| Forum Index > SC2 General |
|
Deleted User 3420
24492 Posts
On January 25 2019 22:23 Ej_ wrote: Its not. The point is to make AI that can compete with humans execution... Where are you getting that from? You can read DeepMind's mission statement here: https://deepmind.com/about/ The APM limitation was just to make it more fair. If they wanted to make it imitate human execution they would need to use a robotic hand, (and i guess a camera, too). | ||
Poopi
France12906 Posts
On January 25 2019 22:17 travis wrote: lol @ people complaining about an AI doing the things an AI can do yes it's not fair, that's the point, that's why an AI can exceed humans. duh It's not a battle of intelligence if you just outplay hard your opponent mechanically. Them trying to make it a bit fair is because they understand otherwise it would not mean much. So far they have made good progress compared to previous gaming AI but it's still far from robust, as we have seen. | ||
|
mishimaBeef
Canada2259 Posts
| ||
|
Haukinger
Germany131 Posts
On January 25 2019 22:40 Poopi wrote: It's not a battle of intelligence if you just outplay hard your opponent mechanically. It's a battle of Starcraft. That a game not about intelligence, but mechanical execution. If else, there would be APM-limits in the game like minimal cooldowns on all commands. | ||
|
Acrofales
Spain18139 Posts
On January 25 2019 22:36 travis wrote: Where are you getting that from? You can read DeepMind's mission statement here: https://deepmind.com/about/ The APM limitation was just to make it more fair. If they wanted to make it imitate human execution they would need to use a robotic hand. That's kinda the point being made though. We can make a machine that can do surgery with far more precision than a human could ever hope to achieve, but we still need the medic controlling it, because the medic knows *what* to do, even if the robot is far better at actually doing it. Creating a bot that has absolutely perfect control over the units and beats humans based (mostly) on that control, is not "solving intelligence", it is using a superhuman unit control to win. Now I don't know what inputs AlphaStar receives, whether it's the videofeed and it needs to process that, or whether it's a list of everything that is happening in the game in some kind of symbolic format (e.g. an XML of all positions of all units it can see, all structures, what they are producing, how far along they are, etc. etc.). But clearly to make the competition "fair" in a real-time game, you cannot treat it in the same way as a turn-based game like chess or go: you need to factor in real-time limitations of humans, which includes things like screen lag and time to physically move an arm which moves the mouse, as well as the associated imprecision in both processes. Now if you say "computer vision is a hard problem that we don't want to deal with", and "robotics is a hard problem that we don't want to deal with", that's fine, and I agree that that is not really the point of a Starcraft bot. But then the simulation scenario should add in an approximately human level latency and add random noise to "clicks" that approximates human error. A start was made by removing the "zoom out hack", which I don't believe really should affect much, because superhuman minimap awareness + superhuman speed at clicking there and back on the minimap when a red blip appears is about the same, however in Mana's game with the DT, the superhuman map awareness almost certainly allowed AlphaStar to build an observer in time, whereas a human would *probably* only have noticed them by the time they were in his base, attacking shit. Consider the difference in reaction to DTs to the presence of an observer in the AI's base for the entire bloody game in the live showmatch without the zoom hack  Now I may sound negative about AlphaStar. I am not. I think it is mindblowing what they managed to achieve, and I think it's time I start looking at LSTMs myself to see if I can apply them (in a far more limited scale) to some of my research questions. I fear I don't have the data available to train them (my main critic of deep learning approaches in general). And I also wish I had a casual 16 tensorflow units standing by for me But AlphaStar was far ahead of what I was expecting from an SC2 bot. I really thought the presentation would be some very limited scenarios, not full game play level (even if only PvP and on a limited map pool). It is truly impressive what they accomplished in at most 2 years of work on this.I also really really liked their visualizations of what the network was "thinking". Very very cool stuff. | ||
|
deacon.frost
Czech Republic12129 Posts
