastar comparable to humans.
AS does not have the same abilities it had in it's matches vs MaNa and TLO. AS's APM/APS, and reaction time in these agents were capped to equal GM level players. DeepMind also did the same to the agents that played on the ladder.
*Agents were capped at a max of 22 agent actions per 5 seconds, where one agent action corresponds to a selection, an ability and a target unit or point, which counts as up to 3 actions towards the in-game APM counter. Moving the camera also counts as an agent action, despite not being counted towards APM."
@@fupopanda They never actually fixed the APM (that I've heard of)... they set a pretty high APM limit competitive with pro player APM while ignoring that even for pros their effective APM is much lower and that the additional spammy extra right clicks to move aren't really comparable to a precision selected stalker blink in actual execution time. TLO and Mana almost look like paid shills if they brought them on board and then used them for the exhibition matches... if what you're saying is true then that's really bad.
@@elleshar666 so... 1v1 matches are 'faster' while people play it? ... it kind of makes sense, but really _is weird._ :D -- " Normal speed for 1v1 matches is not normal, but faster. " .......... right...... XD I guess if I sad "Live speed for 1v1 matches is not normal, but faster. " it almost makes sense...
@@Verrisin When you start playing the game campaign for the first time you have an option to choose a difficulty. The "normal" difficulty is played on normal speed. Also if you play 1v1 and start from the lowest of the leauges the game speed is normal there too (unless it's changed later on) . So if you get a sense of the game and choose a harder campaign difficulty or move up the 1v1 ladder then game gets faster. Then the faster speed is what you will encounter all the time hence making it the normal game speed from your perspective. edit: Yes 1v1 matches are played on "faster" speed. you can't change it unless you play with a friend or form a custom party. and if you wonder, you can't play the game at 4x speed.
@@Verrisin sc2 is not a moba it's a RTS :) the campaigns well worth the price. very fun and not difficult to get into . if you want to try 1v1 or team games it's free to play.
I support Alphastar and I think it would become the best without doubt, BUT... If Deepmind someday achieve this, I wish it would be without extra advantages like the ones it has today. It should do it with a camera watching a monitor as visual input and one mechanic hand with 5 fingers using a keyboard and another interface to control a mouse. In that way, any screen or control cheats are not possible, like: 1- selecting single units without control groups under the view of other flying units. 2- not wasting time to enclose a group of units with the mouse. 3- not need for control groups or waste actions by minute in no trivial stuffs. 4- winning most of the matches with all races in all maps. If all that is fulfilled, then we can said that Alphastar strategies and decision making are better than human level. To achieve the point I am asking, it needs to use another neural networks to command the keyboard and mouse in base to what alphastar needs, about the camera, Alphastar already has image recognition, but we cannot know if it cheats somehow.
Alright, I'll address each point one by one. But firstly I'd recommend you read the paper published by deapmind about 2 days ago about the starcraft AI, it covers almost everything you need to know. So for the robotic hand and camera, there is a field of computer science called computer vision wherein one can take an image from a camera and translate it to a simplified view of unit positions. This typically uses AI but in fact does not require it in simple situations like starcraft units, which have a set icon. Simply scan the image for each icon in turn and hilight that section of the screen with that unit type. This can easily be done in real time with a 100% accuracy for unit positions given the simplicity of starcraft rendering. A similar thing can be done for the hand. For this absolutly no AI in needed, only classical control theory. Now if implimented both of these would have absolutly no impact on the AI, as it would still have the same info about unit positions and the same output to the computer (note that command latency was implimented for the AI to simulate button press times, in fact a real-world robot would have less latency as robots are significantly faster then humans). The way this was implimented mearly saves the work of implimenting these other systems that would take a lot of time for absolutly no payoff, as it mearly makes things identical to how they are now. Also it would require a large robot instead of a flashdrive, and would not be able to play against multiple oponents at once. Now for the rest of your points about unit selection, the AI actually selects units the same way a human does in all situations. Command delay is incorperated into this so the whole "mouse drag time" this is not an issue. Overall the AI has no advantage in unit selection. The AI also has no advantage in terms of command rate (even during spikes) and in fact has a slightly lower command rate then human plays at a similar level at all times. The paper mentioned a 99.5 or higher win rate in all possible matchups between races on all maps. Note this is against any players, not just pros, but its still better then most professionals. Overall, what the AI is right now meets all your criteria for a better then human AI. This is with the exception of completly ussless addons that would merly take engineers away from the important points of AI research and have no effect on gameplay.
I shoud also mention that the AI is mearly giving commands to the actual game, it does not command the game for things like resources or unit health. In the sence it is impossible for it to cheat. There was also a previous itteration of this AI that had no restrictions on command rate or vision (saw the whole map instead of a regular view) and it absolutly demolished all grandmasters. It did have the restriction of only playing protos V protos on one map, and heavily overused micro to win where it probably should have lost. But thats the advantage of AI!
@@whyOhWhyohwhy237 *Simply scan the image for each icon in turn and hilight that section of the screen with that unit type. This can easily be done in real time with a 100% accuracy for unit positions given the simplicity of starcraft rendering* not sure if that methods allows to identify a soap of units (many overlords and mutalisk), invisible units, units dying, explosions, special effects, etc. I would bet that this is not the method used, I imagine a neural network is used for map and unit feature recognition, besides, without a neural net approach, any illusion or trick exploited by humans to hide units or nuclear launches may not work vs a brute force recognition if does not learn with patterns and neurons. *A similar thing can be done for the hand. For this absolutly no AI in needed, only classical control theory* They are studying learning machine and you propose the use of math and normal algorithms? A neuralnet can make mistakes (the same we do), a neuralnet should learn the keyboard reaction time, it should repeat an action in case this one did not execute, etc. We are talking about a tiny neuralnet by the way, no more than 50 neurons for that task. With this the AI would have a playstyle or behavior close to humans. *The way this was implimented mearly saves the work of implimenting these other systems that would take a lot of time for absolutly no payoff* The way is currently implemented looks completely different to human control behavior. First fact! The AI is amazing in micro! I saw her fight hand to hand with 30% less income over 15 minutes using warp prism super micro. Players spam APM, giving a false APM which AlphaStar kinda match. Players also lost time selecting individual units one far of the other due mouse moving time (that is why nobody can match that stalker blink or warp prism micro! That is why all agents choose warp prism and stalkers as their main unit. *Also it would require a large robot instead of a flashdrive, and would not be able to play against multiple oponents at once.* You really think they use the image recognition neural net to play between agents? Noo.. in that case they use unit position code, not image, not nothing. That is why they made an special sc2 version, to train and play thousands of games by hour without the need to render graphics, about the keyboard and mouse could be simulated in virtual environment, it would not consume much processor, Or it can be approximated and learned with the latest agents, the most hard thing to do is to train the neural net for clever decision making, control and image recognition are just neuralnets interfaces that can be trained independent in some cases. *the AI actually selects units the same way a human does in all situations* No, it does not. The paper tells the difference and you can see it in the replays played in battlenet. Few delays setted does not equal to human speed. I saw the IA spotting (due minimap) when a single zerling jumped from the fog zone to the viewed pixel, and 100 ms later the two oracle from the other side of the map where recall, a pro player would have take 2 seconds or more to react. Another real disadvantage for real players is eyes focus, we can see all the screen but we only see clear in the center of our sight focus. Sometimes the AI select an unit in the edge of the screen when only few pixels can been seen. *The paper mentioned a 99.5 or higher win rate in all possible matchups between races on all maps* Just in 4 maps, Serral may have 100% win rate vs master or below, and around more than 90% rate in grandmaster. Zerg Agent may be in the worst places from 200 of grandmaster, they only play mostly with zerlings, roaches and ravagers. They dont even know how to predict shoots with the ravager. *that would merly take engineers away from the important points of AI research and have no effect on gameplay.* Proving that a computer is faster than US does not prove nothing, they need to prove that an IA could make better choices than US, for that they need to play as I said to match a more natural play style, something excited to see and erase any other cheat they may use. Because the true is: that I don't really trust in all the words they said. Many of them could have their job at risk if they dont accomplish results, and already 2 years or more had past since they started. Alphastar also needs to learn to said GG when a game is lost, the chess version of deepmind already measures winning probabilities according to position and pieces, this is done with another neural net.
