That apm limit really is brutal for alphastar. It looks like that attack move command was it's 20th action in that period and did not have enough apm left in the 5 second period to re-target the disruptor shot. I'm surprised the developers didn't do an override to like 50 or 75 apm in combat involving more than 10 units or something.
I just want to point out that whoever it is that was playing the zerg player didn't play his swarm host/nydus nearly as well as the players in Blizzcon have, nonetheless, well played by AlphaStar, I think we can learn a lot from what it's doing
Well, it's *some* GM on the European ladder. Check out Beastyqt's videos on getting into GM league with meme strategies like marines only, it's clear that not everyone on there is Serral.
Indeed, Alphastar's current achievement is reaching GM but, as Winter has pointed out many times, there's about as much difference between a top GM and a low GM than there is between a low GM and a Diamond 3 player - a ton of difference
I love how much progress the AlphaStar project has made in 2019. It was a little disheartening to see such a "weak" version debut awhile back (the one that could get into 2000 APM in short bursts), but seeing this new version that is more greatly confined to human-level inputs still dominate the game is amazing.
This isn't serral. I'm sure of it. Serral plays zerg like he's conducting a symphony of death. You can see it in the minimap. This player doesn't have the flow.
A small remark on the statement that AlphaStar learns from professional human players: the machine learning model actually has been trained on human replays in the very beginning but this is not really for actually "learning" from humans but rather to speed up the training process. The problem is, that due to the fact, that the model is initialized randomly in the beginning it has no intuition of vision, logic, units or what so ever. It would simply do completely random actions and take a ton of time until it eventually finds out some meaningful actions by chance. By training it on replays, this difficult early phase can be significantly speed up but it's not like, that it actually really learns human tactics or behavior.
Personal control for each blink stalker would make them the best unit in the game.. The AI would crush everything and everyone without limitations with only gateway units.
Fascinating video! Thanks for making this. Just subscribed and I can't wait to see more AlphaStar Final analysis videos! And thanks for putting "Final" in your title so we can search for AlphaStar Final videos.
no, I'm pretty sure mass stalker warping in still counts as apm, and is basically spam. Only early game human APM is spam, for like the first 3 minutes.
computers apm is OP, its actions are all calculated and chosen in nanoseconds. google overstates its fairness. literally its like being able to think in spiderman matrix slow motion and choosing to make mistakes because google sets limitations but already all the correct answers and probabilities are calculated for every single possible move.
But how to come up with fair limits? Human capability varies greatly even among pros. How many deliberate and sensible actions can a human reach for 1s, 5s, the whole game ....
@@ekrotte8714 it isn't even about how many actions to take. Humans have limits to how accurately we can move and click the mouse, and almost every player clicks multiple times when move commanding, stuff like that. Even things like being able to see exactly where each unit is going to path, so for example you can see a million bikes coming down in the middle of the stalker center of mass but somehow exactly the right stalker takes a step to the side and almost everything misses... Almost no actions but not very realistic for human ability
@@emissarygw2264 yes, because blowing up a third of your army with your own ruptor is not very realistic for human ability, and then getting like 2 good shots over the whole game and sacrificing your mineral line to keep your army alive isn't realistic for human ability... McCanning can guarantee better ruptor records if he's given that much leeway. (not worrying about harass at home) High GM humans average 300 easily, which, if we say a quarter is unit production, which is almost guaranteed to be on point, that's 65/75, if we say movement is 1/3 efficient and that's a half of the 300 actions, that's another 50/150, and if we say the remaining is abilities and camera movement and other shit which are usually quite efficient, that's say, 50/75, that's 165 effective APM, a VERY low estimate, that hits AlphaStar's average, though not its max. Note, if we use higher efficiency numbers, because these are quite low, we can easily pull out 200 APM.
@@soju69jinro Not for long. Computers went from an actual joke to beating professionals in a few years. Beating the best pros under tournament conditions will be only a few more.
