You asked for it, here we go! AlphaStar vs MaNa the 5 game showmatch. I didn't end up adding the game with AlphaStar vs MaNa that MaNa ended up winning. The reason for that is there isn't much to say. The AI got confused by the Warp Prism movement from MaNa and that bought him enough time to build up an army and push across the map and defeat AlphaStar. One thing that I didn't know/didn't mention at the time is next. AlphaStar for the first 5 matches had overview of the WHOLE map, the AI didn't see through fog of war..but it could see whole map at once without having to move the screen. In the rematch that it played vs MaNa live it had normal screen where it had to move and played visibly worse because of it. Hopefully one day I can face AlphaStar as the highest ranked Random player in the world :P Would be a lot of fun!
When i began playing sc/broadwar and then sc 2.i always keep wondering why only 16 probe at minerals and not more,because i did compare the gas and mineral and saw that in gas, like you said ,there was always a probe waiting when its fully saturated wich 3 and the're now lost of gas and at mineral that didn't happen..so when i play i was going always 19 or 20 probe because i thought that like gas,you wouldt mine more minerals but always hear people saying your doing it wrong it need to be 16 and not worth it..i'm not playing sc now but enjoy your vid and when i saw this one i was like..finaly,proof that what i was doing wasn't wrong.
BeastyqtSC2 have to tell you something very very important. Maybe the AI do better economy than us but 89% of why the Ai won is because it had and unrealistic micro. If they address that, and the AI plays with human micro then the strategy have to change and therefore the whole game change and style.
I just figured something out -- In the Q&A one of the Deepmind guys said _"We keep old versions of each agent as competitors in the AlphaStar League. The current agents typically play against these competitors in proportion to the opponents' win-rate. This is very successful at preventing catastrophic forgetting, since _*_the agent must continue to be able to beat all previous versions of itself._*_ "_ So, the reason "Agent Troll" builds on the low ground and sends his zealot off at 1:12:34 is to counter previous versions of himself. Note that: he sends his zealot to where he builds his proxy robo at the same timing he places it.
"Agen Troll" is supposed to be in the top 5 agents tho. they said they selecct what they think are the op 5 agents so it must win against the others AI somehow
4:25 deep has 2 stalkers in production 5 probes on gas 4:40 queues 3rd stalker 4:45 queues 4th stalker 4:49 last probe, halts production... prioritizes mineral income for 3rd & 4th stalker and 28th probe, then puts 6th probe into gas 21 probes on minerals, 6 on gas, 1 probe for proxy gates. 2 stalkers, 2 more stalkers in production. builds 1 more probe 5:52 proxy 3rd gateway 6:11 proxy 4th gateway 6:59 proxy probe builds pylon 7:21 queues stalker 7:24 queues stalker 7:30 looses 4 probes.. Zero probes were replaced, Zero fucks were given.. 5 probes in gas
trust me scII community, as someone who had my Go community blown open by deepmind only years earlier, this is only the beginning. the ai will get so good that no humans will be able to even play a competitive game against it... after a couple of exhibition shows against top players of each race and proving that their AI is the number 1 player in every matchup, they will take the ai and go and you won't see it again. on the bright side, deepmind will probably release batches of alpha v alpha games for us to review and blow our minds. :)
Oh? why they put so much work into it and then simply leave? they should have it around for the public to play against/practice if wish ??. I'm only diamond 1-2 tbh but i guarantee you id beat this ai, id abuse it because the multiple complexities of the game.. where the whole board is not shown (as in GO) multiplies the number of moves 1000 fold even that of GO . The ai will never win for this game. And as state at the start, it is a great ai never the less.. hope they KEEP it around for practice.
I think all games were won due to micro. I hoped that AI will challange human in strategical thinking, and thats not the case here. But they did a good job and I'm hopefull for the future.
@@SamuelGrguric haha yeah fair enough, oh man! everyone just wants a crack at this deepmind. But hey, this just ONE race!, i think it will take them a long time as well to configure it to zerg and terran, and then to make it learn how to fight different races. Long long road ahead to be considered complete.
C'mon, AlphaGo brought a revolution in the go theory and enabled any go player access to superhuman evaluation with free software. I'm not a startcraft player but I, as profane, would expect that this ai will bring radical new ideas scII community
Something I notice about the final phoenix battle at 27:00 - they pick up far fewer stalkers at a time than they could have, and fewer than I think a human would. This leads me to believe it is better to maximize the phoenix-to-lifted-stalker damage than to lift them all and risk not killing many before they drop.
I wonder if in game 3, the AI is willing to risk being supply blocked because it feels a fight is imminent from happening and would prefer that minerals to be available for unit production during the fight.
The EPM of human players is not 100% beneficial -- the actions aren't doing nothing, but they don't necessarily do something helpful. With the AI, almost every action is beneficial. So this experiment is actually WAAAAY more tilted in the AI's favor than it appears. In other words, a 120 EPM AlphaStar would have more beneficial actions per minute than the best pro.
APM = clicks + commands EPM = commands Therefore if a player selects a unit and right-clicks somewhere 5 times, that would be 6 APM and 5 EPM (and one truly effective action). Also I believe it's somewhat incorrect to say that the AI has 100% effective actions, for it was initially developed based on mimicking actual 1v1 replays. I believe the developers said it spammed a bit.
I dont understand why the community always thought that 2 workers/patch was the optimum way to go. There is an old teamliquid post by PiousFlea which clearly show these different stats and one can easily see that there is an advantage on saturating each patch with 3 workers. It does not take long to get the money back from what you spend on the workers, but sure a lot can happen during that period. www.teamliquid.net/forum/sc2-strategy/140055-scientifically-measuring-mining-speed
I dont play SC2 anymore but in WoL it was common knowledge that 24 workers per base is best. At least until you have a lot of bases and you need army supply. I remember that when i saw HoTS trailers and that there is this counter of workers and it says x/16 i thought it is really stupid and people will think that 16 is best because of it.
We know that you get more money. Our preference was the have 16 workes which mine more effectifly and get a faster base. Alphastar gets more workers but build the expansion later.
True. This is really where learning ai really shines; it does not rely on rules, but everything it does is contextual. (Sometimes it will not over saturate probes). Humans will create another problem when they see this and think "I should oversaturate my probes". No, this is a contextual question. Humans use thinking shortcuts to master complex /dangerous situations, whereas ai needn't do that. It's something we are designed to do in high pressure / high stakes situations because it aided our survivability, but it's where a lot of the mistakes creep in.. sometimes really consequential ones in real life situations.