@@AngelLestat2 Neural networks are not a magical fix-all. They are good at tasks where we do not know even the first step of programing something to accomplish a task, but they are quite bad at tasks with known solutions. Image recognition can be done by taking unit images (invisible images, dying images, everything) and scanning the image input with a "search square". Take the RGB values from this search square, put them into 3 matrices and find the dot product between those matrices and the desired unit images. the resulting number is a similarity score which can be compared to a pre-set threshold to create a boolean value for whether a unit is there or not. Its one of the first things you learn in a computer vision class. They are studying neural networks for how they make decisions in star craft. This does not mean they have to use neural nets for every mundane calculation they ever need to do. Robots can be made VERY fast with classical control theory, in fact using the root-locust technique we can optimally tune PID controllers for the fastest possible controller to get to a desired position, such as hitting a key. This is done FAR faster then humans. Robots have been faster then humans for years ( th-cam.com/video/-KxjVlaLBmk/w-d-xo.html ). If you have an actual position and a desired position, as is the case with a typing robot (or using a mouse), you should absolutely never use AI. AI are best used as high level controllers, low level controllers should not use them for best results are these are solved fields. We know almost everything about the robot so AI is not necessary in the slightest and in fact always yields worse solutions and the know solutions are optimal. The previous version of the AI was godly at micro, and demolished everyone because of it. This version has significant limitations on the APM. In fact the pro players are quoted as saying its micro is on par with humans and does not feel massively better. Again, there are videos of AlphaStar playing with great micro, but those are previous versions. "You really think they use the image recognition neural net to play between agents???" That's literally the exact opposite of what I'm saying. I'm saying that a robot would be impractical as a counterargument to you saying they should use a robot. I'm the one advocating for using a purely simulated AI. I don't actually play StarCraft that much (I'm more interested in the paper as an AI researcher then a gamer), so your probably right about the reaction/focusing time. they might have to account for that more in subsequent iterations. You're right in saying that they are trying to measure the high level play of the starcraft AI more so then the low level microing. However I believe that these results are incredibly significant in terms of overall strategy. Again, I don't actually play the game that much so I have to go on the word of professions, but they say its not superhuman in terms of micro and I tend to trust them, thus any wins are more so due to strategy then speed. I absolutely agree that any superhuman micro needs to be removed to properly test the strategy of the AI, but things like using robots and image recognition are not the best way to achieve this. However I don't think a "natural play-style" is the metric we should use to determine success. AI are notorious for crazy counter-intuitive ideas that somehow work. You can see them in chess and GO AI. I would advise you to trust researchers, peer review is a hell of a process and replication is key. Alphastar will never "learn" to say GG. Its goal is to win, and even if you have only a 1% chance of winning saying GG lower that to 0% and is thus never a good option. If given the option the AI will literally never surrender, as their only goal is to win. I believe Alphastar already has a win probability calculation network, but if the researchers were to program it to say GG after a certain point it would be just that, programing. Right now the researchers are just manually telling it to say GG when they know the game is lost, which is fine for now.
Why is Serral off creep... You're down units and tech... you could easily work that choke with the buildings and buy time... I feel like the 4 games I've watched with Alphastar were thrown. Serral could have taken some games.
It can be tempting to anthropomorphize, especially as the AI becomes increasingly complex, but it's not a person. AlphaStar doesn't 'want' to do anything. It is responding to stimuli. I would also be careful with what you call a 'mistake.' Some moves may appear to be a mistake on their face, but on reflection there may be a logical explanation. I understand you may be taking creative license with some of the phrasing. I just find it a bit jarring.
@@Sid11358 Fair enough still there is a point where it gets really hard to tell if something is conscious or not. I mean what about chimps, dogs, snakes, ants? At what point is it conscious and at what point is it just responding to stimuly?
@@petrkinkal1509 I agree there is no clear delineation. It may be that Deepmind becomes conscious (in its own fashion - analogous to how dolphins or rats or crows are conscious) in the near future and we don't notice or misinterpret its consciousness as formulaic responses. On the other hand it may never become conscious, but we misinterpret its responses for consciousness, though this difference may have little practical value (if it appears conscious it may as well be). The crossover point will be difficult to observe.
@@Sid11358 "difference may have little practical value" This I agree with this so much. But that is a big part of my point it doesn't matter that much if it wants to win or not if will still do best it can to win so it effectively acts like if it wants to win. The anthropomorphising would start to matter if people started doing something like for example having it also play few matches with lover tier players so it doesn't feel bad for always loosing (or expecting it to feel bad about a player it is constantly trashing) that would be pretty ridiculous.
@@StevenCasteelYT What are you trying to say? I am sure that there are far more people who is interested in Deepmind but never even heard of starcraft than the number of sc2 players.
Steven Casteel wake up Steven, the “community” of SC players you’re reffering to is only getting smaller and smaller, and this is a great way to bring new people in and make them understand what the game is even about. All of the “newbie friendly” content is outdated, everything is absolutely different than it used to be even like 2 years ago and everything is constantly changing. I’ve been with the game for 20 years now and absolutely love the idea of making content for people that simply haven’t been around for this long. Nothing wrong with this, and videos of the match in real-time are avaliable as well. So stop hiding behind the “community” just because you don’t feel like watching something that is not targeted at you.