On January 25 2019 22:56 Acrofales wrote: That's kinda the point being made though. We can make a machine that can do surgery with far more precision than a human could ever hope to achieve, but we still need the medic controlling it, because the medic knows *what* to do, even if the robot is far better at actually doing it. Creating a bot that has absolutely perfect control over the units and beats humans based (mostly) on that control, is not "solving intelligence", it is using a superhuman unit control to win. Now I don't know what inputs AlphaStar receives, whether it's the videofeed and it needs to process that, or whether it's a list of everything that is happening in the game in some kind of symbolic format (e.g. an XML of all positions of all units it can see, all structures, what they are producing, how far along they are, etc. etc.). But clearly to make the competition "fair" in a real-time game, you cannot treat it in the same way as a turn-based game like chess or go: you need to factor in real-time limitations of humans, which includes things like screen lag and time to physically move an arm which moves the mouse, as well as the associated imprecision in both processes. Now if you say "computer vision is a hard problem that we don't want to deal with", and "robotics is a hard problem that we don't want to deal with", that's fine, and I agree that that is not really the point of a Starcraft bot. But then the simulation scenario should add in an approximately human level latency and add random noise to "clicks" that approximates human error. A start was made by removing the "zoom out hack", which I don't believe really should affect much, because superhuman minimap awareness + superhuman speed at clicking there and back on the minimap when a red blip appears is about the same, however in Mana's game with the DT, the superhuman map awareness almost certainly allowed AlphaStar to build an observer in time, whereas a human would *probably* only have noticed them by the time they were in his base, attacking shit. Consider the difference in reaction to DTs to the presence of an observer in the AI's base for the entire bloody game in the live showmatch without the zoom hack Now I may sound negative about AlphaStar. I am not. I think it is mindblowing what they managed to achieve, and I think it's time I start looking at LSTMs myself to see if I can apply them (in a far more limited scale) to some of my research questions. I fear I don't have the data available to train them (my main critic of deep learning approaches in general). And I also wish I had a casual 16 tensorflow units standing by for me But AlphaStar was far ahead of what I was expecting from an SC2 bot. I really thought the presentation would be some very limited scenarios, not full game play level (even if only PvP and on a limited map pool). It is truly impressive what they accomplished in at most 2 years of work on this.I also really really liked their visualizations of what the network was "thinking". Very very cool stuff. I take it shortly 1) Don't project human emotions and ways of work on machines. That doesn't work, that's why they're machines and not humans(duh) 2) AlphaStar knew. YOur surgeon example is wrong, machines whcih make decisions based on machine learning experience know. It's not human knowing per se, it's machine knowing, but they know. 3) Because machines don't have emotions and they have bigger experience(also faster thinking) they can tell which fights are worth it and which aren't on more precise scale. Similarly a pro can tell this while you wouldn't be able to tell it on the same scale. And because they don't have the emotions they don't fear of losing, because they don't fear of losing they are doing humanly insane things(e.g. the ramp things). (just imagine what some pros would be able to do without the fear of losing the units which is always there, even if it's on the background with low priority(to use machine terms  ))To me all the threads about the games are full of big misunderstanding how such machine operates and projections of humanism on machines. | ||
|
Haukinger
Germany131 Posts
On January 25 2019 22:56 Acrofales wrote: Creating a bot that has absolutely perfect control over the units and beats humans based (mostly) on that control, is not "solving intelligence", it is using a superhuman unit control to win. That's like asking it to win with only Zealots... if it's in the game, the AI may use it. Only winning matters, if it wins by controlling stalkers individually with 10000apm, it's a win nevertheless. | ||
|
counting
11 Posts