AlphaStar only played protoss vs protoss I think. It decides to spend resources to mine more minerals. But it is more vulnerable to early attacks. Its strategy might evolves when it starts playing zergs (or as you mentioned a Tank Rush). We'll see if this strategy persists.
thank you for showing finally ALL the games how could they ever dare to not show them this is freaking interesting. very nice thoughts of you i just wanted to add 34:43 : when the ai was buidling the expansion and then stargate it made sense since overdroning will instantly give you more income the moment the expansion finishes which can be spend through additional production building. i always thought phoenixes are great xD (but of course with this insane micro its maybe op)
I think that the apm cap on the AI is intentionally not capped, this way now that they have an AI that can beat pro gamers with brute micro they can train another AI that will be limited to 200apm or something against the insane one, and that is how we will get innovative smart plays they can't train an AI against pro players from the get go, because no one has 200 years to spare. Also keep in mind that the AI skill progression is not linear it's more like exponential, first 500 000 games it probably learned how to make pylons and where to make them and thats it.
Beastyqt, Florencio human last defense against AI. Best of Hollywood Starcraft 2 movie starring Beastyqt, scene 06:00 21:06 39:37 and ofc For an Oscar Nominee scene at.. 1:13:57 Enjoy.
One thing I'm curious about is at the very start of the games how the IA spread the probes. Since it has 1500 eAPM, it can probably send each probe to individual mineral patch; another thing that human can't do.
I thought move click on to the ground 5 times count as both apm and epm, but those clicks that are undoable(not enough mineral,not enough gas, you can't place there) count as apm but not epm
That war really interesting to think about. There are points where the AI knows where it is going to win and can just push, it seems to pick up cues and make better decisions.
they should not change anything , the ai is showing us the right way to play starcraft, alpha star kills you in 1 base by over saturating , while we out here not investing on probes
Any person, at least a little familiar with ML, will feel what tremendous work has been done here. This is the decision making of the AI, this is his tactical "thinking", this analysis of data on limited data. Guys from DeepMind deserved a lot of respect. On the other hand, what we have: AI, that: 1) learned about 200 (!!!) years how to play a gem 2) have an instant reaction 3) APM that is more than 1000, lost to a man as soon as he haven't ability to review the entire map (or just playing without fenix). So the players are still far away. Moreover, the second factor. People are learning too. And they do it many times faster. I bet that played with a dozen more games, Mana or even TLO, will understood how to play against this AI, and wins games without any problems
About the part where you said that TLO and Mana learns faster, you are both right and wrong. You are right in that the humans learn more for every game played beacouse they are better att analysing stratagies and their opponents but the AI learn faster if we use the unit of time. If the average SC2 gae is 15 min a human can play 96 games a day if they don't eat, sleep or go to the toilet. In that time the AI will have played (rough estimate) 1000 games. You are basicly saying that humans are better at chess becouse they can ignore all the obviusly bad moves and are ignoring the fact that the AI plain simply have the computing power to analyse every possible move no mather how stupid in seconds.
The early pylon in the last game gives cheap scouting and encourages MaNa to build zealots and fail to scout his natural/third. Plus mind games. And yeah, 1200 EPM is superhuman, therefor cheating. Should Cap it at Serral levels (don't know the fastest Protoss offhand).
@@leolampret804 and how, imagine if we in future can learn humans fast as neural networks. For example will invent some king of neural interface that allow to learn ten thousand times faster. I think projects like this can help with that in furure. If AI alone so strong, imagine what can do human + AI hybrid
21:48 deep lifts 2 stalkers with phoenix drops them immediately, then focus fires phoenix on manas phoenix. the lift interrupts the stalkers focus fire, they were targeting deeps phoenix.. when the stalkers start firing again, they are targeting deeps 1 and only zealot, deeps phoenix are basically uncontested. 22:00 beastys facial expression change..priceless
Where is the repetition of this game? I want to watch only what alphastar see. Camera alphastar all time@t I can't open maps in starcraft 2, I put maps on folder but I can't open it. I believe is about ancient version map. Can I upgrade to current version?
One way to even things up for playing against AI would be including basic micro stuff as selectable unit behavior. Stuff like "stutterstep retreat" or "blink at x% health" or "keep healthiest unit forward".
Some of the mineral patches are farther away from the nexus than others. (It's like that on every map). Those patches would need another worker to maximize mining. The additional worker doesn't add very many minerals per minute, but it's something. So you would need about 16+4=20 workers for maximum mining, and then the extra workers are for rebuilding after harass or to transfer to the expand as soon as it is done.
Maybe they'll let people adjust the difficulty with apm. "Just beat alphastar at 100apm" "lol I can only do it at 50" "maybe someday I can beat it at 300 like beasty"
Micro aside, what the AI seems to do is the same thing it usually does in TBS, generates early on a lot of low tier units and rush with a stronger army. Players sometimes tend to economy or research as priority. But what the AI does is just generate army, more army, lots of army, and then push. It's zerging.
Just a correction, they didnt program this. it is a neuralnetwork wich learns. So every decision the AI made is based on previous experinces combined with the situation it finds it self in. This is accully a bit scarier then most people realize. This is only the beginning of AI and its so impressive. The good part is that humans will be able to learn from the ai, at least early on. once it becomes crazy smart then we wount even know what we are looking at anymore. :) I loved these games so cool! hope they do more. But would like them to invite Korean players :)
well they needed to programm it somewhat. like rewards and what the AI can see and the controlls etc. but how it plays isnt programmed. aswell as it was feeded games from actuall pro games. i dont have 100% facts if thats true but all the programes that invold 0 human help were named something with "Zero" (AlphaGoZero, AlphaZero) and maybe they do another one wich will be "AlphaStarZero".
If you know your opponent is over making workers. Wouldn't you skip the harsh and go for a quicker expand? Because the opponent wouldn't be able to contest? Not really a pro or even Diamond player but just wanted to say. :P
The canon rush thing is the one thing I was wondering about... IF someone would tell me I am about to play SC2 against an learning A.I I would NEVER test my micro against it... Preferably I also wont test my macro against A.I.... I would check how cheese-proof it is ;) Honestly we all know how good A.I CAN be... It strenghts are clearly related to inhuman speed... Lets test it on a human level, who can be the biggest troll? ;)
TLO btw had APM from him that spiked to around 1800 APM. Obviously not EPM, and likely not as effective as what the AI is actually doing with its APM, but just want that there for context.
Even though I consider AlphaStar to be to one of the most wonderful piece of software that has ever been written, guys from Deepmind don't have achieved their goal of beating humans "the right way", as they say. AlphaStar is superior to human at micro, multitasking and optimisation, it reacts faster, and it probably have a better estimation of the situation. But all these abilities are typical computer skills. AlphaStar has shown pretty poor strategies skills, inspired by human replays. What I wanna see now is less "computer" skills and real strategic capabilities. Also, the final goal should be ONE agent, able to play multiple strategies and adapting to the player it is playing with, like humans do.
you can literally stick together all the agents and that's one agent. It wouldn't matter. Also much like with AlphaGo and AlphaZero, it starts largely as imitation learning and slowly transitions towards learning the game from scratch and throwing away any conceptions and strategies that it picked up from humans.