Artosis is great but I really think pausing the game every 3 seconds is a terrible idea. I know these videos are for people who never played Starcraft but people just can't get how fast both players have to think to plan and execute all the actions when you slow down the games like this.
but that shows, if game would be slowdown 3 to 5 times, then humans would win with severe difference, alphastar plays like big subconscious, no logical thinking there at all(only reflexes and deep interpolation between known outcomes/actions), people just can't do much logical thinking on those speeds, but if they would have such time...
sure, for players that want to learn to play on their own. but this is analysis/commentary. you can get live casting by watching pro games live. also, please use speed up/down and just skip ahead features. it's 2019, i think most people shouldn't only be limited to watching things at 1x anymore, especially if you're already familiar with the basic content. at least that's how i watch talks and stuff now. i get it if you're on your phone though, but on the computer, it's so much time saved.
By that logic DeepBlue ruined chess in 1996. And recently backgammon. Except that it didn't, and humans still do chess tournaments. And those are still fun. Just because there is an entity, like an AI or a strong player somewhere out there better than you, doesn't mean you have to stop playing, or stop enjoying a game. Whether or not these advanced machine learning techniques will be used for the betterment of all humanity or just a few wealthy individuals is the real issue here.
Anybody who do not play the game cannot keep up at all with the regular casts. These breakdowns are brilliant, it separates from the other replay channels and very insightful.
Might be a spoiler but it seems an Ai can beat a human with everything but Terran. Moral of the stroy humans are bad. Force A.I. to play Terran and it's all fine, never gonna win. Perfectly balanced for both humans and ai alike, as all things should be. xD xD xD
I think the reason it lost with terran was because of the first four lings that messed with Alphastars build. In every game it controlled the "flow" of the game except in the terran game were Serral did whatever he wanted.
@@Zettern96 Might be, but the reason it happened this way might as well be because Terran is in a shithole and even an AI can't muster a win with it, when it managed to do it quite comfortably with the other 2 races versus the same opponent.
@@Figgy20000 Oh yeah, i did imagine these 5 games against serral were how the AI rebellion defeats our best generals. But in reality, the AI rebellion would look a lot more like Automaton 2000 playing with it's 300'000 apm
Just for the record. Serral isn't even a champion in any of the recent Premier tournament. Far from "best player in the world". It's good that they choose Serral to feed AlphaStar, minor players before the actual players with actual resullts.
cmon Artosis. Starcraft players are fastest and smartest gamers in the entire world! And you pause it like every 30 seconds to explain extrimely simple stuff.
Pavel Averyanov I wrote it in a different response, but am gonna do it again. Stop whining just because you don’t feel like watching something that is not targeted at you, simple as that.
I absolutely hate hate HATE this commentary style... I know how a ravenger works and that it evolved from a roach. This is was so boring it made me angry. Watch this video to help you sleep or if you have never played StarCraft before and are curious.
Yes, seriously. You can make a SC2 tutorial OR make an Alphastar replay. But don't do both at the same time. Make 2 separate videos. One for newbs and one for pros. This is like watching Deep Blue play chess and then the commentator pauses the game every 2 seconds to explain the basics of chess. We aren't cavemen. Is there really people who like Alphastar but don't even know how to play SC2? Sigh...
@@weestro7 Well the whole point of the comment section is for people to voice their opinion, which helps give feedback to the creator to make more entertaining content so that they can make more money. So criticism is of high value when it is constructive.
Alphastar sees the 3rd base and is immediately like "OH NO BRO NOT AGAINST ME!!"
build eco is the ultimate cheese lol
wrgg
After seeing how defensive alpha star plays in the first games I would’ve never guessed it was going to do a roach all in. Amazing series
It is not a single "agent" as they call it. I would guess each game was played by a different AI agent, each having different tendencies they build up over their training.
+Chris McGarrah Alpha ALWAYS roach ravager queen all ins with zerg...
Eye of Horus And how would I know that if this is the only set I’ve seen of Alpha. Try reading what I wrote
@@goodguycwyzz4768 I'm just telling you... I know exactly what you wrote
@@eyeofhorus1301 then why did you add the ellipsis as if what you said was obvious? rofl
These breakdowns are brilliant Artosis, separates you from the other replay channels and very insightful. Really made me appreciate the game. 👍👍👍👍👍
sarcasm right
Thanks for taking the time to explain what's happening. It helps me better appreciate the technology that the Deepmind team is creating!
I like the pausing and the deep analysis, rather than just superficially watching the match at regular speed.
by deep analysis you mean starcraft 101 right?
Why did they allow AlphaStar to bring it's own mouse and keyboard if Serral didn't have his ?
Uh, no? They didn't let AlphaStar use any mouse or keyboard. Can you imagine how hard it is to play StarCraft 2 just using the power of your mind to move units? That's a huge handicap and AlphaStar still wins.
@@michaelbuckers without keybinds too. Mad
@@michaelbuckers Imagine how fucken scary it would be if Alpha star played with two robot arms holding onto a mouse and keyboard. You'd look over to an empty chair and suddenly lock eyes with Nintendo's R.O.B
kek 1tap I been playing sc2 for 6 years, never knew there such a thing, until today...... game std all the way,I guess that’s different between noob and pro, lol=)
@@miner4236 Source?
Artosis I'd love to see more commentary like these on normal high level games.
Check out his In Depth series with NoRegret. It's similar: deep and well thought out
I've just stumbled on your analysis videos and watched them all in one go because I am generally interested in everything AI. Big thanks for explaining it in a noob-friendly manner, I don't even play Starcraft.
Same
Great content, especially for people who don't have the slightest idea of what starcraft is.:) i'd suggest to not repeat the concepts that you've already tackled in one video to lessen the redundancy, but then again, it's very intuitive especially for new players.:) thanks for this serral game review
I thought the bit of redundancy seemed helpful for non-starcraft players. Can't absorb all these details by watching one game.
@@Singularitarian good to hear that.:) it's a great thing the video can reach out to spectator that aren't playing the game . hope you'll be able to enjoy and learn more of StarCraft
:)
I just want to say that I really like this casting style, especially the pauses.
I prefer without pauses tbh
Irony?
@@diegomieresherrera5920 No Irony, comments can be made while game on
@@emiliomagallon9490 i mean it for who made the main comment
It's clearly intended for all the people interested in AI who've never played.
so the moral of the story is, even AlphaStar can't win games as terran where he dominates with other races.
Yeah, I dont recall seeing serral lose a late game to terran
Spoken like a true Terran
I like how you made this comment on z v z.
Maybe they now can make a balanced game making use of alphastar to test balance.
AlphaStar also played basically Marines only and drop only. And threw away a lot of units.
New meta: when you scout the third, cancel everything and roach all-in
The change from the old meta is waiting until you see a third instead of doing it from the start I guess...
i havent played since WoL and i come back regulary just to hear your casting. best ever PERIOD.