To calrify the APM restriction. The agent is allowed to do 22 actions per 5 seconds, but we aren't talking about Starcraft actions here. Think of an agent action like a command. So selecting a high templar, selecting storm then clicking somewhere counts as 1 agent action. Whereas Starcraft would count this as 3 actions. Additionally, Starcraft does not count moving the camera as an action (by scrolling or clicking). Alphastar considers camera movement an action. So Alphastar is limited to an agent APM of 264, but this inlcudes moving the camera. From analysing the replays, Alphastar averages 190-210 in-game APM. This seems to indicate that a large proportion of APM is for moving the camera. Alphastar's APM in-game should be some number higher than 264 minus the actions needed to move the camera.' Another interesting thing I noticed from the replays is that Alphastar's timing is perfect on the macro side. Orbital command, chrono, and inject efficiency is 100%. This is potentially undesirable as this would imply that in the middle of a fight Alphastar might leave to go do some macro. This might explain Alphastar's sloppy micro. It would make sense that if limited, Alphastar would tend towards winning games through macro. I haven't done any replay analysis by watching them yet but I really want to see if Alphastar uses command queuing. Seems like a great way to maximise APM (I assume that it does not always do 22 actions per 5 seconds which would mean it is wasting potential). Could overcome the issue of it leaving to do macro by, for example, giving a queen a move command then queuing the inject after that to time the inject out perfectly ahead of time.
I'm not familiar with the starcraft environment but it seems silly to restrain the algorithm with the camera shenanigans. I would assume that they could simply input the map in full size instead of the game's typical viewpoint. But all of this is to replicate a player's constraints.
They call deepmind neural learning or artificial neural network because that's exactly how a human learns stuff , thank you Winter for pointing the real question and the real point of developing a deep mind a.i , not to be shocked by blink micro (I saw some other casters focusing on b.s) that's why I love your channel more than any other sc2 channel , you are super smart with a greater vision , even your bronze League hero jokes reflect your great personality , the fact is this deepmind gives unlimited calculation power to people who don't deserve it in banking finance and any other job made for men and I believe it's dangerous for most of us normal human beings , they can do anything a human can do with better precision with no moral values.
Reacting to the end of your comment: you wouldn't use a neural net for calculations though. You might use that to analyse signals from news, Twitter etc, and give you hints, but certainly not for pure number problems. That would be the most convoluted way to get a wrong result.
@@ryousenketsu6053 there's a world of difference between "4+7" and "given what's happening on Twitter at the moment, should I buy/sell any stock" or "given the state of this StarCraft 2 game, which move should I do next". In one case, a neural net would give great results. In the other, not so much.
@@kernald everything can be transcribed to set theory and logic , everything you do also has a pattern , math is not your school's 7+4 , this a.i can make movie trailers but does it watch a movie when he comes back from work? No it makes logical assumptions based on a pattern to understand a movie. You define a goal and calculate the possible ways to reach a goal that's why military strategy is so complicated and requires genius work like Napoleon , a.i can see all the possible ways , We have to mention that the a.p.m was limited in this game but in reality it can perform 1000s of apm in second
I think to select specific army units, alpastar presses select all and control clicks the units he wants to use. Pretty sure beasty confirmed it doesn't use a mouse for army micro, and moves it's camera via hotkeys and camera hotkeys
Those shield batteries were I think a backup so that with it's expansion if he got rushed then he'd still be okay, letting him expand a little more greedily and if the expansion lives then the cost is basically mitigated.
NoobStarcraftPlayer question: Why does everyone stay so far away from Hydras regardless of what the enemy has? I've seen every other unit utilized but Hydras, are they terrible for any situation or something?!
Prot i think they are hard to balance. A while ago hydras were buffed and everyone was using them. I think they got a dps nerf and they fell out of favor. They’ll become much better with the newly proposed microbial shroud ability
Ahhhh cool cool, ok thanks guys. Im only gold Zerg (Hydras work in gold btw xD) but I hear you guys, i just didn't know because... uh... gold xD Thanks again for the info
@@joethestrat One more thing, Hydras are fine for the metal leagues, what matters the most is that you macro well. I find Roach Hydra comfortable to make (the amount of gas lines up quite nicely) and I win tons of games with it on Plat. They don't require much micro, so they're probably much more effective in lower leagues than a unit like swarm hosts would be.
@@Brainth_1780 Oh awesome, thanks for the tip. I kinda guessed Hydras had some inherent flaw I couldn't see since no pros used them in any tourneys I watched... You guys helped clear it up though, thanks again. I'll admit... I kinda like them aesthetically a lot xD hahaha
Hydras do have inherent flaws = expensive and yet so fragile. They are glass cannons. Think of steamed marines that don’t have medivac support and are way more expensive
Alpha star is so good at long term economical planning, for human in this stressful situation, human usually make more units than necessary and you look at all the alpha star replays, alpha star always have a better economy than human players
Other advantages for the AI is that it doesn't get tired, sick or tilted. Nor does it worry much about unwinnable situations. On the other hand, if Alphastar had a choice, would it even play Starcraft? If not, maybe we're safe after all.