@@saurabhjhanjee2408 the reason why the agents go down 1 strategy is because they have overtrained to remove any other possible strategy. 'Strategy' is a bad term.. all it knows is given this current state, doing is the best chance of being rewarded with a win.
Jeff Taylor you could literally add the networks together and add an input node (or a dimension to the CNN) which indicates which agent to use. They would just be subnetworks.
Sticking together the AIs would be ugly from a theoretical and practical point of view. Training multiple agents should be a first step, you don't want to split TMU resources to train multiple agents when you can focus on one. Also you don't want a system that would have to randomly pick an agent from a list BEFORE the game starts, because human players would find a way to identify which agent they are facing early in the game and counter it. The goal to achieve is one unique, more complex neural network, able to apply multiple strategies. An ideal situation would be the AlphaStar league electing such a dominant agent, but I'm afraid it will keep being a Rock-paper-scissors situation, with poorly evolved strategies, like we've seen. Also we don't know if it will be possible to make a from-scratch learning considering the infinite gap of complexity between SC2 and Go and Chess.
Youpla unlike Go and Chess the agent for StarCraft 2 is largely based off imitation learning and hence the agents are in very suboptimal local minima, so the best next step would be to create an architecture which would reduce the state space sufficiently so that a reinforcement learning agent can be trained from scratch, or at least with minimal influence from the performance of humans (maybe only use it as data for off policy learning?)
In co-op, the nexus label suggests 3 workers per patch, and almost every player I see blindly follows that. I always put 2 per close patch and 3 per far patch, and force workers to go to their correct patches until none are running back and forth. However, I do think you gain slightly more income with fully 3 workers per patch. With Tychus, who doesn't lose supply for going overboard with scvs, sometimes I oversaturate beyond 3/patch and I think that gives even slightly more minerals, because some patches are too far away to be perfectly mined even by 3 workers.
quick tip: if u want to watch many old replays, u may want to consider closing the game but keeping blizzard app running and then starting the replay file from windows . this way you dont have to type your password and u can save urself a few seconds, and also if u'r a streamer u can protecc ur password more (it becomes somewhat easier to hack your password if the number of characters is known)
I like to compare this apm situation to chess. In chess you basically do one move per "frame" of the game, and very importantly, you can not make no move. While in starcraft you basically have the decision to do something during 1 frame which would be 1/60th of a second or not to do anything(because you are too slow). The game still goes on and does not wait for you. So if the AI has twice the amount of EPM than the humam it can do twice as much. While in chess we dont allow it to move twice in a row🤔
True but you can't make your units mine faster or shoot faster with apm, 1zealot with 10000apm will never beat an immortal. Not saying it doesn't matter of course, but it's not as strong as free moves in chess. I could probably beat a chess grandmaster if I had double moves. I can't see myself beating a sc2 grandmaster even if they played only with their off-hand or had a broken computer which dropped half their input
To people comparing the 50 ms vs 350 ms times: They were different things. The 50 ms was the time to process the situation and make a decision. The 350 ms was reaction time.
I am missing the last 10 minutes but what has not been discussed is also the quality of the actions. It is not only 4 times faster than a human but also all actions are on point. Even in the heat of the battle it perfectly picks up the right stuff wit Phoenixes, turns one Phoenix for 1 shot to finalise the low health prism, has all cool downs of units on point. A human even if fast will make more mistakes on top.
It seems to me that in order to get the benefit of over-probing you'd need to be constantly shuffling your probes around to keep the extras on the far patches. This is something a human might have more trouble with than an AI.
The difference between this and how the casters/people from deepmind look at it is because Beasty can afford to be wrong, he can afford to be more critical and honest about what he sees. Deep mind can beat human players, no doubt. But if you wanted to get a REAL AI, you would cap it at 150 APM during the entire game. This is because you can play sc2 at that level with 150, and the games that are won are won on strategy, the games that are lost are lost on strategy, build orders and timings. (or at the very minimum allow apm spikes to be as short as human are). If you play VERY FAST (and efficient and that is where a computer is miles ahead of a human), you can make up for glaring mistakes that would cost you the game over time. wait.... from 16 to 21 workers. Thats 5 workers. Thats 250 minerals. its an additional 100 minerals per minute. It pays for itself in 2 minutes 30 seconds. (include the build time of a nexus and the added extra mining time to get the extra 150 minerals.... where as the probes increase the effectiveness from the first probe... 12 (minus chronoboost) seconds in... that is 19 minerals per minute more, each 12 seconds, so by the time you make the 5th probe you already mined an additional .... ? Lets LOOK at the time when you finish 16 probes and the time when you finish 21 probes. IF in that time you CANNOT make a nexus due to the nature of the match up OR due to how you need to produce.... holy shit. I gotta test this in broodwar. I know its different there, but I wonder how much. Its actually more than 100 minerals, closer to 200. EDIT: Yeah right... Dat micro..... Holy Fucking meatballs... I want to see agent 5 vs agent 1 :D.
here is the thing about epm: we know the ai can micro faster. as beasty said, ai could probably run over to the human opponent with probes at the beginning of the game and win with micro. deepmind is focused on decision-making, so winning with 1500apm and blink stalkers doesnt proof the ai is better than mana in decision-making. it could be, because its very close anyways. good job deepmind.
Imagine AlphaStar in archon mode, with one agent responsible for perfect micro and the other for perfect macro. With AI-speed of communication between them. That would be scary to watch.
last game - making a stargate makes the enemy come out... and you were saying those things were in shooting range, but mana thought so too and moved units there, they couldn't do anything. *shrug* good show.
this AI vs Human on Starcraft reminds me of a Korean Drama about someone who can travel to the future and watch the AI beat the best SC player in the world 3 - 0
The reason why A.I would over probe and all in when it has a small lead is because it is so confident with its micro knowing it has high rate of winning. Beast is right, avoid or minimize micro engagements, just cheese or confuse A.I with dumb strats like multi prong harass, cannon rush, force field baits etc.
That Stargate is what made him win, though. What if he knew Immortals would come? And then a Warp Prism. And what if he knew Mana could juggle Immortals and kill him? As it turns out, the Phoenix was a good counter to the Warp Prism cause Mana had nothing to shoot up.
I think the AI won only by a surprise effect. With enough experience MaNa should have figured out how to counter it even without knowing which one he is playing against. It's the same with regular AI's. You start playing, Hard AI beats your ass, then you notice some flaws in it and abuse them. Also maybe the devs have put some limitations to MaNa's strategies... like "no cheese"
@@ra6865 It still has the programmed part of what signals it can recognize. The player needs to figure out which ones it can not understand and abuse it.