4:00 what AS did there was to have queens stop anything from going inside and seeing the roach warren early on, that's very smart.
Dan I want to say I have played StarCraft since StarCraft one and have truly appreciated your casting and now you’re content the entire time thank you for sticking it out and your love for StarCraft
Another strength of alpha star which hasn’t been mentioned is that unlike a human player which can have a bad game, you’re always playing the best version of alpha star, it’s never having a bad game and it’s always playing at it’s best
Watching these five games was a really wild ride. Thanks for taking us along and explaining things!
Terran confirmed hardest race for AI to master and thus hardest race in game. Buff terran plz blizz
zynius not a buff but a total rework of the race
This doen't mean anything. Terran is the hardest race to master without end units. And AI still has ways to go. One AI can beat humans with all the races then we can see which race is superior based on how aphastar performs versus itself
@@colemantanner9360 You realize how long it took to refine the three races into what they are right now? If you'd do a complete rework of terran you'd be able to play it in the year 2040 mate. Terran is great the way it is. All that's missing is a bit of very fine buffing and finetuning.
if Alphastar had a few ghosts vipers are almost useless
just what?! I'm actually in awe but in the back of my mind it feels like a meta change... hilariously the AI is showing what is now possible and what human players will do with it. i LOVE this game and I really like how the AI is approaching it. It makes US better. gogogo re
When Serral saw that it was an all in due to the queens marching across the map, is there a reason that he didn't build 2-3 spines to supplement his roaches? I think they'd make the defense a lot more tenable, but since Serral is infinity better than me, I'm curious if that would have been a clearly losing play in a way I don't see.
It's the presence of ravagers. Corrosive bile are great against static defense.
@@友-16 - Thanks for the simple explanation. Makes sense - I haven't actively played since WoL, so I always forget about the units added in the expansions.
Christopher Gallagher plus, the spines would have to be defending the natural, which would pretty much mean just giving up the third. And if that happens, then Alphastar can just contain and choke him out. That’s why Serral chose to fight between the natural and third initialy, because simply giving up the third would mean he just dies slower.
Two reasons.
The first is you'd have to give up your third as well as your Evo and Roach Warren wall at your natural, and it would be very difficult for you to move out. This alone would give Alphastar a lot of damage and the attack would not longer be as "all in"
The second reason is Ravagers are great against static defense anyways. They aren't great for their cost thanks to Corrosive Bile.
@@veidro no, alphastar was all in so even if serral sacrificed his 3rd he would still be ahead on economy. He would then be forced to take favorable engagements because he would be low on army.
The only reason serral chose to make units over spines is because he is a pro gamer which means he would rather micro the few extra roaches.
And in roach vs roach battles it's 90% numbers 10% upgrades.
Someone please teach AlphaStar the battle cruiser rush and proper BC micro with repair... it will then have a winning terran AI as well
Thanks for the in-depth newb friendly commentary.
And this is what AlphaStar was afraid in the previous game, where it played as terran. It also probably used low tech units cause it never made it so far with terran in a game against zerg.
AMAZING COMMENTATING ON ALL 5 GAMES. THANKS.
AlphaStar rushed this silver player and showed how to play zerg
Wonder if pulling the boys earlier would have made a difference vs that all in attack.
AlphaStar real savage😂 Was not at all expecting that 2 base all in
If you're looking for more commentary like this, then check out the Day9 daily archive.
So all Serral had to do was to stall right? The economy advantage was about to kick in in seconds right?
So did Deepmind do it ?
Is AlphaStar finally able to win a GSL ?
Now I feel alright with my playstyle considering Serral vs AS - ZvZ plays almost exactly like my ZvZ games vs Elite AI XD
Aah, this reminds me of super old school ZvZ where powering too hard gets you overrun.
What we can conclude is, Alphastar has achieved superhuman level with Protoss at its final iteration.
Which is interesting because (especially with the latest balance pack) protoss is not as popular as terran or zerg due to the fact the higher unit investment doesn't compare to the efficiency of the other races. The ai managed to master protoss competatively
@@Enkarashaddam I don't think you know what efficiency means given the results.
@@zvxcvxcz Efficiency having to do with being able to produce more units faster for less resources/supply. I am not saying protoss can't beat terran and zerg, only that they are popular because you can field an impressive amount of units as zerg very quickly. Even if they are weak and un-upgraded their function as scouts and sappers is undeniable.
I wonder how alpha star choses to go for a certain strategy like an all-in vs. a defensive play from the get go. Does it model the oponents mindset and make assumptions about what the oponent expects? Or is it just random? Or does it follow past experiences?
Its an AI, so essentially it takes what it's learned from previous matches. Pretty much it knows when it sees X it can expect Y, so Z is the strategy its had most success with in that scenario. Hard to explain, but that's how it thinks in a nutshell.
the ai played millions of games so it knows what happens in most of the scenarios but if it didnt it just make its best guess and the guesses it makes are very good, this AI first learned from pros replays then started playing vs itself and learned from that, its the best way to train an AI currently and everyone is doing it in complicated games like dota and sc2
less complicated games like chess and go that has a number of states (sc2 and dota have infinite numbers of states) can be learned by only self learning and doesnt require replays.
@@hamzanasab1713 It still learns from human replays via supervised learning. After learning from human replays it refines it's technique via reinforcement learning which is A* playing against itself.
@@kylemcconnell378 actually alphastar doesnt learn from playing against humans. it trains itself in games against itself.
Alphastar would've won the ZvT game had splitting been added to Alphastar's abilities. Fights would've looked much different if Alphastar didn't clump the siege tanks together, and didn't A-move the marines into banelings/ultras. I could see Alphastar beating Serral in that match with proper splitting and tank formation due to its macro.
You’re right, but keep in mind that AlphaStar doesn’t really have “abilities” that it can be intentionally taught. Everything you see it doing (macro, micro, harass, drops, all-ins) are emergent properties of a model that is defined at a more fundamental level of inputs and outputs.
It's official - all in zerg is the best strategy
Do you know if Serral prepared for AlphaStar at all? Did he watch replays and know it's strategies and tendencies, like he would for another top level opponent?
@Cédric Actually... Alphastar has been training on pro game replays (including serral's games) since it started. Alphastar would still be a worthless AI without the pro game training.
Its initial knowledge of the game was trained based on pro-level replays. However, since then, it has learned and evolved exclusively from self play. It wasn't made specifically to prefer strategies that counter Serral's play style for this match.
@@letao12 Incorrect. Everytime they pull it down and tweak it (which has happened multiple times now), they feed it pro gamer replays.
I never said it was made to prefer strategies that counter Serral. However by watching his replays it made the exact correct call based on almost no information (nothing but knowing he made an 2nd base). Sorry but that is quite fishy.
