|
United Kingdom13775 Posts
On November 08 2016 03:41 aQuaSC wrote:Show nested quote +On November 08 2016 03:32 LegalLord wrote:On November 08 2016 03:27 aQuaSC wrote:On November 08 2016 03:22 LegalLord wrote:On November 08 2016 03:14 aQuaSC wrote:On November 08 2016 03:06 LegalLord wrote:On November 08 2016 03:01 aQuaSC wrote:On November 08 2016 02:27 LegalLord wrote:On November 08 2016 02:22 Jett.Jack.Alvir wrote:
Well this would be a huge PR stunt for both Blizzard and Google. SC2 would get huge coverage, look at Go, and Google would push the boundaries of AI learning.
So Blizzard is allowing DeepMind to access SC2 api? I thought that DeepMind was looking at BW because there is no api restrictions. On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way On the other hand you may just hate Blizzard too much, why BW is better than SC2 for potential AI? Just because of the fact that the scene is now established? I think people at Deepmind are much more qualified to make such assumptions or comparisons, if they really thought BW would be better for it they would do it in BW. Or maybe evil Mike bought Google and their Deepmind team? I don't hate SC2 at all. I simply don't see their rationale, from an academic perspective, of actually doing SC2, precisely because they are throwing out years of progress by those who made AI before them. Among other things, there is a 24/7 AI test stream for BW and countless works by Berkeley and others, all done with BW. DeepMind can make their own decisions, but I do not believe they are made from an academically wise perspective. They are basically retreading old ground for no academically justifiable reason. I'd love to hear why is it wise to make the research behind each choice of games, and it's weird to me to think that the progress made in BW has to be largely scrapped and can't be moved over to a game of the same genre, but which has more complex and challenging choices on the example of macro mechanics which BW doesn't have, BW macro mechanic is ordering newly trained workers to mine which is not choice-based but mechanically demanding. I can't understand how changing the game but staying within the same genre is "throwing out years of progress" in other thing than maybe code alone. And I suppose fourth is availability - access to BW copies is much easier than access to SC2 copies, which makes it easier to collaborate with people who are conscious about spending money on lots of copies of SC2. This is a missed argument since starter edition provides all the potential AI research needs. You just don't need to buy the full game and the API will be free and made - as they said - accessible to variety of people, not just academic researchers so the interest in potential research (or the game itself by the way) may rise among many people. I completely disagree with your opinion that they made the deal to promote SC2 first. One may think that people doing research would want access to multiplayer since playing ladder games is a perfectly valid approach to training your AI. It's not an insurmountable issue, but it's just not one that BW has. It's a game easily available for free. They said they will never let the API to be used on ladder, but focus more on letting it interpret replays gathered from actual ladder games played by people and go from there.
Then that is a reduction of AI capabilities for stupid reasons. BW AI do sometimes train themselves in ladder games. At any rate they are capable of doing so.
|
On November 08 2016 03:45 LegalLord wrote:Show nested quote +On November 08 2016 03:41 aQuaSC wrote:On November 08 2016 03:32 LegalLord wrote:On November 08 2016 03:27 aQuaSC wrote:On November 08 2016 03:22 LegalLord wrote:On November 08 2016 03:14 aQuaSC wrote:On November 08 2016 03:06 LegalLord wrote:On November 08 2016 03:01 aQuaSC wrote:On November 08 2016 02:27 LegalLord wrote:On November 08 2016 02:22 Jett.Jack.Alvir wrote:
Well this would be a huge PR stunt for both Blizzard and Google. SC2 would get huge coverage, look at Go, and Google would push the boundaries of AI learning.