@@tft_heart i think it does recognize, but if it hasnt had the exact experience it might not know how to accuratly respond. But it will respond. As we saw in the game when Mana won. But in that game they added a "camera" funktion for the ai so it became a bit more fair.
MaNA could have cheesed if he wanted to. im pretty ysure.. i dont know an facts bt in the rematch game he did abuse the AI as the AI didnt knew how to handle the "transporter"? ii dont play SC2 so my apologies if it isnt correct.
@@joi1794 True but u can clearly see that in the end it was adapting by keeping an eye on the prism. It did not react the optimal way though. it did stop the harras. U could also see the Ai was confused, it tried to blink to an area that it could not blink to. Try looking at a human play starcraft that have never played and you will see the same reactions. Watch bronze league heros :)
You can ignore the haters who want to all be critics because I think it's an interesting topic and I agree with you with the mistakes that Mana made. Mana was playing as if he was playing a human. Your viewers that made nasty comments did not realize that like you did. wish I had an AI like this to daytrade my account for me lol. Since that's what I do and I just watch sc2 casually. If I could teach a bot to trade for 200 years and than daytrade the markets I bet I'd be wealthy from it lol. I have a person who codes that I can show your video too :)
That dude who said "no logic AI" He didn't even saw the live stream about alphastar right? Cause the developers said that is a machine that makes it own decisions... like, wtf dude, imagine that, a fucking machine that just goes "welp, i can fuck u up right now" and thats what this is
It's a shame they overtrained the AI so that they had to have 5 different agents... they could overcome this by having each agent remember it's previous games, and reduce the expected reward it would receive based on how many times it opened the same way
that one cheats and query the raw game input to see which zergling is being targetted and then create a "force-field" around the targetted zergling to repell other zerglings away from it
There are two ways to poke this bear. Have an AI that cannot peak above the human in EPM that may effectively only control what the screen is showing, then it will have to adopt creative strategies to get ahead, this is where you can see its creativity. Can it out-think a human and not merely out-perform. At this point it operates with effectively an over sized mini-map with fog of war. Then when its mastered all races, take the gloves off and allow it full reign of its capabilities and play 1 v 7 pro-gamers. Would it be that good? Sure, why not. Essentially it wouldn't loose a single engagement because of insane micro of 16k effective EPM. By the way, I watched the cast games closely and every single time it was always ahead in minerals / min, every game it had superior income, so progressively every objective came sooner. Oh, and Beasty, stop using the Lord's name, you should be bigger than that. Otherwise you're heaping judgment. Why, well you brought it up!
Kind of amusing: When I suck with my terran units if there is one HT or 1 baneling on the map and my whole, bigger army just dies to those splash units, "better" players keep telling me I am a noob because I know nothing about the game... and forget about my 50-60 APM. But once a pro gets beaten by a faster AI it is all about APM and EPM. Maybe - in the end the game is not as deep as many keep postulating and it is more about fast clicking than strategy after all.
1) make extra probes 2) mass stalkers 3) micro like a fucking god 4) ??? 5) profit! Clearly SC2 AIs will go the way of chess, into a league of their own separate from humans. It's an inevitability, given the different mechanics used by human vs AI (e.g. AI can click anywhere instantly while a human has to move their mouse first, AI never misclicks, etc) and the power of micro in mirror matchups. It will still be very, very interesting to watch and study however. Also of note is the use of this in balancing the game. Apparently the DeepMind team found no strategy that was strictly dominating over another in the PvP mirror, which suggests that PvP is well-balanced. When AIs expand to other matchups and races, this can also be used to balance the game. I predict that Zerg AIs will find an advantage simply due to the native imbalance of Zerg macro mechanics (lategame replenishing army in under 1 minute which T can't do and P can only do if having invested in like 20+ gates), which is only counterbalanced in human play by the high APM requirement of Zerg (more expanding, creep spread, queen inject).
Up next: Google makes 10 ton AI tank and runs over human samurai and proclaims our AI is better at martial arts! A computer that is allowed to surpass all limitations of the game and do over 1000EPM isn't playing Starcraft; It's hacking, as beastyqtSC2 rightly points out.
Last game: makes one phoenix didnt do anything... um...it was THE KEY unit to that push killing or being held. Am i the only one saw it immediately shoo away the immortal juggling prism??? No juggle, no hold. Phoenix over VR b/c void is too slow and cannot lift an immortal if needed. Its not crap, its really thinking ahead; "whats his one big chance to hold this? A prism keeping immortals alive...if i stop that my immortals w multiple batteries cant loose."
You asked for it, here we go! AlphaStar vs MaNa the 5 game showmatch.
I didn't end up adding the game with AlphaStar vs MaNa that MaNa ended up winning. The reason for that is there isn't much to say. The AI got confused by the Warp Prism movement from MaNa and that bought him enough time to build up an army and push across the map and defeat AlphaStar.
One thing that I didn't know/didn't mention at the time is next. AlphaStar for the first 5 matches had overview of the WHOLE map, the AI didn't see through fog of war..but it could see whole map at once without having to move the screen. In the rematch that it played vs MaNa live it had normal screen where it had to move and played visibly worse because of it.
Hopefully one day I can face AlphaStar as the highest ranked Random player in the world :P Would be a lot of fun!
i literally just posted a request for this a couple hours ago and already wish granted!!! Thank you so much you're the best SCII youtuber! ^_^
When i began playing sc/broadwar and then sc 2.i always keep wondering why only 16 probe at minerals and not more,because i did compare the gas and mineral and saw that in gas, like you said ,there was always a probe waiting when its fully saturated wich 3 and the're now lost of gas and at mineral that didn't happen..so when i play i was going always 19 or 20 probe because i thought that like gas,you wouldt mine more minerals but always hear people saying your doing it wrong it need to be 16 and not worth it..i'm not playing sc now but enjoy your vid and when i saw this one i was like..finaly,proof that what i was doing wasn't wrong.
Beatyqt vs agent troll, the real meaning of why it all exists.
TLO tried to cannon rush but never got a cannon up.
BeastyqtSC2 have to tell you something very very important. Maybe the AI do better economy than us but 89% of why the Ai won is because it had and unrealistic micro. If they address that, and the AI plays with human micro then the strategy have to change and therefore the whole game change and style.
I just figured something out -- In the Q&A one of the Deepmind guys said _"We keep old versions of each agent as competitors in the AlphaStar League. The current agents typically play against these competitors in proportion to the opponents' win-rate. This is very successful at preventing catastrophic forgetting, since _*_the agent must continue to be able to beat all previous versions of itself._*_ "_
So, the reason "Agent Troll" builds on the low ground and sends his zealot off at 1:12:34 is to counter previous versions of himself. Note that: he sends his zealot to where he builds his proxy robo at the same timing he places it.
