EPTBoxeR: "AlphaGo won't beat humans in StarCraft" - Page 14
| Forum Index > SC2 General |
|
Tuczniak
1561 Posts
| ||
|
suid
11 Posts
On March 14 2016 18:26 papaz wrote: Rofl @humans and their pride. Why does some many people think humans have some kind of magic that can't be replicated by computers? Is it because people are religious and think we have some kind of "soul" or something along those lines that makes us unique? We are just advanced biological machines as a result of evolution. We will of course sooner or later develop AI and computers that outshines us in every single way. It's just a matter of time. What argument is there that this won't happen? It's just a matter of science and enough time. Science fighting! I agree that, in general, humans overestimate their position in the animal kingdom, but there are non-religious, professional philosophers and cognitive scientists who believe that human consciousness is an emergent phenomena from impossibly complicated circuits at the neurobiological level that they believe cannot be completely comprehended, let alone reproduced, by theoretical algorithms. Dennett says this consciousness is an illusion, which I typically feel is probably an accurate statement of what most people think of. Any species with an endocrine system, a sufficiently complex endbrain, and a few other structures/connections probably experiences some feeling of "sentience" or "consciousness." But, again, I think those words probably don't actually mean anything scientifically. It's very easy to discuss the components of consciousness without actually even realizing it's consciousness being discussed. Sure, humans are "just advanced biological machines as a result of evolution." How long did that process take again? 3.5 billion years? You're view of the complexity of the human species seems very degraded. The theoretical/computational neuroscientists don't really even have a falsifiable theory of the brain yet. It's a gigantic pile of anatomical details and physiological details. There are still many, many experiments to be done, and that alone can take extremely long. And, fwiw, the "Turing test" is an idiotic metric of machine intelligence. A couple areas that machines will have a very difficult time "outshining" human expertise or biology are energy conservation (metabolism), regardless of species; least for a very, very, very long time. Evolution optimized that process very well. That's an opinion I just formed while writing this, maybe someone else has actually studied it. My view of the SC2 AI is that this could already happen today very easily. I don't know why anyone cares about Boxer's opinion. | ||
|
Asbury
2 Posts
| ||
|
Ciaus237
South Africa288 Posts
| ||
|
guN-viCe
United States687 Posts
The AI will lose in it's current state but if the team persists for a year or two they will get to pro-gamer level and beyond. How does one go about writing a bot for SC and what language is optimal? Is the bot designed to be like a human and play from the top down? Or is it somehow tied to the SC files in a deep rooted fashion? | ||
|
danielias
Chile67 Posts
| ||
|
JimmyJRaynor
Canada17682 Posts
On March 14 2016 23:58 Tuczniak wrote: I think the most interesting set of limitations will be the one that produce the best strategies that could be theoretically used by humans. But of course I would like to see AI vs AI without limitations, even if it's irrelevant to the game we play. interesting idea but the game would have to be re-balanced from the ground up. in SC2 as an example, the relationship between marines and banelings is very different in AI v. AI at unlmited APM. there are dozens of as yet undiscovered micro-techniques at 3,000 APM that some day AIs may help humans discover. we do not know much about Starcraft at 3,000 APM. we do not know what we don't know. | ||
|
Makro
France16890 Posts
On March 15 2016 01:47 danielias wrote: Wow...Spechless. Robots will take control of everything in the future, hospitals, highways, buildings, food, etc. In our lifes era when we get old we´ll see that. Maybe in 50 years from now. i would say 150 years, 50 years not enough | ||
|
danielias
Chile67 Posts
Maybe, but i think we will see some of this in our lifes. In 150 years robots will definitly have a bigger rol in humanity. Michiu Kaku teach this stuff. I think this DeepMind is a huge thing, like A.C and B.C. the same will hapen. Before self learning robot and before self learning robot. Jobs will be very different in 100 or 150 years. | ||
|
Tenks
United States3104 Posts
| ||
|
beheamoth
44 Posts
| ||
|
TheZov
Russian Federation34 Posts
| ||
|
ZAiNs
United Kingdom6525 Posts