You can actually see that it does not use any control groups because it has an API to access buildings and units. For example when you were showing the awkward recall - protoss does not have any control groups and does not move the camera. Yet he is able to select phoenix that is not in his view and then circle through his Gateways to order units, again without going back to them and then using the recall without switching camera to the nexus. Alphastar does incredible job, but it has much more control capabilities that the player does.
You can't see the players control groups with the WCS overlay that was used in the video. However, if you watch closely you can see that right before the recall all its units on screen become deselected indicating that AlphaStar has selected the nexus to access recall. Whether that's through a specific control group or an API action it still had to do effectively the same thing as a human would do.
@@Obzerver All the commentators noted that AlphaStar does not use control groups. If it does, I have no questions. However, from what we have seen from all casters, it doesn't. There is a huge difference between player using control groups (having to constantly add and remove units from there) and AlphaStar being able to select any subset of units/buildings on the map at will. It makes both Macro and Micro much much easier.
It might look at locusts and think "if I fight them I am guaranteed to win". Since every time Alphastar fights locusts, no matter how bad the fight is going it magically wins the fight after long enough. I think it would make sense that Alphastar hasn't figured out that it can also beat locusts by doing nothing. It's reinforcement learning. Every locust wave convinces it more and more than the chance Alphastar wins (i.e. kills all the locusts) is 100%.
@@andrenowaczek4909 On the other hand, in the early stages of learning, the AI has a lot of random permutations to be more "curious". It would have a fight kiting locusts and figure out that this overall meant not losing any units or HP on them. Either this was so rare it never really happened, or it values the postioning on the map over the lost health.
Remember how some pro players would flex winning against 1 v 7 hardest ai? In short future we will see how ai would be destroying 7 humans opponents alone. Time to get that sweet revenge!
I wonder what race would be the most effective from an AI with or without restrictions. I'm thinking Terran, if every marine can be controlled simultaneously and kite mostly everything.
Tools have always been our true strength. Not how well we use the tools. Only makes sense that a tool we make to use our tools would do it better than us.
That would be stupid. Look up Automaton 2000 videos to see how utterly broken machine-controlled micro can be - you don't need any of Alphastar's strategy learning to beat a human player.
In the Blizzard forums, people trash Alphastar because it was unable to beat any of the major pro players. But since it had something between 95% and 99% win ratio, they are actually thrashing most of the SC2 community, including themselves, without realizing it. They always say how flawed the bot is, or how easly it could be countered if this or that was done, as if they were worthy of teaching the AI how to play.
You think it's disturbing to have a machine which learnins like a human, but that's what its supposed to be. Scientists are trying to reproduce the learning process, the human type of a learning process. There's still a huge difference between being able to learn like a human and being a human like being. First of all, choosing the goals. The non obvious will to win a game is hardcoded into AlphaStar, it didn't come up with that. On top of that, human needs a couple of games to get a grasp of whats going on - AlphaStar needs hundreds of games to get it. Still nothing to worry about. It's still just a set of optimization techniques.
This was a surprisingly close match! If those swarm hosts hadn't gotten caught at the end, we could have had an even longer game o: I wonder who that zerg player was...
Idk if it's just me... but anyone else interested to see what alphastar could do when the take off the human restrictions winter mentioned? Just curious.. would alphastar be able to make a bazillion apm or display a terrifying level of control? What would someone do against say a grandmaster or higher level terran alphastar that could control large numbers of individual marines or groups of army at the same time? Would it be able to narrow its focus to multiple spots on the map at once all while microing workers or training units? A multi pronged attack using multiple types of units.. it would steamroll pro players. Lol would be cool
Wondering why they only taught alpha star how to play the easiest race? I've only ever seen it play toss. Zerg would be fun to see how it decides to play...
9:54 and that's why i love Winter ,his not even trying to humor us but still i crackle alot on his streams exactly for this 9:54 Biggest dude in starcraft we love u Winter
Terrifyingly, unlike the previous super-micro-stalker, I cannot tell what makes the Protoss so good in this game. It looks like if Serral main race is Protoss :/
Very strong macro never missing warpins constant production the micro and army control isnt great but it dosent need great micro when your even or ahead on supply vs swarmhosts
The program has perfect memory can do instant math on damage and can count all your units and never forgets. Still not quite an even match but impressive nonetheless that it learns so quickly.