"Agen Troll" is supposed to be in the top 5 agents tho. they said they selecct what they think are the op 5 agents so it must win against the others AI somehow
4:25 deep has 2 stalkers in production 5 probes on gas
4:40 queues 3rd stalker
4:45 queues 4th stalker
4:49 last probe, halts production...
prioritizes mineral income for 3rd & 4th stalker and 28th probe, then puts 6th probe into gas
21 probes on minerals, 6 on gas, 1 probe for proxy gates.
2 stalkers, 2 more stalkers in production.
builds 1 more probe
5:52 proxy 3rd gateway
6:11 proxy 4th gateway
6:59 proxy probe builds pylon
7:21 queues stalker
7:24 queues stalker
7:30 looses 4 probes.. Zero probes were replaced, Zero fucks were given..
5 probes in gas
trust me scII community, as someone who had my Go community blown open by deepmind only years earlier, this is only the beginning. the ai will get so good that no humans will be able to even play a competitive game against it... after a couple of exhibition shows against top players of each race and proving that their AI is the number 1 player in every matchup, they will take the ai and go and you won't see it again. on the bright side, deepmind will probably release batches of alpha v alpha games for us to review and blow our minds. :)
Oh? why they put so much work into it and then simply leave? they should have it around for the public to play against/practice if wish ??. I'm only diamond 1-2 tbh but i guarantee you id beat this ai, id abuse it because the multiple complexities of the game.. where the whole board is not shown (as in GO) multiplies the number of moves 1000 fold even that of GO . The ai will never win for this game. And as state at the start, it is a great ai never the less.. hope they KEEP it around for practice.
I think all games were won due to micro. I hoped that AI will challange human in strategical thinking, and thats not the case here. But they did a good job and I'm hopefull for the future.
@@Kyzik244 yeah we said the same thing about Go comparing to to Chess and other games. :P
@@SamuelGrguric haha yeah fair enough, oh man! everyone just wants a crack at this deepmind. But hey, this just ONE race!, i think it will take them a long time as well to configure it to zerg and terran, and then to make it learn how to fight different races. Long long road ahead to be considered complete.
C'mon, AlphaGo brought a revolution in the go theory and enabled any go player access to superhuman evaluation with free software.
I'm not a startcraft player but I, as profane, would expect that this ai will bring radical new ideas scII community
Very intelligent analysis. Appreciate it! This matchup was so much fun to watch.
this is extremely good analysis beasty. i hope deepmind invites u.
no such thing as extremx or good or invitex or not, cepu, say, think any nmw and any be perfect
39:05 the ai anticipated a units loss of a more capable player and with that the "supply block" is more efficient
That is a very intelligent possible explanation.
Something I notice about the final phoenix battle at 27:00 - they pick up far fewer stalkers at a time than they could have, and fewer than I think a human would. This leads me to believe it is better to maximize the phoenix-to-lifted-stalker damage than to lift them all and risk not killing many before they drop.
I wonder if in game 3, the AI is willing to risk being supply blocked because it feels a fight is imminent from happening and would prefer that minerals to be available for unit production during the fight.
The EPM of human players is not 100% beneficial -- the actions aren't doing nothing, but they don't necessarily do something helpful. With the AI, almost every action is beneficial. So this experiment is actually WAAAAY more tilted in the AI's favor than it appears.
In other words, a 120 EPM AlphaStar would have more beneficial actions per minute than the best pro.
Exactly. There needs to be a better way to measure EPM. Like Optimal Effective Actions Per Minute or something.
APM = clicks + commands
EPM = commands
Therefore if a player selects a unit and right-clicks somewhere 5 times, that would be 6 APM and 5 EPM (and one truly effective action).
Also I believe it's somewhat incorrect to say that the AI has 100% effective actions, for it was initially developed based on mimicking actual 1v1 replays. I believe the developers said it spammed a bit.
I dont understand why the community always thought that 2 workers/patch was the optimum way to go. There is an old teamliquid post by PiousFlea which clearly show these different stats and one can easily see that there is an advantage on saturating each patch with 3 workers. It does not take long to get the money back from what you spend on the workers, but sure a lot can happen during that period.
www.teamliquid.net/forum/sc2-strategy/140055-scientifically-measuring-mining-speed
I dont play SC2 anymore but in WoL it was common knowledge that 24 workers per base is best. At least until you have a lot of bases and you need army supply. I remember that when i saw HoTS trailers and that there is this counter of workers and it says x/16 i thought it is really stupid and people will think that 16 is best because of it.
We know that you get more money. Our preference was the have 16 workes which mine more effectifly and get a faster base. Alphastar gets more workers but build the expansion later.
True. This is really where learning ai really shines; it does not rely on rules, but everything it does is contextual. (Sometimes it will not over saturate probes). Humans will create another problem when they see this and think "I should oversaturate my probes". No, this is a contextual question. Humans use thinking shortcuts to master complex /dangerous situations, whereas ai needn't do that. It's something we are designed to do in high pressure / high stakes situations because it aided our survivability, but it's where a lot of the mistakes creep in.. sometimes really consequential ones in real life situations.
AlphaStar only played protoss vs protoss I think. It decides to spend resources to mine more minerals. But it is more vulnerable to early attacks. Its strategy might evolves when it starts playing zergs (or as you mentioned a Tank Rush). We'll see if this strategy persists.
thank you for showing finally ALL the games how could they ever dare to not show them this is freaking interesting. very nice thoughts of you i just wanted to add 34:43 : when the ai was buidling the expansion and then stargate it made sense since overdroning will instantly give you more income the moment the expansion finishes which can be spend through additional production building. i always thought phoenixes are great xD (but of course with this insane micro its maybe op)
I think that the apm cap on the AI is intentionally not capped, this way now that they have an AI that can beat pro gamers with brute micro they can train another AI that will be limited to 200apm or something against the insane one, and that is how we will get innovative smart plays they can't train an AI against pro players from the get go, because no one has 200 years to spare. Also keep in mind that the AI skill progression is not linear it's more like exponential, first 500 000 games it probably learned how to make pylons and where to make them and thats it.
Not really exponential, there are diminishing returns after some point the ai skill will plateau
Very intelligent analysis. I truly enjoyed that. This a historic moment in human existence.
Beastyqt, Florencio human last defense against AI.
Best of Hollywood Starcraft 2 movie starring Beastyqt, scene
06:00
21:06
39:37
and ofc For an Oscar Nominee scene at..
1:13:57
Enjoy.
Hi Beastyqt, Do you have an Alphastar analysis playlist? Thanks!
where can I watch alphastar replay?