On January 25 2019 22:56 Acrofales wrote: Now I don't know what inputs AlphaStar receives, whether it's the videofeed and it needs to process that, or whether it's a list of everything that is happening in the game in some kind of symbolic format (e.g. an XML of all positions of all units it can see, all structures, what they are producing, how far along they are, etc. etc.). But clearly to make the competition "fair" in a real-time game, you cannot treat it in the same way as a turn-based game like chess or go: you need to factor in real-time limitations of humans, which includes things like screen lag and time to physically move an arm which moves the mouse, as well as the associated imprecision in both processes. From their blog post and video showing the behind the scene videos, AlphaStar certainly used a modified version of PySC2 API as the input and output interface. deepmind.com There is a AMA thread on reddit MachineLearning subreddit If we really want to know the details, maybe we can get some details soon. On January 25 2019 22:56 Acrofales wrote: Now I may sound negative about AlphaStar. I am not. I think it is mindblowing what they managed to achieve, and I think it's time I start looking at LSTMs myself to see if I can apply them (in a far more limited scale) to some of my research questions. I fear I don't have the data available to train them (my main critic of deep learning approaches in general). And I also wish I had a casual 16 tensorflow units standing by for me But AlphaStar was far ahead of what I was expecting from an SC2 bot. I really thought the presentation would be some very limited scenarios, not full game play level (even if only PvP and on a limited map pool). It is truly impressive what they accomplished in at most 2 years of work on this.I also really really liked their visualizations of what the network was "thinking". Very very cool stuff. LSTM is interesting and powerful, but not a silver bullet, and it is famously hard to train (most RNNs have the same problem). It is essentially just a time series pattern recognition mechanism (can be used as a sequence generator). But it is just one component (albeit a crucial one), You will need many more components to solve complex problems. Like if you use an Action-Critic system, the LSTM will most likely be the action/sequence generator, and you still need a critic/evaluation network/system for it to function. | ||
|
Acrofales
Spain18139 Posts
On January 26 2019 00:26 Haukinger wrote: That's like asking it to win with only Zealots... if it's in the game, the AI may use it. Only winning matters, if it wins by controlling stalkers individually with 10000apm, it's a win nevertheless. Not really, I'm saying it should play SC2, not some heavily modified version that you can play by "plugging electrodes into your brain and thinking about exactly where you want each individual stalker to blink to and they do that instantly". SC2 is still a computer game for humans, which means you have to look at the screen, process the screen and then move the mouse and keyboard to perform actions. | ||
|
Acrofales
Spain18139 Posts
On January 26 2019 00:35 counting wrote: LSTM is interesting and powerful, but not a silver bullet, and it is famously hard to train (most RNNs have the same problem). It is essentially just a time series pattern recognition mechanism (can be used as a sequence generator). But it is just one component (albeit a crucial one), You will need many more components to solve complex problems. Like if you use an Action-Critic system, the LSTM will most likely be the action/sequence generator, and you still need a critic/evaluation network/system for it to function. I know. I just happen to be bumping into a rather hard time series pattern recognition problem. Unfortunately, my first problem is the dataset itself. It's small, and unlabelled at small scales: I have labels for hour-long sequences, and then I *definitely* don't have enough samples, and never will, so I need to look at smaller windows, and then they're unlabelled. So that'd mean labelling them manually, and I don't have time nor money at the moment to do so, but if they accept my newest grant proposal, I'll definitely consider this. We'll see  For now, I'll stick with manual feature engineering from the whole sequence and using random forests, which is giving decent results at a more macro level  | ||
|
Charoisaur
Germany16017 Posts
On January 26 2019 00:26 Haukinger wrote: That's like asking it to win with only Zealots... if it's in the game, the AI may use it. Only winning matters, if it wins by controlling stalkers individually with 10000apm, it's a win nevertheless. Of course if it does that it's technically better than humans at sc2. But that misses the point of this competition, the point of this is to show that the AI is smarter than humans, we already know that it's physically far superior. And comparing the strategic thinking/decision-making of humans vs AI can only be done if the AI's physical capabilities are limited. | ||
|
niteReloaded
Croatia5282 Posts
On January 25 2019 05:10 renaissanceMAN wrote: fucking bot LOVES stalkers holy cow haha :'D | ||
|
Polypoetes
20 Posts