This is not how Alphastar works at all. There is no calculating damage or any kind of unit parameter calculations. It takes in raw pixel data from the games camera and makes decisions based on the information. It's only as good as the amount of games it has experienced and learned from.
Your harping on about apm, but have you stopped and considered ALL OF ALPHA STARS APM IS E-apm. IT'S IMMPOSIBLE for humans to have that high of an EFFECTIVE-apm consitantly.
Its explained better in the article. 22 Agent actions per 5 seconds where an agent action is not necessarily an action that counts towards a single in game APM.
2009: how many AI's should I add to make this a challenge?
2019: how many of me would it take to kill 1 Alphastar AI?
The classic real time strategy game featuring three iconic races: the visious Korean, the noble Finnish and mighty Artificialy Intellingent
Nationalities aren't races you fucking dipshit
@@supermalevitality1187 u sure?
@@venyakrabov yes, they're different things
@@supermalevitality1187 are you really really 100% sure?
@@supermalevitality1187 Geez, someone's a little angry today. Sleep ok?
"I'm so good, I'm going to help you kill some of my units and still win." - AlphaStar
I'm so accustomed to Winter's face cam that it feels really weird hearing him just talking without it :D
he's afraid he'll accidentally don a mask and say: "Liberate Hong Kong. Revolution of our times."
There is no face.
That's not Winter talking.
It's AlphaStar.
And sitting in a spawning pool...
"Some incredibly human mistake by alphastar"
I can almost hear serral "you pathetic biologic human"
Do Alphastars dream of Electric Trophies?
Best comment ever !!!
sounds like you've read Isaac Asimov.
A man of culture
@@_Wai_Wai_ I do think it is more from Philip K.Dick,'s novel. This novel was the basis for Blade Runner, a movie I deeply recommend.
@@TheEolee Ok, cool. I thought he was referring to Isaac Asimov's Robot Dreams.
This is not AlphaStar final form.
He doesn't even know how to BM his opponent. I am sure this will be fixed.
it would probably improve the winrate if it tilted the opponents
whats BM?
@@cruentus1816 short for "bad mannors"
Would love to see A-star do a few troll strategies, like a pylon heart, killing a worker early on, playing with a specific set if units, etc.
final form would be it after winning, tabs out the player and the prompt for uninstalling SC2 appears
That apm limit really is brutal for alphastar. It looks like that attack move command was it's 20th action in that period and did not have enough apm left in the 5 second period to re-target the disruptor shot. I'm surprised the developers didn't do an override to like 50 or 75 apm in combat involving more than 10 units or something.
I just want to point out that whoever it is that was playing the zerg player didn't play his swarm host/nydus nearly as well as the players in Blizzcon have, nonetheless, well played by AlphaStar, I think we can learn a lot from what it's doing
Well, it's *some* GM on the European ladder. Check out Beastyqt's videos on getting into GM league with meme strategies like marines only, it's clear that not everyone on there is Serral.
Indeed, Alphastar's current achievement is reaching GM but, as Winter has pointed out many times, there's about as much difference between a top GM and a low GM than there is between a low GM and a Diamond 3 player - a ton of difference
I want to see a moment when AlphaStar starts to write in the chat
offensive GG's are the best
or it can type "humanity ends today".
@@Falzelo or just noob by playing 5000 games with toxic players.
@@Falzelo If Alphastar writes that Alphastar dies, regardless of if it's joking.
"I estimate the probability of winning to be over 99%."
I love how much progress the AlphaStar project has made in 2019. It was a little disheartening to see such a "weak" version debut awhile back (the one that could get into 2000 APM in short bursts), but seeing this new version that is more greatly confined to human-level inputs still dominate the game is amazing.
This isn't serral. I'm sure of it. Serral plays zerg like he's conducting a symphony of death. You can see it in the minimap. This player doesn't have the flow.
yeah, I agree, just look at that nidus spawned in the middle of the map, when has Serral done that? xD
Alphastar was about 5.8k mmr which is lower gm on eu so this may be namshar, lambo or maybe tlo not quite the level of serral or reynor
PaleGhost69 he was clearly joking
Can we report Alphastar for botting?
rofl
it didn't even "gl hf". What kind of monster are they creating?!
I think they removed the chats from the replays
So they admit how foul-mouthed their "monster" is!
@@waxkun what do they want to hide from us?!!??!!