There's a difference between "learned from pro games" and "prepared for match against Serral". Everyone else's talking about the latter here.
Also, source for your claim that "everytime they pull it down and tweak it"?
@@letao12 Look im not gonna go searching for it when you can google it on your own. But yes they feed the AI pro games between versions, they have said as much a few times.
serral is a bot
alphastar is the new version of serral without physical body
While Serral is def the best Zerg in the standard game id much rather see how Rouge does it vs Alphastar especially in ZvZ
It amazes me really that it's just SO bad at terran compared to the other races. But as others have pointed out, it played millions of games and just "chose" those few iterations of itself that had the best chance of survival.
Yeah, it is sorta random (but not really). So it happened to be terran that sucks.
Note: there are thousands other reasons why this might be.
(eg. terrain has some strats that require more exact timing or terrain might be harder to learn, or many other reasons)
It's also possible that the AlphaStar Zerg bots are weak against Terran because they haven't learned to properly use Vipers and Infestors. This would also end up making the Terrain bots weak against human Zerg players.
@@ruanpingshan right it’s agents train against other agents. It’s amazing what it could get done with three of the games most basic units.
Someone else pointed out that AlphaStar was maybe afraid of a 2 base all in just like it did in this game.
Terran needs a buff. Not even A.S can win with it. :) Protos and Zerg is however winning armies under its control.
How can one man have this much sheer KNOWLEDGE?
can anyone explain why Serral chose Lair after recognising roach rush?
Because he wasn't sure how all in the upcoming attack was going to be. He did cancel his Lair at 6:57 when he saw roaches moving out. Since roaches have a build time of 19 seconds that gives him about enough time to have them ready when the attack hits 30 seconds later.
@@moritzsinghartinger5376 You right, I didnt see the lair cancelled. Well spotted.
Why didn't Serral start some spines?
Ravagers bile kill them quick
Someone explained it above, it would mean to give up the third, and Alphastar would be able to contain Serral and choke him out. Basically it's just the slower death. So it's better to fight immediatly.
Because you don't throw pinecones at an armada.
The video is fantastic. If there could be less pauses, and with a bit faster-pace explanation. It would be even better
I mean if you havent noticed this already then you're probably a lil blind but in 2019, half the shit Alphastar was doing was uncommon/weird/not what a pro would do. Go play sc2 in 2022, and every single opening build is a direct copy of alphastar and now the Meta, and has been the meta for the years since 2019.. Basically the pro's watched every single replay, studied them, copied them, turned them into the meta. Only problem is since 2019 we havent seen alphastar... I wonder why? Because if they release it again, the meta will change for the next 5 years and everything will just be copied again.
I have not played or watched in many years so this is gonna be a noob question, but why didn't Serral defend that attack with sunken colonies? Wouldn't they have helped? Seems like he could have turtled a little bit until he had mutas and cleaned up the roaches and then used them to harras the hell out of alpha star. Obviously that didn't happen so I'm guessing I am misremembering a lot of things.
since ur calling them sunken colonies i assume ur a BW player *high five*. the ravagers, the unit morphed from a roach, has an ability that is long range and does a lot of dmg that is on cooldown that wrecks static defense.
shame game is so dependent on actions per minute. wish you could apply serral to a civilizations or other 4x game..
*"Why does AlphaStar cancel building?"* This question is coming from someone who doesn't understand deep learning. Even after hundreds of years of playing against oneself, in some _human VS_ situations, DeepMind will know not to build the structure, but the only record/memory of doing so is to build it and then cancel 20 seconds later (because it's one of billions of possible timelines it has stored.) Remember, DeepMind can think of billions of possibilities for gameplay; in the meantime it can't calculate the _execution_ of each of these possibilities to perfect accuracy. What makes it the best player is mostly (a) it's comprehensiveness and then only (b) it's precision of KEY encounters, not its precision of building. Comprehensiveness means 'outside the box' thinking, not precision. The only time is will have perfect precision is when playing against itself.
TBF, real starcraft strategies can have similar forks. For example, in the eras I played, ZvZ was a very sharp game. One of the most critical strategic skills was to recognize your opponent is going a lot harder on economy than you, at which point you need to immediately switch gears to all-in attack to damage your opponent before your window of opportunity rapidly closes... but in that window of opportunity you can be extremely effective. Otherwise, you are at a major disadvantage and can only really hope to catch your opponent in a mistake.
Alphastar is ready for tournament play. I want to see Alphastar compete in real tournaments like WCS, Nation Wars, Blizzcon, etc. AI should be allowed to compete in tournaments.
And mb give it the right to vote? It's much better to stay in the entertainment/science business.
The main issue is that it has to relearn based on map pool.
@@Th3Sh1n1gam1 I don't think maps would be such a huge deal, it'd matter for TvX more, with siege tank and lib positioning, seeing how he did defeat a top tier player 4-1 And it gets better in days, not weeks, like humans. My capacity to advance through sc2 ladder is very slow. This beast doesn't get tilted, doesn't get ladder anxiety. It just keeps getting better every hour of every day.
Not really, while A* is very impressive in a straight up slugging match, Alphastar has been shown to totally suck against cloaked units (at least the protoss iteration). I suspect it'd get shredded by pros that takes advantage of that.
and when you are at it also let an motorcycle compete in tour de france.
Look I like these matches and it could be a nice sideshow but it should not be allowed to compete in the actuall tournament.
Do the same with chess and go and the humans may as well not bother.
Terran have been clearly underpowered for years unless your superpower is the world's best micro, macro, and multi-tasking aka Maru.
Hmm the micro is pretty sick, would not be able to do it, maybe pros can quickly select individual units and pull back and morph to ravagers. The PC will always be better at micro.
AlphaStar Final has restrictions that make it not better than the top humans at micro.
AS does not have the same abilities it had in it's matches vs MaNa and TLO. AS's APM/APS, and reaction time in these agents were capped to equal GM level players.
@@Singularitarian still its not perfect sometimes it micros too much on protos
@@Singularitarian Matching average and max APM doesn't make the micro equal. For instance a computer can essentially freeze time, assess the whole screen and issue the perfect order on the ideal unit/s. A human at max APM could never have the same precision and will lose concentration over time.
Yeah, apparently being able to mass micro blink troops basically is 100% victory.
Wow! I did not expect this outcome. I wonder what Alphastars APM is though? In the matches against TLO I think they put a limit on it, but it was still able to output extreme APM in short bursts. Also, for an AI APM=EPM, which is the relevant number for human players.
Alpha star is limited in every match that it plays. They didn't want alpha star to be overpowering it's opponents just with more APM or Screens per minute. Also it's reaction time is pretty slow compared to a pro player. It's just very very accurate with everything it does and it makes it's limited actions count.
They fixed the APM thing like ages ago. How have so many people following this not heard of this yet. After the exhibition in Jan, they brought TLO into the Alphastar team, and they also consulted Mana and Kelazhur for restrictions that will make Alphastar comparable to humans.