On March 15 2016 03:29 TheZov wrote: Oh come on guys, what are we talking about here? The AI of tomorrow is not the Deep Blue archetype that (I might point out) DID beat a professional chess player at his own game. Dynamic thinking machines that have unlimited APM and unlimited ability to process, calculate, analyze and mimic every play in the history of the game (both pro and amateur, including every ladder game ever played in the history of the game), on a micro-second basis, selecting situational solutions in real time with perfect execution and an unlimited potential for multi-prong execution... Does that sound like something anyone else can do? First of all, it's well agreed upon that if the AI exploits infinite APM it will win easily. What's in question is whether an AI will be able to beat a human any time soon, when the AI is subject to physical limitations similar to that of a human. Also your understanding of AI right now isn't very accurate. The AI won't have 'analyzed every play in the history of the game including every pro, amateur and ladder match', it couldn't for Go because there were simply too many game states, and StarCraft has several magnitudes more different game states. Brute force worked for chess but not for Go which needed a neural network AI to 'solve'. For AlphaGo, they fed the AI around 150,000 pro-games (there's no way they will be able to feed a StarCraft AI anywhere close to that, and even if they could I'm pretty sure StarCraft would required waaaay more games than Go to get it to the same place). AlphaGo then played against itself for months using a crazy amount of computing power (which amounts to a completely insane number of games). In StarCraft, every tiny change in any unit's positioning is a different game-state (or even the position of the mouse cursor), so it'll constantly be playing game-states that it's never played before. You also say the AI can perform this super complex analysis 'on a micro-second basis' which isn't the case at all. AlphaGo does not make decisions instantaneously, AlphaGo and Lee Sedol in their ongoing match are both given 2 hours of thinking time per game, and Go is a game with around 200 discrete moves, if you consider every action made in StarCraft there are probably tens of thousands of discrete moves from both sides. Making the AI able to make decisions in 1/60th of a second is going to be a huge challenge, and there is no way it can perform the same level of analysis in 1/60th of a second as it could with even 1 second. | ||
|
bluQ
Germany1724 Posts
On March 15 2016 00:04 suid wrote: I agree that, in general, humans overestimate their position in the animal kingdom, but there are non-religious, professional philosophers and cognitive scientists who believe that human consciousness is an emergent phenomena from impossibly complicated circuits at the neurobiological level that they believe cannot be completely comprehended, let alone reproduced, by theoretical algorithms. Dennett says this consciousness is an illusion, which I typically feel is probably an accurate statement of what most people think of. Any species with an endocrine system, a sufficiently complex endbrain, and a few other structures/connections probably experiences some feeling of "sentience" or "consciousness." But, again, I think those words probably don't actually mean anything scientifically. It's very easy to discuss the components of consciousness without actually even realizing it's consciousness being discussed. Sure, humans are "just advanced biological machines as a result of evolution." How long did that process take again? 3.5 billion years? You're view of the complexity of the human species seems very degraded. The theoretical/computational neuroscientists don't really even have a falsifiable theory of the brain yet. It's a gigantic pile of anatomical details and physiological details. There are still many, many experiments to be done, and that alone can take extremely long. And, fwiw, the "Turing test" is an idiotic metric of machine intelligence. A couple areas that machines will have a very difficult time "outshining" human expertise or biology are energy conservation (metabolism), regardless of species; least for a very, very, very long time. Evolution optimized that process very well. That's an opinion I just formed while writing this, maybe someone else has actually studied it. My view of the SC2 AI is that this could already happen today very easily. I don't know why anyone cares about Boxer's opinion. Let me help you with your argument: They estimate that simulation of the whole human brain would require supercomputer with about 500 petabytes of memory. Current record in one system is 1.5 petabytes (Sequoia supercomputer). So we need a system over 300 times larger. Such machines are not expected in this decade. Human Brain Project expects that such machines will be available around 2023. Source: here People oversimplify the problem. Without a quantum computer no AI will outshine a humans brain. And one simple fact most are missing anyways: An AI designed to ONLY be superb at a certain thing just can't be compared to a human who dedicates their brain-power to a diversity of things (social life, sports, school, studies etc. etc.). To speak of "superior" AIs is the real illusion here. Science hwaiting. | ||
|
L_Master
United States8017 Posts