All chat+names were removed from the replays for anonymity
It probably begged for help
A small remark on the statement that AlphaStar learns from professional human players: the machine learning model actually has been trained on human replays in the very beginning but this is not really for actually "learning" from humans but rather to speed up the training process. The problem is, that due to the fact, that the model is initialized randomly in the beginning it has no intuition of vision, logic, units or what so ever. It would simply do completely random actions and take a ton of time until it eventually finds out some meaningful actions by chance. By training it on replays, this difficult early phase can be significantly speed up but it's not like, that it actually really learns human tactics or behavior.
AlphaStar is the ultimate life form in SC2. It will be the end of us all.
only John Conn...er Serral, Maru and Stats can lead the human resistance
I want to see an AlphaStar on full potential doing splits against Banelings!
Personal control for each blink stalker would make them the best unit in the game.. The AI would crush everything and everyone without limitations with only gateway units.
13:15 was quite painful
Fascinating video! Thanks for making this. Just subscribed and I can't wait to see more AlphaStar Final analysis videos! And thanks for putting "Final" in your title so we can search for AlphaStar Final videos.
You commentary is easier to follow and more informative than any gsl/global or other SC2 videos I have watched ever.
Note that while AlphaStar is being APM limited, each action of the AI is deliberate while a good chunk of human APM is spam.
no, I'm pretty sure mass stalker warping in still counts as apm, and is basically spam. Only early game human APM is spam, for like the first 3 minutes.
computers apm is OP, its actions are all calculated and chosen in nanoseconds. google overstates its fairness. literally its like being able to think in spiderman matrix slow motion and choosing to make mistakes because google sets limitations but already all the correct answers and probabilities are calculated for every single possible move.
But how to come up with fair limits? Human capability varies greatly even among pros. How many deliberate and sensible actions can a human reach for 1s, 5s, the whole game ....
@@ekrotte8714 it isn't even about how many actions to take. Humans have limits to how accurately we can move and click the mouse, and almost every player clicks multiple times when move commanding, stuff like that. Even things like being able to see exactly where each unit is going to path, so for example you can see a million bikes coming down in the middle of the stalker center of mass but somehow exactly the right stalker takes a step to the side and almost everything misses... Almost no actions but not very realistic for human ability
@@emissarygw2264 yes, because blowing up a third of your army with your own ruptor is not very realistic for human ability, and then getting like 2 good shots over the whole game and sacrificing your mineral line to keep your army alive isn't realistic for human ability...
McCanning can guarantee better ruptor records if he's given that much leeway. (not worrying about harass at home)
High GM humans average 300 easily, which, if we say a quarter is unit production, which is almost guaranteed to be on point, that's 65/75, if we say movement is 1/3 efficient and that's a half of the 300 actions, that's another 50/150, and if we say the remaining is abilities and camera movement and other shit which are usually quite efficient, that's say, 50/75, that's 165 effective APM, a VERY low estimate, that hits AlphaStar's average, though not its max. Note, if we use higher efficiency numbers, because these are quite low, we can easily pull out 200 APM.
Looking forward to the final showdown of Serral vs. AlphaStar!
Serral wins.
@@soju69jinro Not for long. Computers went from an actual joke to beating professionals in a few years. Beating the best pros under tournament conditions will be only a few more.
Yes, more Alphastar footage! Keep em comin
very well informed presentation, thank you
I still think that presenters are showing AlphaStars camera too rarely.
This is such good content. Your casting is just amazing as usual
To calrify the APM restriction. The agent is allowed to do 22 actions per 5 seconds, but we aren't talking about Starcraft actions here. Think of an agent action like a command. So selecting a high templar, selecting storm then clicking somewhere counts as 1 agent action. Whereas Starcraft would count this as 3 actions. Additionally, Starcraft does not count moving the camera as an action (by scrolling or clicking). Alphastar considers camera movement an action. So Alphastar is limited to an agent APM of 264, but this inlcudes moving the camera. From analysing the replays, Alphastar averages 190-210 in-game APM. This seems to indicate that a large proportion of APM is for moving the camera. Alphastar's APM in-game should be some number higher than 264 minus the actions needed to move the camera.'
Another interesting thing I noticed from the replays is that Alphastar's timing is perfect on the macro side. Orbital command, chrono, and inject efficiency is 100%. This is potentially undesirable as this would imply that in the middle of a fight Alphastar might leave to go do some macro. This might explain Alphastar's sloppy micro. It would make sense that if limited, Alphastar would tend towards winning games through macro.