AS does not have the same abilities it had in it's matches vs MaNa and TLO. AS's APM/APS, and reaction time in these agents were capped to equal GM level players. DeepMind also did the same to the agents that played on the ladder.
*Agents were capped at a max of 22 agent actions per 5 seconds, where one agent action corresponds to a selection, an ability and a target unit or point, which counts as up to 3 actions towards the in-game APM counter. Moving the camera also counts as an agent action, despite not being counted towards APM."
@@fupopanda They never actually fixed the APM (that I've heard of)... they set a pretty high APM limit competitive with pro player APM while ignoring that even for pros their effective APM is much lower and that the additional spammy extra right clicks to move aren't really comparable to a precision selected stalker blink in actual execution time. TLO and Mana almost look like paid shills if they brought them on board and then used them for the exhibition matches... if what you're saying is true then that's really bad.
why do you call the speeds "Faster" normal and "Normal" slowed down? XD
1v1 matches played at "faster" speed. So "faster" is the "live/normal" speed so to speak.
@@elleshar666 so... 1v1 matches are 'faster' while people play it? ... it kind of makes sense, but really _is weird._ :D
-- " Normal speed for 1v1 matches is not normal, but faster. " .......... right...... XD
I guess if I sad "Live speed for 1v1 matches is not normal, but faster. " it almost makes sense...
@@Verrisin When you start playing the game campaign for the first time you have an option to choose a difficulty. The "normal" difficulty is played on normal speed. Also if you play 1v1 and start from the lowest of the leauges the game speed is normal there too (unless it's changed later on) . So if you get a sense of the game and choose a harder campaign difficulty or move up the 1v1 ladder then game gets faster. Then the faster speed is what you will encounter all the time hence making it the normal game speed from your perspective.
edit: Yes 1v1 matches are played on "faster" speed. you can't change it unless you play with a friend or form a custom party. and if you wonder, you can't play the game at 4x speed.
@@elleshar666 Thanks. Makes a lot more sense now. ^^ - Also, I had no idea the game had a campaign, I thought it was just this moba we always see. ^^
@@Verrisin sc2 is not a moba it's a RTS :) the campaigns well worth the price. very fun and not difficult to get into . if you want to try 1v1 or team games it's free to play.
Is there a version of this series without all the pauses?
Makes me feel uncomfortable when Artosis says, AlphaStar is trying ______ . Scary
How TF is there no Serral Vs alphastar without this stupid bullshit pauses???
Stop pausing.
這就是腦蟲跟主宰的差別嗎?
I support Alphastar and I think it would become the best without doubt, BUT...
If Deepmind someday achieve this, I wish it would be without extra advantages like the ones it has today.
It should do it with a camera watching a monitor as visual input and one mechanic hand with 5 fingers using a keyboard and another interface to control a mouse.
In that way, any screen or control cheats are not possible, like:
1- selecting single units without control groups under the view of other flying units.
2- not wasting time to enclose a group of units with the mouse.
3- not need for control groups or waste actions by minute in no trivial stuffs.
4- winning most of the matches with all races in all maps.
If all that is fulfilled, then we can said that Alphastar strategies and decision making are better than human level.
To achieve the point I am asking, it needs to use another neural networks to command the keyboard and mouse in base to what alphastar needs, about the camera, Alphastar already has image recognition, but we cannot know if it cheats somehow.
Alright, I'll address each point one by one. But firstly I'd recommend you read the paper published by deapmind about 2 days ago about the starcraft AI, it covers almost everything you need to know.
So for the robotic hand and camera, there is a field of computer science called computer vision wherein one can take an image from a camera and translate it to a simplified view of unit positions. This typically uses AI but in fact does not require it in simple situations like starcraft units, which have a set icon. Simply scan the image for each icon in turn and hilight that section of the screen with that unit type. This can easily be done in real time with a 100% accuracy for unit positions given the simplicity of starcraft rendering. A similar thing can be done for the hand. For this absolutly no AI in needed, only classical control theory. Now if implimented both of these would have absolutly no impact on the AI, as it would still have the same info about unit positions and the same output to the computer (note that command latency was implimented for the AI to simulate button press times, in fact a real-world robot would have less latency as robots are significantly faster then humans). The way this was implimented mearly saves the work of implimenting these other systems that would take a lot of time for absolutly no payoff, as it mearly makes things identical to how they are now. Also it would require a large robot instead of a flashdrive, and would not be able to play against multiple oponents at once.
Now for the rest of your points about unit selection, the AI actually selects units the same way a human does in all situations. Command delay is incorperated into this so the whole "mouse drag time" this is not an issue. Overall the AI has no advantage in unit selection. The AI also has no advantage in terms of command rate (even during spikes) and in fact has a slightly lower command rate then human plays at a similar level at all times.
The paper mentioned a 99.5 or higher win rate in all possible matchups between races on all maps. Note this is against any players, not just pros, but its still better then most professionals.
Overall, what the AI is right now meets all your criteria for a better then human AI. This is with the exception of completly ussless addons that would merly take engineers away from the important points of AI research and have no effect on gameplay.
I shoud also mention that the AI is mearly giving commands to the actual game, it does not command the game for things like resources or unit health. In the sence it is impossible for it to cheat.
There was also a previous itteration of this AI that had no restrictions on command rate or vision (saw the whole map instead of a regular view) and it absolutly demolished all grandmasters. It did have the restriction of only playing protos V protos on one map, and heavily overused micro to win where it probably should have lost. But thats the advantage of AI!
@@whyOhWhyohwhy237 *Simply scan the image for each icon in turn and hilight that section of the screen with that unit type. This can easily be done in real time with a 100% accuracy for unit positions given the simplicity of starcraft rendering*
not sure if that methods allows to identify a soap of units (many overlords and mutalisk), invisible units, units dying, explosions, special effects, etc.
I would bet that this is not the method used, I imagine a neural network is used for map and unit feature recognition, besides, without a neural net approach, any illusion or trick exploited by humans to hide units or nuclear launches may not work vs a brute force recognition if does not learn with patterns and neurons.
*A similar thing can be done for the hand. For this absolutly no AI in needed, only classical control theory*
They are studying learning machine and you propose the use of math and normal algorithms?
A neuralnet can make mistakes (the same we do), a neuralnet should learn the keyboard reaction time, it should repeat an action in case this one did not execute, etc.
We are talking about a tiny neuralnet by the way, no more than 50 neurons for that task. With this the AI would have a playstyle or behavior close to humans.
*The way this was implimented mearly saves the work of implimenting these other systems that would take a lot of time for absolutly no payoff*
The way is currently implemented looks completely different to human control behavior.
First fact! The AI is amazing in micro! I saw her fight hand to hand with 30% less income over 15 minutes using warp prism super micro.