One thing I'm curious about is at the very start of the games how the IA spread the probes. Since it has 1500 eAPM, it can probably send each probe to individual mineral patch; another thing that human can't do.
where did you get that shiba doggo pillow
I thought move click on to the ground 5 times count as both apm and epm, but those clicks that are undoable(not enough mineral,not enough gas, you can't place there) count as apm but not epm
That war really interesting to think about. There are points where the AI knows where it is going to win and can just push, it seems to pick up cues and make better decisions.
That agent troll must have won so many games using that technique if he was picked.
The fifth agent ist inspired by the one and only Sewer Mermaid, aka. Florencio
they should not change anything , the ai is showing us the right way to play starcraft, alpha star kills you in 1 base by over saturating , while we out here not investing on probes
Any person, at least a little familiar with ML, will feel what tremendous work has been done here. This is the decision making of the AI, this is his tactical "thinking", this analysis of data on limited data. Guys from DeepMind deserved a lot of respect.
On the other hand, what we have: AI, that: 1) learned about 200 (!!!) years how to play a gem 2) have an instant reaction 3) APM that is more than 1000, lost to a man as soon as he haven't ability to review the entire map (or just playing without fenix).
So the players are still far away.
Moreover, the second factor. People are learning too. And they do it many times faster. I bet that played with a dozen more games, Mana or even TLO, will understood how to play against this AI, and wins games without any problems
About the part where you said that TLO and Mana learns faster, you are both right and wrong.
You are right in that the humans learn more for every game played beacouse they are better att analysing stratagies and their opponents but the AI learn faster if we use the unit of time.
If the average SC2 gae is 15 min a human can play 96 games a day if they don't eat, sleep or go to the toilet. In that time the AI will have played (rough estimate) 1000 games.
You are basicly saying that humans are better at chess becouse they can ignore all the obviusly bad moves and are ignoring the fact that the AI plain simply have the computing power to analyse every possible move no mather how stupid in seconds.
Last game was Florencio Neeb in Archon mode
The early pylon in the last game gives cheap scouting and encourages MaNa to build zealots and fail to scout his natural/third. Plus mind games. And yeah, 1200 EPM is superhuman, therefor cheating. Should Cap it at Serral levels (don't know the fastest Protoss offhand).
@@leolampret804 and how, imagine if we in future can learn humans fast as neural networks. For example will invent some king of neural interface that allow to learn ten thousand times faster. I think projects like this can help with that in furure. If AI alone so strong, imagine what can do human + AI hybrid
21:48 deep lifts 2 stalkers with phoenix drops them immediately, then focus fires phoenix on manas phoenix.
the lift interrupts the stalkers focus fire, they were targeting deeps phoenix..
when the stalkers start firing again, they are targeting deeps 1 and only zealot, deeps phoenix are basically uncontested.
22:00 beastys facial expression change..priceless
The madman did it, thanks fam I've been excited for you analysis.
Where is the repetition of this game? I want to watch only what alphastar see. Camera alphastar all time@t
I can't open maps in starcraft 2, I put maps on folder but I can't open it. I believe is about ancient version map. Can I upgrade to current version?
One way to even things up for playing against AI would be including basic micro stuff as selectable unit behavior. Stuff like "stutterstep retreat" or "blink at x% health" or "keep healthiest unit forward".
Thanks for posting great content
Especially in early game more workers have always been worth it. You perfectly described the situations for that :)
Just enjoy the beauty of the 3-front attack. It''s pretty cool that it could figure out that strategy even if it had high apm.
this is an interesting breakthrough! 21 workers! not even starcraft devs catched that one..
Some of the mineral patches are farther away from the nexus than others. (It's like that on every map). Those patches would need another worker to maximize mining. The additional worker doesn't add very many minerals per minute, but it's something. So you would need about 16+4=20 workers for maximum mining, and then the extra workers are for rebuilding after harass or to transfer to the expand as soon as it is done.
Maybe they'll let people adjust the difficulty with apm. "Just beat alphastar at 100apm" "lol I can only do it at 50" "maybe someday I can beat it at 300 like beasty"
Micro aside, what the AI seems to do is the same thing it usually does in TBS, generates early on a lot of low tier units and rush with a stronger army. Players sometimes tend to economy or research as priority. But what the AI does is just generate army, more army, lots of army, and then push. It's zerging.
Did anyone notice how much alpha star plays like florencio
1:14:30 Agent Wimp Lo. They purposely trained it wrong, as a joke.
Researcher 1: It doesn't even converge.
R2: Fuck it, we're doing this live
Just a correction, they didnt program this. it is a neuralnetwork wich learns. So every decision the AI made is based on previous experinces combined with the situation it finds it self in. This is accully a bit scarier then most people realize. This is only the beginning of AI and its so impressive. The good part is that humans will be able to learn from the ai, at least early on. once it becomes crazy smart then we wount even know what we are looking at anymore. :) I loved these games so cool! hope they do more. But would like them to invite Korean players :)
well they needed to programm it somewhat. like rewards and what the AI can see and the controlls etc. but how it plays isnt programmed. aswell as it was feeded games from actuall pro games. i dont have 100% facts if thats true but all the programes that invold 0 human help were named something with "Zero" (AlphaGoZero, AlphaZero) and maybe they do another one wich will be "AlphaStarZero".
This means we need a 1v7 extreme extreme challenge with you and the ai now.
If you know your opponent is over making workers. Wouldn't you skip the harsh and go for a quicker expand? Because the opponent wouldn't be able to contest? Not really a pro or even Diamond player but just wanted to say. :P
The canon rush thing is the one thing I was wondering about... IF someone would tell me I am about to play SC2 against an learning A.I I would NEVER test my micro against it... Preferably I also wont test my macro against A.I....
I would check how cheese-proof it is ;)
Honestly we all know how good A.I CAN be... It strenghts are clearly related to inhuman speed... Lets test it on a human level, who can be the biggest troll? ;)
mana did troll it well with the prisim last game
great video!
"I'd cannon rush that motherfucker". ~ Beasty in Terminator 15
Bob built the Stargate as a hedge against something of course. It has lost more before when it doesn't do this.
Great video. Really interesting.
TLO btw had APM from him that spiked to around 1800 APM. Obviously not EPM, and likely not as effective as what the AI is actually doing with its APM, but just want that there for context.
So are we going to see @beastyq vs 7 alpha star
Even though I consider AlphaStar to be to one of the most wonderful piece of software that has ever been written, guys from Deepmind don't have achieved their goal of beating humans "the right way", as they say.
AlphaStar is superior to human at micro, multitasking and optimisation, it reacts faster, and it probably have a better estimation of the situation. But all these abilities are typical computer skills. AlphaStar has shown pretty poor strategies skills, inspired by human replays. What I wanna see now is less "computer" skills and real strategic capabilities.