I haven't done any replay analysis by watching them yet but I really want to see if Alphastar uses command queuing. Seems like a great way to maximise APM (I assume that it does not always do 22 actions per 5 seconds which would mean it is wasting potential). Could overcome the issue of it leaving to do macro by, for example, giving a queen a move command then queuing the inject after that to time the inject out perfectly ahead of time.
I'm not familiar with the starcraft environment but it seems silly to restrain the algorithm with the camera shenanigans. I would assume that they could simply input the map in full size instead of the game's typical viewpoint. But all of this is to replicate a player's constraints.
They call deepmind neural learning or artificial neural network because that's exactly how a human learns stuff , thank you Winter for pointing the real question and the real point of developing a deep mind a.i , not to be shocked by blink micro (I saw some other casters focusing on b.s) that's why I love your channel more than any other sc2 channel , you are super smart with a greater vision , even your bronze League hero jokes reflect your great personality , the fact is this deepmind gives unlimited calculation power to people who don't deserve it in banking finance and any other job made for men and I believe it's dangerous for most of us normal human beings , they can do anything a human can do with better precision with no moral values.
Reacting to the end of your comment: you wouldn't use a neural net for calculations though. You might use that to analyse signals from news, Twitter etc, and give you hints, but certainly not for pure number problems. That would be the most convoluted way to get a wrong result.
@@kernald everything is calculation
@@ryousenketsu6053 there's a world of difference between "4+7" and "given what's happening on Twitter at the moment, should I buy/sell any stock" or "given the state of this StarCraft 2 game, which move should I do next". In one case, a neural net would give great results. In the other, not so much.
@@kernald everything can be transcribed to set theory and logic , everything you do also has a pattern , math is not your school's 7+4 , this a.i can make movie trailers but does it watch a movie when he comes back from work? No it makes logical assumptions based on a pattern to understand a movie. You define a goal and calculate the possible ways to reach a goal that's why military strategy is so complicated and requires genius work like Napoleon , a.i can see all the possible ways , We have to mention that the a.p.m was limited in this game but in reality it can perform 1000s of apm in second
Mixing up with some DT is just a great idea.
Random question: do hallucinated Archons destroy Force Fields?
they dont
no
they don`t destroy force fields but can go through them
@@jovantrajcevski2140 im pretty sure they dont
@@fishlobs he is right. They dont break it but go through
Hopefully blitz can get some alpha star agents running with various levels of constraints and conditioned bias for a next level vs ai mode
I think to select specific army units, alpastar presses select all and control clicks the units he wants to use. Pretty sure beasty confirmed it doesn't use a mouse for army micro, and moves it's camera via hotkeys and camera hotkeys
Those shield batteries were I think a backup so that with it's expansion if he got rushed then he'd still be okay, letting him expand a little more greedily and if the expansion lives then the cost is basically mitigated.
How where can I watch the finals?
NoobStarcraftPlayer question: Why does everyone stay so far away from Hydras regardless of what the enemy has? I've seen every other unit utilized but Hydras, are they terrible for any situation or something?!
Prot i think they are hard to balance. A while ago hydras were buffed and everyone was using them. I think they got a dps nerf and they fell out of favor. They’ll become much better with the newly proposed microbial shroud ability
Ahhhh cool cool, ok thanks guys. Im only gold Zerg (Hydras work in gold btw xD) but I hear you guys, i just didn't know because... uh... gold xD
Thanks again for the info
@@joethestrat One more thing, Hydras are fine for the metal leagues, what matters the most is that you macro well. I find Roach Hydra comfortable to make (the amount of gas lines up quite nicely) and I win tons of games with it on Plat. They don't require much micro, so they're probably much more effective in lower leagues than a unit like swarm hosts would be.
@@Brainth_1780 Oh awesome, thanks for the tip. I kinda guessed Hydras had some inherent flaw I couldn't see since no pros used them in any tourneys I watched...
You guys helped clear it up though, thanks again. I'll admit... I kinda like them aesthetically a lot xD hahaha
Hydras do have inherent flaws = expensive and yet so fragile. They are glass cannons. Think of steamed marines that don’t have medivac support and are way more expensive
Alpha star is so good at long term economical planning, for human in this stressful situation, human usually make more units than necessary and you look at all the alpha star replays, alpha star always have a better economy than human players
Other advantages for the AI is that it doesn't get tired, sick or tilted. Nor does it worry much about unwinnable situations. On the other hand, if Alphastar had a choice, would it even play Starcraft? If not, maybe we're safe after all.
Alphastar's canon rush defense as zerg in a very recent video was amazing to watch.