So Blizzard is allowing DeepMind to access SC2 api? I thought that DeepMind was looking at BW because there is no api restrictions. On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way On the other hand you may just hate Blizzard too much, why BW is better than SC2 for potential AI? Just because of the fact that the scene is now established? I think people at Deepmind are much more qualified to make such assumptions or comparisons, if they really thought BW would be better for it they would do it in BW. Or maybe evil Mike bought Google and their Deepmind team? I don't hate SC2 at all. I simply don't see their rationale, from an academic perspective, of actually doing SC2, precisely because they are throwing out years of progress by those who made AI before them. Among other things, there is a 24/7 AI test stream for BW and countless works by Berkeley and others, all done with BW. DeepMind can make their own decisions, but I do not believe they are made from an academically wise perspective. They are basically retreading old ground for no academically justifiable reason. I'd love to hear why is it wise to make the research behind each choice of games, and it's weird to me to think that the progress made in BW has to be largely scrapped and can't be moved over to a game of the same genre, but which has more complex and challenging choices on the example of macro mechanics which BW doesn't have, BW macro mechanic is ordering newly trained workers to mine which is not choice-based but mechanically demanding. I can't understand how changing the game but staying within the same genre is "throwing out years of progress" in other thing than maybe code alone. And I suppose fourth is availability - access to BW copies is much easier than access to SC2 copies, which makes it easier to collaborate with people who are conscious about spending money on lots of copies of SC2. This is a missed argument since starter edition provides all the potential AI research needs. You just don't need to buy the full game and the API will be free and made - as they said - accessible to variety of people, not just academic researchers so the interest in potential research (or the game itself by the way) may rise among many people. I completely disagree with your opinion that they made the deal to promote SC2 first. One may think that people doing research would want access to multiplayer since playing ladder games is a perfectly valid approach to training your AI. It's not an insurmountable issue, but it's just not one that BW has. It's a game easily available for free. They said they will never let the API to be used on ladder, but focus more on letting it interpret replays gathered from actual ladder games played by people and go from there. Then that is a reduction of AI capabilities for stupid reasons. BW AI do sometimes train themselves in ladder games. At any rate they are capable of doing so.
You don't know if there is not going to be a way of automating or simulating the matchmaking the ladder has in the package they are going to release. Why do you think letting bots into a game where people play is good?
|
United Kingdom13775 Posts
On November 08 2016 03:49 aQuaSC wrote:Show nested quote +On November 08 2016 03:45 LegalLord wrote:On November 08 2016 03:41 aQuaSC wrote:On November 08 2016 03:32 LegalLord wrote:On November 08 2016 03:27 aQuaSC wrote:On November 08 2016 03:22 LegalLord wrote:On November 08 2016 03:14 aQuaSC wrote:On November 08 2016 03:06 LegalLord wrote:On November 08 2016 03:01 aQuaSC wrote:On November 08 2016 02:27 LegalLord wrote:
[quote]
On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way On the other hand you may just hate Blizzard too much, why BW is better than SC2 for potential AI? Just because of the fact that the scene is now established? I think people at Deepmind are much more qualified to make such assumptions or comparisons, if they really thought BW would be better for it they would do it in BW. Or maybe evil Mike bought Google and their Deepmind team? I don't hate SC2 at all. I simply don't see their rationale, from an academic perspective, of actually doing SC2, precisely because they are throwing out years of progress by those who made AI before them. Among other things, there is a 24/7 AI test stream for BW and countless works by Berkeley and others, all done with BW. DeepMind can make their own decisions, but I do not believe they are made from an academically wise perspective. They are basically retreading old ground for no academically justifiable reason. I'd love to hear why is it wise to make the research behind each choice of games, and it's weird to me to think that the progress made in BW has to be largely scrapped and can't be moved over to a game of the same genre, but which has more complex and challenging choices on the example of macro mechanics which BW doesn't have, BW macro mechanic is ordering newly trained workers to mine which is not choice-based but mechanically demanding. I can't understand how changing the game but staying within the same genre is "throwing out years of progress" in other thing than maybe code alone. And I suppose fourth is availability - access to BW copies is much easier than access to SC2 copies, which makes it easier to collaborate with people who are conscious about spending money on lots of copies of SC2. This is a missed argument since starter edition provides all the potential AI research needs. You just don't need to buy the full game and the API will be free and made - as they said - accessible to variety of people, not just academic researchers so the interest in potential research (or the game itself by the way) may rise among many people. I completely disagree with your opinion that they made the deal to promote SC2 first. One may think that people doing research would want access to multiplayer since playing ladder games is a perfectly valid approach to training your AI. It's not an insurmountable issue, but it's just not one that BW has. It's a game easily available for free. They said they will never let the API to be used on ladder, but focus more on letting it interpret replays gathered from actual ladder games played by people and go from there. Then that is a reduction of AI capabilities for stupid reasons. BW AI do sometimes train themselves in ladder games. At any rate they are capable of doing so. You don't know if there is not going to be a way of automating or simulating the matchmaking the ladder has in the package they are going to release. Why do you think letting bots into a game where people play is good?