Players spam APM, giving a false APM which AlphaStar kinda match.
Players also lost time selecting individual units one far of the other due mouse moving time (that is why nobody can match that stalker blink or warp prism micro!
That is why all agents choose warp prism and stalkers as their main unit.
*Also it would require a large robot instead of a flashdrive, and would not be able to play against multiple oponents at once.*
You really think they use the image recognition neural net to play between agents?
Noo.. in that case they use unit position code, not image, not nothing.
That is why they made an special sc2 version, to train and play thousands of games by hour without the need to render graphics, about the keyboard and mouse could be simulated in virtual environment, it would not consume much processor, Or it can be approximated and learned with the latest agents, the most hard thing to do is to train the neural net for clever decision making, control and image recognition are just neuralnets interfaces that can be trained independent in some cases.
*the AI actually selects units the same way a human does in all situations*
No, it does not. The paper tells the difference and you can see it in the replays played in battlenet.
Few delays setted does not equal to human speed.
I saw the IA spotting (due minimap) when a single zerling jumped from the fog zone to the viewed pixel, and 100 ms later the two oracle from the other side of the map where recall, a pro player would have take 2 seconds or more to react.
Another real disadvantage for real players is eyes focus, we can see all the screen but we only see clear in the center of our sight focus. Sometimes the AI select an unit in the edge of the screen when only few pixels can been seen.
*The paper mentioned a 99.5 or higher win rate in all possible matchups between races on all maps*
Just in 4 maps, Serral may have 100% win rate vs master or below, and around more than 90% rate in grandmaster.
Zerg Agent may be in the worst places from 200 of grandmaster, they only play mostly with zerlings, roaches and ravagers. They dont even know how to predict shoots with the ravager.
*that would merly take engineers away from the important points of AI research and have no effect on gameplay.*
Proving that a computer is faster than US does not prove nothing, they need to prove that an IA could make better choices than US, for that they need to play as I said to match a more natural play style, something excited to see and erase any other cheat they may use.
Because the true is: that I don't really trust in all the words they said. Many of them could have their job at risk if they dont accomplish results, and already 2 years or more had past since they started.
Alphastar also needs to learn to said GG when a game is lost, the chess version of deepmind already measures winning probabilities according to position and pieces, this is done with another neural net.
@@AngelLestat2 Neural networks are not a magical fix-all. They are good at tasks where we do not know even the first step of programing something to accomplish a task, but they are quite bad at tasks with known solutions. Image recognition can be done by taking unit images (invisible images, dying images, everything) and scanning the image input with a "search square". Take the RGB values from this search square, put them into 3 matrices and find the dot product between those matrices and the desired unit images. the resulting number is a similarity score which can be compared to a pre-set threshold to create a boolean value for whether a unit is there or not. Its one of the first things you learn in a computer vision class.
They are studying neural networks for how they make decisions in star craft. This does not mean they have to use neural nets for every mundane calculation they ever need to do. Robots can be made VERY fast with classical control theory, in fact using the root-locust technique we can optimally tune PID controllers for the fastest possible controller to get to a desired position, such as hitting a key. This is done FAR faster then humans. Robots have been faster then humans for years ( th-cam.com/video/-KxjVlaLBmk/w-d-xo.html ). If you have an actual position and a desired position, as is the case with a typing robot (or using a mouse), you should absolutely never use AI. AI are best used as high level controllers, low level controllers should not use them for best results are these are solved fields. We know almost everything about the robot so AI is not necessary in the slightest and in fact always yields worse solutions and the know solutions are optimal.
The previous version of the AI was godly at micro, and demolished everyone because of it. This version has significant limitations on the APM. In fact the pro players are quoted as saying its micro is on par with humans and does not feel massively better. Again, there are videos of AlphaStar playing with great micro, but those are previous versions.
"You really think they use the image recognition neural net to play between agents???"
That's literally the exact opposite of what I'm saying. I'm saying that a robot would be impractical as a counterargument to you saying they should use a robot. I'm the one advocating for using a purely simulated AI.
I don't actually play StarCraft that much (I'm more interested in the paper as an AI researcher then a gamer), so your probably right about the reaction/focusing time. they might have to account for that more in subsequent iterations.
You're right in saying that they are trying to measure the high level play of the starcraft AI more so then the low level microing. However I believe that these results are incredibly significant in terms of overall strategy. Again, I don't actually play the game that much so I have to go on the word of professions, but they say its not superhuman in terms of micro and I tend to trust them, thus any wins are more so due to strategy then speed. I absolutely agree that any superhuman micro needs to be removed to properly test the strategy of the AI, but things like using robots and image recognition are not the best way to achieve this. However I don't think a "natural play-style" is the metric we should use to determine success. AI are notorious for crazy counter-intuitive ideas that somehow work. You can see them in chess and GO AI.
I would advise you to trust researchers, peer review is a hell of a process and replication is key.
Alphastar will never "learn" to say GG. Its goal is to win, and even if you have only a 1% chance of winning saying GG lower that to 0% and is thus never a good option. If given the option the AI will literally never surrender, as their only goal is to win. I believe Alphastar already has a win probability calculation network, but if the researchers were to program it to say GG after a certain point it would be just that, programing. Right now the researchers are just manually telling it to say GG when they know the game is lost, which is fine for now.
@@whyOhWhyohwhy237 your link about robots faster than humans doesn't work
Terran weakest confirmed!
Why is Serral off creep... You're down units and tech... you could easily work that choke with the buildings and buy time... I feel like the 4 games I've watched with Alphastar were thrown. Serral could have taken some games.
It can be tempting to anthropomorphize, especially as the AI becomes increasingly complex, but it's not a person. AlphaStar doesn't 'want' to do anything. It is responding to stimuli. I would also be careful with what you call a 'mistake.' Some moves may appear to be a mistake on their face, but on reflection there may be a logical explanation. I understand you may be taking creative license with some of the phrasing. I just find it a bit jarring.
Would you say that a plant wants to turn in such a way it maximizes the amount of sunshine it recieves?
en.wikipedia.org/wiki/Heliotropism
@@petrkinkal1509 No I would not. Plants, like AlphaStar, are not conscious.
@@Sid11358 Fair enough still there is a point where it gets really hard to tell if something is conscious or not.
I mean what about chimps, dogs, snakes, ants? At what point is it conscious and at what point is it just responding to stimuly?
@@petrkinkal1509 I agree there is no clear delineation. It may be that Deepmind becomes conscious (in its own fashion - analogous to how dolphins or rats or crows are conscious) in the near future and we don't notice or misinterpret its consciousness as formulaic responses. On the other hand it may never become conscious, but we misinterpret its responses for consciousness, though this difference may have little practical value (if it appears conscious it may as well be). The crossover point will be difficult to observe.