Is it just me or is it really weird not seeing winter in the vid? Lol
Yess another one I love your comments on AlphaStar games
You can actually see that it does not use any control groups because it has an API to access buildings and units. For example when you were showing the awkward recall - protoss does not have any control groups and does not move the camera. Yet he is able to select phoenix that is not in his view and then circle through his Gateways to order units, again without going back to them and then using the recall without switching camera to the nexus. Alphastar does incredible job, but it has much more control capabilities that the player does.
You can't see the players control groups with the WCS overlay that was used in the video. However, if you watch closely you can see that right before the recall all its units on screen become deselected indicating that AlphaStar has selected the nexus to access recall. Whether that's through a specific control group or an API action it still had to do effectively the same thing as a human would do.
@@Obzerver All the commentators noted that AlphaStar does not use control groups. If it does, I have no questions. However, from what we have seen from all casters, it doesn't.
There is a huge difference between player using control groups (having to constantly add and remove units from there) and AlphaStar being able to select any subset of units/buildings on the map at will. It makes both Macro and Micro much much easier.
I feel like it was treating the locusts as if they were actual units. Rather than kiting and letting them time out, it would fight them.
It might look at locusts and think "if I fight them I am guaranteed to win". Since every time Alphastar fights locusts, no matter how bad the fight is going it magically wins the fight after long enough. I think it would make sense that Alphastar hasn't figured out that it can also beat locusts by doing nothing. It's reinforcement learning. Every locust wave convinces it more and more than the chance Alphastar wins (i.e. kills all the locusts) is 100%.
@@andrenowaczek4909 On the other hand, in the early stages of learning, the AI has a lot of random permutations to be more "curious". It would have a fight kiting locusts and figure out that this overall meant not losing any units or HP on them. Either this was so rare it never really happened, or it values the postioning on the map over the lost health.
@@DrZaius3141 Or it knows that 90+ probes can harvest enough to rebuild faster than the swarmhosts, anyway overal it looked quite efficient.
Hey do you know what's going on with Alphastar? Why are we not having any news since 4 years?
Remember how some pro players would flex winning against 1 v 7 hardest ai? In short future we will see how ai would be destroying 7 humans opponents alone. Time to get that sweet revenge!
Creating a neural network that plays SC2 is amazing.
Having that AI show human mistakes is unbelievable
I wonder what race would be the most effective from an AI with or without restrictions. I'm thinking Terran, if every marine can be controlled simultaneously and kite mostly everything.
Great commentary.
can't wait for next commentary
Alphastar didn't commit with the oracles ti'll the 7min. mark when he got 4 oracles to slaughter more drones and a queen :O
This was not Serral.
"22 actions per minute per 5 seconds" lol XD
hahaha that nova on his own army got me laughing so hard :D up to that point i thought it was doing so great.
I just got my logitech Artemis spectrum wireless head phones and your commentary sounds extra vivid. Unless you got a new mike. Love the channel
Tools have always been our true strength. Not how well we use the tools. Only makes sense that a tool we make to use our tools would do it better than us.
They should do a match without any restrictions on AS..... just to see what it could actually do when unleashed.
The first version of alphastar is basically what you described. However, it was only playing protoss at the time
We will know alphastar has made it when it throws out the offensive GG
Just read the articles about alphastar
Who reads in 2019 lul
@@speedfastman is this irony or do you mean it ?
@@popkhorne5372 its irony he had to read in order to know what to write
@@popkhorne5372 I would bet a dollar he is serious. The people that do not read in 2019 are the same ones that did not read in any other time period.
Unfortunately, illiteracy is a huge issue for all of humanity right now as we speak.
I would like to see Alpha Star without restriction.... How insane it is.
Only vs Serral, the man without restrictions
That would be stupid. Look up Automaton 2000 videos to see how utterly broken machine-controlled micro can be - you don't need any of Alphastar's strategy learning to beat a human player.
In the Blizzard forums, people trash Alphastar because it was unable to beat any of the major pro players. But since it had something between 95% and 99% win ratio, they are actually thrashing most of the SC2 community, including themselves, without realizing it.
They always say how flawed the bot is, or how easly it could be countered if this or that was done, as if they were worthy of teaching the AI how to play.
1:23 22? did you mean 220?
If he hadn't lost his swarm hosts at that point, he'd have won. He'd just killed 40 probes and the army count had been even.