Do you have a problem with playing with a robot that is about your skill level if you have no way of knowing whether or not it is actually a robot you're playing?
No one in BW ever has had a problem with that. It's so inconsequential relative to the size of the playerbase that it really just doesn't matter. But for training the AI it absolutely does matter.
|
On November 08 2016 03:28 imp42 wrote:Show nested quote +On November 08 2016 03:01 aQuaSC wrote:On November 08 2016 02:27 LegalLord wrote:On November 08 2016 02:22 Jett.Jack.Alvir wrote:
Well this would be a huge PR stunt for both Blizzard and Google. SC2 would get huge coverage, look at Go, and Google would push the boundaries of AI learning.
So Blizzard is allowing DeepMind to access SC2 api? I thought that DeepMind was looking at BW because there is no api restrictions. On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way On the other hand you may just hate Blizzard too much, why BW is better than SC2 for potential AI? Just because of the fact that BW AI scene is now established? Can you provide solid reasons instead of what you believe is true? I think people at Deepmind are much more qualified to make such assumptions or comparisons, if they really thought BW would be better for it they would do it in BW. Or maybe evil Mike bought Google and their Deepmind team? I think I can give you plenty reasons: - BW is much more stable. Who knows if Blizzard is going to release a new patch. Optimistically you could say that's the challenge that Deepmind wants. Realistically you have to admit it if they really want to deal with changing rules it would be much better for them to have those rules under their own control (by e.g. tweaking units in a map editor) without having any conflict of interest created by the fact Blizzard also has to support an active player base. - Guess which program is going to be more light-weight to execute. A program developed in 1998, runnable on Windows98, or Sc2? There already exists a "headless" version of BW (no graphics) and you can easily create pretty much any API you want. - compared to Sc2 BW is simpler in terms of possible moves, but arguably* deeper strategically. If you're really interested in "real" AI you would want to strip any unnecessary complexity and focus on the core issue. - BW has a low resolution, making the jump from Atari games more reasonable if you want to go the pixel interpretation way. - As has been said before, there is already research available on BW. The statement at the Blizzard panel that "all BW bots are scripted" is not a 100% true. Approaches using Neural Nets have been explored as well. * arguably deeper: my personal opinion, no need to discuss it. If anybody thinks otherwise that's fine with me
Stability is actually a point I agree with you on. But maybe Blizzard considers SC2 in its current version stable enough that big balance tweeks to multiplayer will no longer happen, just as they haven't happened to BW. Is SC2 perfectly balanced? No. But BW isn't either. Not sure what balance patches are planned, but major balance updates would mean AI needs retraining, which is a major pain in the ass.
SC2 is not heavy-weight enough to really worry about. Yes, SC2 will eat more CPU cycles. But when you're using a couple of orders of magnitude MORE cycles for your AI, the difference between SC2 and BW becomes a rounding error.
Strategic depth is a matter of taste. Lets not bring that discussion over here.
Resolution is not all that relevant for image recognition: it's more about how much stuff is on the screen, which would be about the same, given that the increased resolution is used by making the units look prettier, rather than fitting (significantly) more units on the screen. In any case, computer vision algorithms would have a bloody hard time on SC, and the initial phase will almost certainly be the API providing information about what is on the screen directly from the engine, rather than a CV module analyzing that. I do agree that the simplicity of being able to match a (very restricted) number of sprites makes the CV easier than the 3D model approach of SC2, but I am not enough of a CV expert to be able to say how much easier BW is in this department.