@@Sid11358 "difference may have little practical value"
This I agree with this so much.
But that is a big part of my point it doesn't matter that much if it wants to win or not if will still do best it can to win so it effectively acts like if it wants to win.
The anthropomorphising would start to matter if people started doing something like for example having it also play few matches with lover tier players so it doesn't feel bad for always loosing (or expecting it to feel bad about a player it is constantly trashing) that would be pretty ridiculous.
Could make a link to the video of this game WITHOUT the commentary ?
take of your headphones. Boom! no more commentary
Please just let the game run without pausing, it really takes away from the game when viewing it.
I think Artosis is making the cast accessible to people who are only familiar with machine learning and not with starcraft 2
it's an analysis for non-players to understand, not a GSL stream
@@LSZreduction The community has kinda moved past that point a while ago.
@@StevenCasteelYT What are you trying to say? I am sure that there are far more people who is interested in Deepmind but never even heard of starcraft than the number of sc2 players.
Steven Casteel wake up Steven, the “community” of SC players you’re reffering to is only getting smaller and smaller, and this is a great way to bring new people in and make them understand what the game is even about. All of the “newbie friendly” content is outdated, everything is absolutely different than it used to be even like 2 years ago and everything is constantly changing. I’ve been with the game for 20 years now and absolutely love the idea of making content for people that simply haven’t been around for this long. Nothing wrong with this, and videos of the match in real-time are avaliable as well. So stop hiding behind the “community” just because you don’t feel like watching something that is not targeted at you.
Dont stop the game so often
Artosis is great but I really think pausing the game every 3 seconds is a terrible idea. I know these videos are for people who never played Starcraft but people just can't get how fast both players have to think to plan and execute all the actions when you slow down the games like this.
but that shows, if game would be slowdown 3 to 5 times, then humans would win with severe difference, alphastar plays like big subconscious, no logical thinking there at all(only reflexes and deep interpolation between known outcomes/actions), people just can't do much logical thinking on those speeds, but if they would have such time...
sure, for players that want to learn to play on their own. but this is analysis/commentary. you can get live casting by watching pro games live.
also, please use speed up/down and just skip ahead features. it's 2019, i think most people shouldn't only be limited to watching things at 1x anymore, especially if you're already familiar with the basic content. at least that's how i watch talks and stuff now. i get it if you're on your phone though, but on the computer, it's so much time saved.
It's an analysis not a cast
talk too much. damm
If you are thinking on record more alphastar games, pls stop pausing
So much is going on in a high level SC2 game how can you absorb everything in real time?
@@dannygjk osmosis learning, very usefull
it's to help those of us who haven't played SC since BW but want to see what AI can do and also understand what's going on. you ppl... I mean really.
I know Terran is bad at the moment, it turns out...humans, in general, are bad....oh well, another sport ruined by AI.
By that logic DeepBlue ruined chess in 1996. And recently backgammon. Except that it didn't, and humans still do chess tournaments. And those are still fun. Just because there is an entity, like an AI or a strong player somewhere out there better than you, doesn't mean you have to stop playing, or stop enjoying a game.
Whether or not these advanced machine learning techniques will be used for the betterment of all humanity or just a few wealthy individuals is the real issue here.
I just can not watch your old replays, the constant stopping, and over explaining, just makes me fast forward threw the whole game
very very very slow you pause the game too much
it's for the people who have never played the game
I really struggle to understand how you people saying this don't know that already
It's not a RTS, it's a SlowTS
Anybody who do not play the game cannot keep up at all with the regular casts.
These breakdowns are brilliant, it separates from the other replay channels and very insightful.
Might be a spoiler but it seems an Ai can beat a human with everything but Terran. Moral of the stroy humans are bad. Force A.I. to play Terran and it's all fine, never gonna win. Perfectly balanced for both humans and ai alike, as all things should be. xD xD xD
I think the reason it lost with terran was because of the first four lings that messed with Alphastars build. In every game it controlled the "flow" of the game except in the terran game were Serral did whatever he wanted.
@@Zettern96 Might be, but the reason it happened this way might as well be because Terran is in a shithole and even an AI can't muster a win with it, when it managed to do it quite comfortably with the other 2 races versus the same opponent.
Just let AlphaStar unlock it's full POTENTIAL!!! th-cam.com/video/0EYH-csTttw/w-d-xo.html and th-cam.com/video/DXUOWXidcY0/w-d-xo.html
@@Figgy20000 Oh yeah, i did imagine these 5 games against serral were how the AI rebellion defeats our best generals. But in reality, the AI rebellion would look a lot more like Automaton 2000 playing with it's 300'000 apm
Frankly, that Terran play was bad from AS.
Just for the record. Serral isn't even a champion in any of the recent Premier tournament. Far from "best player in the world". It's good that they choose Serral to feed AlphaStar, minor players before the actual players with actual resullts.
He won WCS Fall and GSL vs. The World. How is he not a champion in any of the recent Premier tournaments?
@@Skobbolop GSLvtW is a showmatch in the middle of Code S. WCS Fall, never heard of that.
GSLvtW and WCS finals are Premier Tournaments according to liquipedia:
liquipedia.net/starcraft2/Premier_Tournaments
Serral won 3 of them in 2019
cant watch ur replays, cuz u keep pausing the game....its just so annoying!
no it's not, he's doing in depth analysis
What :D Even at his speed calling "faster" as "normal" I watched at 0.75 to understand what is going on
cmon Artosis. Starcraft players are fastest and smartest gamers in the entire world! And you pause it like every 30 seconds to explain extrimely simple stuff.
AI enthusiasts aren't all Starcraft players.
He's targeting a general audience.
Pavel Averyanov I wrote it in a different response, but am gonna do it again. Stop whining just because you don’t feel like watching something that is not targeted at you, simple as that.
not all of us are geniuses keep that in mind.
First of all: This vid is NOT meant for SC players only.
Secondly: SC are not all smart judging from your comment.
.
I absolutely hate hate HATE this commentary style... I know how a ravenger works and that it evolved from a roach. This is was so boring it made me angry. Watch this video to help you sleep or if you have never played StarCraft before and are curious.
Yes, seriously. You can make a SC2 tutorial OR make an Alphastar replay. But don't do both at the same time. Make 2 separate videos. One for newbs and one for pros. This is like watching Deep Blue play chess and then the commentator pauses the game every 2 seconds to explain the basics of chess. We aren't cavemen. Is there really people who like Alphastar but don't even know how to play SC2? Sigh...
There are literally dozens and dozens of alphastar casts.
@@PRepublicOfChina Yeah, and we paid so much money for this.
@@weestro7 Well the whole point of the comment section is for people to voice their opinion, which helps give feedback to the creator to make more entertaining content so that they can make more money. So criticism is of high value when it is constructive.
@@zakreiner2469 I think I agreed with your last sentence.