On January 25 2019 22:40 Poopi wrote: It's not a battle of intelligence if you just outplay hard your opponent mechanically. Them trying to make it a bit fair is because they understand otherwise it would not mean much. So far they have made good progress compared to previous gaming AI but it's still far from robust, as we have seen. Weren't you the guy that claimed just a year ago that no AI/Bot/NN would be able to play Starcraft? Playing Starcraft isn't a test of intelligence. It is a test of playing and winning. I don't get how people expect a neural network to optimize winning at Starcraft to also have an eerie ability to (seemingly) read the minds of people. On January 26 2019 03:32 Charoisaur wrote: Of course if it does that it's technically better than humans at sc2. But that misses the point of this competition, the point of this is to show that the AI is smarter than humans, we already know that it's physically far superior. And comparing the strategic thinking/decision-making of humans vs AI can only be done if the AI's physical capabilities are limited. No. If you think an AI that plays Starcraft isn't 'interesting' to you because it turns out the best way to play SC2 is to mass stalkers and blink micro like crazy, then that means that you don't really like SC2 the way it is meant to be played. So you try a different more interesting game. Getting a finely tuned unit composition and moving across the map trying to outflank your opponent is objectively a stupid way to play if you can just mass stalkers, a1a2a3, micro like an AI, and win. But it seems people have some alternate axis of stupid-smart completely independent of winning-losing. Which is interesting, and very confusing, in itself. So is a bot that plays 'like a human' but plays weaker, smarter than a bot that plays 1-dimentionally, 'like an AI', but players very strong? BTW, we already know humans are smarter than AI because humans create and use AIs. AIs don't create humans and use them for their purposes. This is a silly line to even do down. | ||
|
Poopi
France12906 Posts
On January 27 2019 05:41 Polypoetes wrote: Weren't you the guy that claimed just a year ago that no AI/Bot/NN would be able to play Starcraft? Playing Starcraft isn't a test of intelligence. It is a test of playing and winning. I don't get how people expect a neural network to optimize winning at Starcraft to also have an eerie ability to (seemingly) read the minds of people. No. If you think an AI that plays Starcraft isn't 'interesting' to you because it turns out the best way to play SC2 is to mass stalkers and blink micro like crazy, then that means that you don't really like SC2 the way it is meant to be played. So you try a different more interesting game. Getting a finely tuned unit composition and moving across the map trying to outflank your opponent is objectively a stupid way to play if you can just mass stalkers, a1a2a3, micro like an AI, and win. But it seems people have some alternate axis of stupid-smart completely independent of winning-losing. Which is interesting, and very confusing, in itself. So is a bot that plays 'like a human' but plays weaker, smarter than a bot that plays 1-dimentionally, 'like an AI', but players very strong? I'm the guy that said it's basically impossible to have a completely fair fight between an AI and a human, because of the heavy emphasis on mechanics in starcraft 2, and that's hard to make things fair in that department. And the goal of deepmind is general AI so it's a small step towards that but far from enough, that's why fair things regarding mechanics is important if you want a robust and adaptable AI. | ||
|
Polypoetes
20 Posts
I don't know what to say to you that can convince you now that there is no such a thing, by definition, of a 'fair' competition between an AI and a human. What would that even entail? You simulate a human brain inside a computer so that you know that genetically/biochemically, such a human as the one you simulate could theoretically exist? An AI is an AI and a human is a human. The question is if we humans can create AI to do tasks. And games are nice benchmarks because they are well-defined problems. There is plenty of test data and it is easy to come up with a cost/objective function. What is next? People here arguing that an AI should miss-click like a human would? Rage and get 'emotional' after being cheesed? Get 'nervous' for important matches? Actually, those may be interesting AI challenges down the road for AI's that should be able to engage socially with humans better than humans are. But right now when the question is if AIs can beat humans in RTS games, that seems silly. I guess people who were convinced that AIs wouldn't be able to play RTS have to move the goalpost somewhere to keep their peace of mind. You want an SC2-based Turing test? As for SC2. It turns out mechanics is at the core of the game. Who would have thought! That is why many of us knew that AI would have good chances of taking games off humans. Exactly because many of these very human soft and subtle skills aren't that important. | ||
|
Charoisaur
Germany16017 Posts