Also, the final goal should be ONE agent, able to play multiple strategies and adapting to the player it is playing with, like humans do.
you can literally stick together all the agents and that's one agent. It wouldn't matter. Also much like with AlphaGo and AlphaZero, it starts largely as imitation learning and slowly transitions towards learning the game from scratch and throwing away any conceptions and strategies that it picked up from humans.
@@saurabhjhanjee2408 the reason why the agents go down 1 strategy is because they have overtrained to remove any other possible strategy. 'Strategy' is a bad term.. all it knows is given this current state, doing is the best chance of being rewarded with a win.
Jeff Taylor you could literally add the networks together and add an input node (or a dimension to the CNN) which indicates which agent to use. They would just be subnetworks.
Sticking together the AIs would be ugly from a theoretical and practical point of view. Training multiple agents should be a first step, you don't want to split TMU resources to train multiple agents when you can focus on one. Also you don't want a system that would have to randomly pick an agent from a list BEFORE the game starts, because human players would find a way to identify which agent they are facing early in the game and counter it. The goal to achieve is one unique, more complex neural network, able to apply multiple strategies. An ideal situation would be the AlphaStar league electing such a dominant agent, but I'm afraid it will keep being a Rock-paper-scissors situation, with poorly evolved strategies, like we've seen.
Also we don't know if it will be possible to make a from-scratch learning considering the infinite gap of complexity between SC2 and Go and Chess.
Youpla unlike Go and Chess the agent for StarCraft 2 is largely based off imitation learning and hence the agents are in very suboptimal local minima, so the best next step would be to create an architecture which would reduce the state space sufficiently so that a reinforcement learning agent can be trained from scratch, or at least with minimal influence from the performance of humans (maybe only use it as data for off policy learning?)
In co-op, the nexus label suggests 3 workers per patch, and almost every player I see blindly follows that. I always put 2 per close patch and 3 per far patch, and force workers to go to their correct patches until none are running back and forth. However, I do think you gain slightly more income with fully 3 workers per patch. With Tychus, who doesn't lose supply for going overboard with scvs, sometimes I oversaturate beyond 3/patch and I think that gives even slightly more minerals, because some patches are too far away to be perfectly mined even by 3 workers.
quick tip: if u want to watch many old replays, u may want to consider closing the game but keeping blizzard app running and then starting the replay file from windows . this way you dont have to type your password and u can save urself a few seconds, and also if u'r a streamer u can protecc ur password more (it becomes somewhat easier to hack your password if the number of characters is known)
I like to compare this apm situation to chess. In chess you basically do one move per "frame" of the game, and very importantly, you can not make no move. While in starcraft you basically have the decision to do something during 1 frame which would be 1/60th of a second or not to do anything(because you are too slow). The game still goes on and does not wait for you.
So if the AI has twice the amount of EPM than the humam it can do twice as much. While in chess we dont allow it to move twice in a row🤔
True but you can't make your units mine faster or shoot faster with apm, 1zealot with 10000apm will never beat an immortal. Not saying it doesn't matter of course, but it's not as strong as free moves in chess.
I could probably beat a chess grandmaster if I had double moves. I can't see myself beating a sc2 grandmaster even if they played only with their off-hand or had a broken computer which dropped half their input
To people comparing the 50 ms vs 350 ms times:
They were different things. The 50 ms was the time to process the situation and make a decision. The 350 ms was reaction time.
I am missing the last 10 minutes but what has not been discussed is also the quality of the actions. It is not only 4 times faster than a human but also all actions are on point. Even in the heat of the battle it perfectly picks up the right stuff wit Phoenixes, turns one Phoenix for 1 shot to finalise the low health prism, has all cool downs of units on point. A human even if fast will make more mistakes on top.
The ai csnnot control army though. It can only control individual unit so it need to have higher apm
It seems to me that in order to get the benefit of over-probing you'd need to be constantly shuffling your probes around to keep the extras on the far patches. This is something a human might have more trouble with than an AI.
i liked when you would get triggered and have your bits of anger, reminds me of myself
The difference between this and how the casters/people from deepmind look at it is because Beasty can afford to be wrong, he can afford to be more critical and honest about what he sees.
Deep mind can beat human players, no doubt. But if you wanted to get a REAL AI, you would cap it at 150 APM during the entire game. This is because you can play sc2 at that level with 150, and the games that are won are won on strategy, the games that are lost are lost on strategy, build orders and timings. (or at the very minimum allow apm spikes to be as short as human are).
If you play VERY FAST (and efficient and that is where a computer is miles ahead of a human), you can make up for glaring mistakes that would cost you the game over time.
wait.... from 16 to 21 workers. Thats 5 workers. Thats 250 minerals. its an additional 100 minerals per minute. It pays for itself in 2 minutes 30 seconds. (include the build time of a nexus and the added extra mining time to get the extra 150 minerals.... where as the probes increase the effectiveness from the first probe... 12 (minus chronoboost) seconds in... that is 19 minerals per minute more, each 12 seconds, so by the time you make the 5th probe you already mined an additional .... ?
Lets LOOK at the time when you finish 16 probes and the time when you finish 21 probes.
IF in that time you CANNOT make a nexus due to the nature of the match up OR due to how you need to produce.... holy shit. I gotta test this in broodwar. I know its different there, but I wonder how much.
Its actually more than 100 minerals, closer to 200.
EDIT: Yeah right... Dat micro..... Holy Fucking meatballs...
I want to see agent 5 vs agent 1 :D.
1:13:45 , the ai knows what it is doing beasty, don't need to backseat drive xd.
here is the thing about epm: we know the ai can micro faster. as beasty said, ai could probably run over to the human opponent with probes at the beginning of the game and win with micro. deepmind is focused on decision-making, so winning with 1500apm and blink stalkers doesnt proof the ai is better than mana in decision-making. it could be, because its very close anyways. good job deepmind.
I want to see the TAS version of SC2 match vs pro
Imagine AlphaStar in archon mode, with one agent responsible for perfect micro and the other for perfect macro. With AI-speed of communication between them. That would be scary to watch.
last game - making a stargate makes the enemy come out...
and you were saying those things were in shooting range, but mana thought so too and moved units there, they couldn't do anything. *shrug* good show.
this AI vs Human on Starcraft reminds me of a Korean Drama about someone who can travel to the future and watch the AI beat the best SC player in the world 3 - 0
Last AI is literally BeastyqtSC2! :)
Am I the only one who would like to see 2 uninhibited AIs duke it out?
they would need to limit Top APM not only Avarage APM
The reason why A.I would over probe and all in when it has a small lead is because it is so confident with its micro knowing it has high rate of winning. Beast is right, avoid or minimize micro engagements, just cheese or confuse A.I with dumb strats like multi prong harass, cannon rush, force field baits etc.