On March 13 2016 03:55 [PkF] Wire wrote: The problem is how do you define human micro (and even human multitask). A simple limit on the APM wouldn't even be enough I think, since the computer doesn't spam and -more importantly- sees all screens at once. It would still make it reasonably fair. If you capped the computer at say, 180-200apm, that would be pretty on par with the number of useful actions taken by progamers. Yea, the computer could certainly be more optimal in its use of those given actions, but it would absolutely shut down almost all the obscene micro possibilities with individual muta micro, instant perfect marine splits vs lurkers, perfectly pulling back all units in battles at once, etc. If the computer learns well enough, I think it will still have a significant control advantage over human players, but probably won't be able to execute game breaking micro. | ||
|
Xyik
Canada728 Posts
Asides from perfect micro which is imo is not that interesting, i don't think an A.I can ever consistently beat a top pro. Like Boxer and Flash have said there are far more variables to SC than Go. 1. It would be difficult to train / test the A.I. Unlike Go, where an A.I could probably simulate 100 moves a second to learn optimal openings, an A.I wouldn't be able to simulate SC games at the same speed. The best it could do is parse out replay-level data (build orders, clicks, etc. - and I'm not sure how detailed that data is) and play real games (and obv the avg SC game length is around ~10-20 minutes, allowing a max of playing ~200 games a day). 2. In general, there is less data on SC that the A.I could learn from, and more importantly, the Google developer team working on the A.I could learn from. While the A.I uses unsupervised learning, the developers on the A.I themselves would still need to know what type of data to feed it, and what type of data is important. They would probably need to recruit a couple high-level SC players to get a sense of whats important in SC, vs what is important in Go. For example, in Go the A.I knows to optimize to net stones. What would it optimize for in SC? 3. There is a factor of luck. An A.I that plays by probability can miss the tiniest things. An easy example that comes to mind is scouting. The scouting path taken can make a huge difference, if some proxy cheese build is found 30 seconds late it could mean the entire game. 4. The number of maps, start positions and race match-up will make the number of situations it has to learn in explode exponentially. Unless they decide to only train the A.I on a select number of maps and match-ups, but then once again there is the problem of having even less data to learn from in the first place. 5. SC is a game of taking risk, making trade-offs with limited information. This makes it very difficult to consistently win games, which is why the state of the game today pretty extremely difficult for any player to have > 70% win rates overall. | ||
|
trulojucreathrma.com
United States327 Posts
On March 15 2016 04:51 Xyik wrote: Seems like most people commenting here aren't aware that many attempts have been made to make bots in BW, in fact there were competitions to do so, and to my knowledge there has never been one to rival a pro level player (of course, there was no big company like Google working on the bots). There have been some people messing around, just for fun, that's correct. Doesn't mean it is easy, but there has been no serious attempt. 3. There is a factor of luck. An A.I that plays by probability can miss the tiniest things. An easy example that comes to mind is scouting. The scouting path taken can make a huge difference, if some proxy cheese build is found 30 seconds late it could mean the entire game. The AI will on average be just as lucky as the human player. I agree though that a very big skill difference is needed in Starcraft to get a significant higher win-rate. | ||
|
JimmyJRaynor
Canada17682 Posts
http://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp/report2013.shtml i believe the university of alberta has constructed a series Foundation Framework Classes for coders to create new AI. waterloo is still better though. just cause. | ||
|
trulojucreathrma.com
United States327 Posts
| ||
|
JimmyJRaynor
Canada17682 Posts
churchill is not messing around. also, when u use a generalized term like "student" you're not saying anything. it could be a 7 year old or it could be a 26 year old phd student defending his or her thesis. http://webdocs.cs.ualberta.ca/~cdavid/research.shtml http://webdocs.cs.ualberta.ca/~cdavid/publications.shtml http://webdocs.cs.ualberta.ca/~cdavid/pdf/combat13.pdf based on what i'm reading the first AI to consistently defeat pro players using human APM limits will probably have to employ at least 1 top level SC player. "As with many game search applications, state spaces are often too large to search completely, so heuristics must be employed to evaluate non-terminal states. In traditional games such as Checkers or Chess these heuristic functions often depend on expertly crafted formula based on intuitive notions such as game positioning, or material counts." they'll need some top level players for that. | ||
| ||