On January 27 2019 05:41 Polypoetes wrote: Weren't you the guy that claimed just a year ago that no AI/Bot/NN would be able to play Starcraft? Playing Starcraft isn't a test of intelligence. It is a test of playing and winning. I don't get how people expect a neural network to optimize winning at Starcraft to also have an eerie ability to (seemingly) read the minds of people. No. If you think an AI that plays Starcraft isn't 'interesting' to you because it turns out the best way to play SC2 is to mass stalkers and blink micro like crazy, then that means that you don't really like SC2 the way it is meant to be played. So you try a different more interesting game. Getting a finely tuned unit composition and moving across the map trying to outflank your opponent is objectively a stupid way to play if you can just mass stalkers, a1a2a3, micro like an AI, and win. But it seems people have some alternate axis of stupid-smart completely independent of winning-losing. Which is interesting, and very confusing, in itself. So is a bot that plays 'like a human' but plays weaker, smarter than a bot that plays 1-dimentionally, 'like an AI', but players very strong? BTW, we already know humans are smarter than AI because humans create and use AIs. AIs don't create humans and use them for their purposes. This is a silly line to even do down. that only means massing blinkstalkers is the smartest way to play if you have the physical capabilities of an AI. It doesn't mean at all that the AI is better at strategizing because a human might do the same if it had the physical capabilities of an AI. If it would only be about beating humans no matter how there wouldn't be the need for the Deepmind team to work on it. Just a standard micro bot executing a basic strategy would be enough. Also a new 1 post user that memorizes what Poopi said 1 year ago................ Hmmmmmm | ||
|
The_Red_Viper
19533 Posts
On January 27 2019 06:22 Charoisaur wrote: that only means massing blinkstalkers is the smartest way to play if you have the physical capabilities of an AI. It doesn't mean at all that the AI is better at strategizing because a human might do the same if it had the physical capabilities of an AI. Also a new 1 post user that memorizes what Poopi said 1 year ago................ Hmmmmmm That is like saying AI's should only be able to calculate X moves in whatever timeframe for chess or go. People do not realize that the point isn't to have an AI which simulates human abilities in everything but Y, the point is to have an AI which is able to learn a task and be more efficient/better at it than the human counter part. Now it is somewhat interesting to bring down the AI's abilities in apm and whatever else you think is important mostly because of the PR and it being more challenging for the developers (that means they learn more about possible other uses), not because it has to be "fair". The interesting part is that you have an AI which learns new tasks without any hardcoded rules, that is what's fascinating. | ||
|
Polypoetes
20 Posts
But how is that not strategizing? The AI has AI mechanics. The proper strategizing therefore is to use blink stlaker with superior AI mechanics. Apparently, it is better at strategizing than you are, because you don't seem to get that. You act as if AIs could beat humans at RTS for years using just some build-in micro. Have you ever considered what kind of code is needed for an AI to outplay a human in a micro battle? And have you considered how hard it is to make an AI that doesn't get stuck or exploited easily by a human? Their strong AI has a 100% winrate vs these players. I don't know how people can be so stubborn and say that the AI doesn't 'think', doesn't 'known', isn't 'intelligent', doesn't 'strategize'. Those are all anthropomorphization of what machine learning is. You optimize it to win games. That's what they did. They did very well. And they have shown the potential to do it even better. In principle, you could also train a neural net to play indistinguishable from a human, passing a SC2 Turing test. But that requires a human agent. You cannot have 200 years of human lifetimes of humans objectively judging if an AI is more human-like than some other iteration. | ||
|
lolfail9001
Russian Federation40190 Posts
| ||
|
ETisME
12588 Posts
On January 27 2019 06:38 lolfail9001 wrote: I mean, i don't get the argument going on. The ultimate point is that AI did figure out a proper decision making in AI terms, but both this decision making was not very satisfying from human's PoV.... nor could be it be in conditions AI was granted. you summed it up perfectly. it's never about winning or losing, the goal isn't to win the game but to learn and solve complex situations with AI. The wins were with too much of mechanics advantage (raw input) and the loss was with player limited vision. | ||
| ||
|
|
Wardi Open
StarCraft2.fi
Replay Cast
The PondCast
OSC
Demi vs Mixu
Nicoract vs TBD
Babymarine vs MindelVK
ForJumy vs TBD
Shameless vs Percival
Replay Cast
Korean StarCraft League
CranKy Ducklings
SC Evo League
BSL 21
Sziky vs OyAji
Gypsy vs eOnzErG
[ Show More ] |
|
|
EPT