Also all people have like 350 effective response time and 100-200 ms reaction time
Beasty is telling the truth. The EPM peak is out of control.
Even Blizzard don't know the maximum mineral and gaz maximum mining numbers ! We have to trust maths only !
The objective was never to beat the computer but to demostrate if an AI can play and win against a pro player
That Stargate is what made him win, though. What if he knew Immortals would come? And then a Warp Prism. And what if he knew Mana could juggle Immortals and kill him? As it turns out, the Phoenix was a good counter to the Warp Prism cause Mana had nothing to shoot up.
I think the AI won only by a surprise effect. With enough experience MaNa should have figured out how to counter it even without knowing which one he is playing against.
It's the same with regular AI's. You start playing, Hard AI beats your ass, then you notice some flaws in it and abuse them.
Also maybe the devs have put some limitations to MaNa's strategies... like "no cheese"
This is not like a normal bot. The bots in starcraft 2 is all programed. these ais have been learning on their own, and can adapt.
@@ra6865 It still has the programmed part of what signals it can recognize. The player needs to figure out which ones it can not understand and abuse it.
@@tft_heart i think it does recognize, but if it hasnt had the exact experience it might not know how to accuratly respond. But it will respond. As we saw in the game when Mana won. But in that game they added a "camera" funktion for the ai so it became a bit more fair.
MaNA could have cheesed if he wanted to. im pretty ysure.. i dont know an facts bt in the rematch game he did abuse the AI as the AI didnt knew how to handle the "transporter"? ii dont play SC2 so my apologies if it isnt correct.
@@joi1794 True but u can clearly see that in the end it was adapting by keeping an eye on the prism. It did not react the optimal way though. it did stop the harras. U could also see the Ai was confused, it tried to blink to an area that it could not blink to. Try looking at a human play starcraft that have never played and you will see the same reactions. Watch bronze league heros :)
Love this video
Why you don't fighting against Deepmind
You can ignore the haters who want to all be critics because I think it's an interesting topic and I agree with you with the mistakes that Mana made. Mana was playing as if he was playing a human. Your viewers that made nasty comments did not realize that like you did. wish I had an AI like this to daytrade my account for me lol. Since that's what I do and I just watch sc2 casually. If I could teach a bot to trade for 200 years and than daytrade the markets I bet I'd be wealthy from it lol. I have a person who codes that I can show your video too :)
Maybe the future is 2 human commander-CoOp (controlling a single army) vs 1 AI.
That dude who said "no logic AI"
He didn't even saw the live stream about alphastar right? Cause the developers said that is a machine that makes it own decisions... like, wtf dude, imagine that, a fucking machine that just goes "welp, i can fuck u up right now" and thats what this is
"but all AI can see the whole map though"...i saw this statement a million times...
So can every friggin player, we call it *t.h.e MINIMAP".
So we need a graph of winning percent in each game for both player !
BeastyQT "I wanna crush it like a bug."
He did it later when he got the chance to play Alphastar. 😃
I want to see more of agent troll
so deep mind will basically define the meta
I want to see u play
The point of the IA (and the marvelous evolution of software) is to 'think' as human, and NOT to execute the commands as humans.
It's a shame they overtrained the AI so that they had to have 5 different agents... they could overcome this by having each agent remember it's previous games, and reduce the expected reward it would receive based on how many times it opened the same way
There are literally micro ai where zerglings dodge siegethank shots. But with like 14 k apm
that one cheats and query the raw game input to see which zergling is being targetted and then create a "force-field" around the targetted zergling to repell other zerglings away from it
There are two ways to poke this bear. Have an AI that cannot peak above the human in EPM that may effectively only control what the screen is showing, then it will have to adopt creative strategies to get ahead, this is where you can see its creativity. Can it out-think a human and not merely out-perform. At this point it operates with effectively an over sized mini-map with fog of war.
Then when its mastered all races, take the gloves off and allow it full reign of its capabilities and play 1 v 7 pro-gamers. Would it be that good? Sure, why not. Essentially it wouldn't loose a single engagement because of insane micro of 16k effective EPM. By the way, I watched the cast games closely and every single time it was always ahead in minerals / min, every game it had superior income, so progressively every objective came sooner.
Oh, and Beasty, stop using the Lord's name, you should be bigger than that. Otherwise you're heaping judgment. Why, well you brought it up!
Kind of amusing: When I suck with my terran units if there is one HT or 1 baneling on the map and my whole, bigger army just dies to those splash units, "better" players keep telling me I am a noob because I know nothing about the game... and forget about my 50-60 APM. But once a pro gets beaten by a faster AI it is all about APM and EPM. Maybe - in the end the game is not as deep as many keep postulating and it is more about fast clicking than strategy after all.
The last game was Epic !!!!!!!!
Think to point out its NOT live pach its preety old one
1) make extra probes
2) mass stalkers
3) micro like a fucking god
4) ???
5) profit!
Clearly SC2 AIs will go the way of chess, into a league of their own separate from humans. It's an inevitability, given the different mechanics used by human vs AI (e.g. AI can click anywhere instantly while a human has to move their mouse first, AI never misclicks, etc) and the power of micro in mirror matchups. It will still be very, very interesting to watch and study however.
Also of note is the use of this in balancing the game. Apparently the DeepMind team found no strategy that was strictly dominating over another in the PvP mirror, which suggests that PvP is well-balanced. When AIs expand to other matchups and races, this can also be used to balance the game. I predict that Zerg AIs will find an advantage simply due to the native imbalance of Zerg macro mechanics (lategame replenishing army in under 1 minute which T can't do and P can only do if having invested in like 20+ gates), which is only counterbalanced in human play by the high APM requirement of Zerg (more expanding, creep spread, queen inject).
Up next: Google makes 10 ton AI tank and runs over human samurai and proclaims our AI is better at martial arts! A computer that is allowed to surpass all limitations of the game and do over 1000EPM isn't playing Starcraft; It's hacking, as beastyqtSC2 rightly points out.
The way you speak and comment, it reminds me of day9
Beasty.. Isn't dark number 1 random player in the world?🙈🙈
Dark actually switched to random for a while. He is like top 100 Korea terram asswell.
Last game: makes one phoenix didnt do anything... um...it was THE KEY unit to that push killing or being held.
Am i the only one saw it immediately shoo away the immortal juggling prism??? No juggle, no hold. Phoenix over VR b/c void is too slow and cannot lift an immortal if needed. Its not crap, its really thinking ahead; "whats his one big chance to hold this? A prism keeping immortals alive...if i stop that my immortals w multiple batteries cant loose."
Random Broken for AI? :D