You think it's disturbing to have a machine which learnins like a human, but that's what its supposed to be. Scientists are trying to reproduce the learning process, the human type of a learning process. There's still a huge difference between being able to learn like a human and being a human like being. First of all, choosing the goals. The non obvious will to win a game is hardcoded into AlphaStar, it didn't come up with that. On top of that, human needs a couple of games to get a grasp of whats going on - AlphaStar needs hundreds of games to get it. Still nothing to worry about. It's still just a set of optimization techniques.
Can't wait for alphastar to pick up on BM and rage quit
Imagine Alphastar learning from IdrA lmao
We need more Probes.
This was a surprisingly close match!
If those swarm hosts hadn't gotten caught at the end, we could have had an even longer game o:
I wonder who that zerg player was...
So when Alphago becomes champion in the overwatch, they can make general from starcraft and soldier from Overwatch. An AI army. Terminator.
AlphaGo only plays Go.
Want to play 3v3 Broodwars Fast Map with 3 of these things.
Idk if it's just me... but anyone else interested to see what alphastar could do when the take off the human restrictions winter mentioned? Just curious.. would alphastar be able to make a bazillion apm or display a terrifying level of control? What would someone do against say a grandmaster or higher level terran alphastar that could control large numbers of individual marines or groups of army at the same time? Would it be able to narrow its focus to multiple spots on the map at once all while microing workers or training units? A multi pronged attack using multiple types of units.. it would steamroll pro players. Lol would be cool
Like your style)
damn it, this is the same one as BaseTrades cast.
that 9:55 made me fall off my chair!
Interesting video
Ok but let's see an AI come up with "no mo-acles, stop it, you're bing bor-icle" hahaha
disruptors are broken...you can't give protoss siege units that can move in tandem with their other army units, has shields, and shoots like reavers
So alpha star is a low apm protoss player? Did we check whether white ra is human?
Winter is a robot confirmed
Wondering why they only taught alpha star how to play the easiest race? I've only ever seen it play toss. Zerg would be fun to see how it decides to play...
More alpha star!
Is zergs 4th off center?
Okay yeah it is wtf
9:54 and that's why i love Winter ,his not even trying to humor us but still i crackle alot on his streams exactly for this 9:54
Biggest dude in starcraft we love u Winter
OH MY GOD HE SHOT HIMSELF IN THE FACE!
je know the finale form, not yet the finale perahps perahps perhaps more of finale finale form
Who won? Whose next? YOU DECIDE? Seriously who won? Did you forget to edit the last part of the video?
Terrifyingly, unlike the previous super-micro-stalker, I cannot tell what makes the Protoss so good in this game. It looks like if Serral main race is Protoss :/
Very strong macro never missing warpins constant production the micro and army control isnt great but it dosent need great micro when your even or ahead on supply vs swarmhosts
What if that disruptor mistake was intentional to bait the zerg player to attack?
Let’s teach an AI how to create and control a killer robot army. This is fine.
The game wasn't anywhere near over when the gm quit.
Please, protoss was still at four bases, and hosts were being killed. Game over.
ugh, Protoss on 4 bases and Zerg on 2 after 15 minutes, u dumb lol.
90 probes by the 9 minute mark, wtf???
Final form protoss VS Final form zerg?
Gogogo
OMG HE SHOT HIMSELF IN THE FACE!
5:27 rap god
I think you got a probelem winter :)
AI is very dangerous, like nuclear weapons. I'm terrified... :/
The next step: Skynet... et the end of humanity... :'(
The program has perfect memory can do instant math on damage and can count all your units and never forgets. Still not quite an even match but impressive nonetheless that it learns so quickly.
This is not how Alphastar works at all. There is no calculating damage or any kind of unit parameter calculations. It takes in raw pixel data from the games camera and makes decisions based on the information. It's only as good as the amount of games it has experienced and learned from.
This is not even my final form!
Bilzzcon better have beastyQT fight alpha star with mass stalkers
Your harping on about apm, but have you stopped and considered ALL OF ALPHA STARS APM IS E-apm. IT'S IMMPOSIBLE for humans to have that high of an EFFECTIVE-apm consitantly.
22 action per minute per 5 seconds ...
Its explained better in the article. 22 Agent actions per 5 seconds where an agent action is not necessarily an action that counts towards a single in game APM.
IMO this is not Serral.
Alphastar looks to be making quite a few human errors. It looks very convincingly human.
Also - Swarm hosts are bullshit.
AlphaStar made a wall? I call fake, it's been Neeb playing this game
9:50
I have no idea WTF is happening (O)_(O)