I already discussed the point about prior research above.
|
On November 08 2016 03:43 LegalLord wrote:Show nested quote +On November 08 2016 03:40 Acrofales wrote:On November 08 2016 02:27 LegalLord wrote:On November 08 2016 02:22 Jett.Jack.Alvir wrote:
Well this would be a huge PR stunt for both Blizzard and Google. SC2 would get huge coverage, look at Go, and Google would push the boundaries of AI learning.
So Blizzard is allowing DeepMind to access SC2 api? I thought that DeepMind was looking at BW because there is no api restrictions. On your question: yes. Blizzard is making an API for DeepMind to use here. Specifically for this purpose. On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way. This is because by choosing SC2, they are throwing out a lot of highly valuable and relevant progress towards a well-developed and mature, if horribly incomplete, project. And Google isn't some genius that can simply toss aside that much progress, "pave their own path," and expect it will somehow work out for the best. It's not true. I kinda disagree that they are throwing it out. If it's interesting algorithms, they can pretty easily be assimilated. It shouldn't take much work to make learning algorithms designed for the BW API to work on the SC2 API (assuming Blizzard makes a decent SC2 API). If it's the work that has been done on perfect muta control bots, then I disagree with the premise, because that is not very interesting from a research point of view anyway (although it's pretty impressive from a mechanical perspective Algorithms are primarily mathematical, so they would be able to use those with some primarily practical tweaking. The codebase already made, and the BW AI community to collaborate with, that is lost by going to SC2.
Well, I presume that having both Blizzard and Google support a new SC2 AI community would cause some drive for the current BW AI community to switch. Note that most research groups working on BW AI are not particularly attached to BW. They just like the tools that the BW community has available. If Blizzard and Google make better tools available (and interesting tournaments), then research groups will switch faster than stalkers can blink.
|
that is only if they release the api to other institutions. it may be that google gets the api only, so no one else can switch from bw
|
On November 08 2016 03:59 Jett.Jack.Alvir wrote:
that is only if they release the api to other institutions. it may be that google gets the api only.
It will be for free for everyone interested with no limitations.
|
On November 08 2016 03:59 Jett.Jack.Alvir wrote:
that is only if they release the api to other institutions. it may be that google gets the api only, so no one else can switch from bw
That would be extremely dumb from a PR point of view. They also said the API would be publicly available.
|
I will give you guys one more reason for BW:
BW is very close to being completely reverse engineered and open sourced.
This is a screenshot the developer just released from his own BW engine: http://i.imgur.com/rdEcBYM.jpg
This is obviously big news and will deserve its own thread in due time. Take it as a teaser for now 
I am not personally involved, just the messenger. The genius behind the project is an active BW Bot developer.
|
United Kingdom13775 Posts
On November 08 2016 03:56 Acrofales wrote:Show nested quote +On November 08 2016 03:43 LegalLord wrote:On November 08 2016 03:40 Acrofales wrote:On November 08 2016 02:27 LegalLord wrote:On November 08 2016 02:22 Jett.Jack.Alvir wrote:
Well this would be a huge PR stunt for both Blizzard and Google. SC2 would get huge coverage, look at Go, and Google would push the boundaries of AI learning.
So Blizzard is allowing DeepMind to access SC2 api? I thought that DeepMind was looking at BW because there is no api restrictions. On your question: yes. Blizzard is making an API for DeepMind to use here. Specifically for this purpose. On your earlier statement: it's a PR stunt, yes, but the focus on SC2 makes it questionable whether or not they are really looking to "push the boundaries of AI learning" in any appreciable way. This is because by choosing SC2, they are throwing out a lot of highly valuable and relevant progress towards a well-developed and mature, if horribly incomplete, project. And Google isn't some genius that can simply toss aside that much progress, "pave their own path," and expect it will somehow work out for the best. It's not true. I kinda disagree that they are throwing it out. If it's interesting algorithms, they can pretty easily be assimilated. It shouldn't take much work to make learning algorithms designed for the BW API to work on the SC2 API (assuming Blizzard makes a decent SC2 API). If it's the work that has been done on perfect muta control bots, then I disagree with the premise, because that is not very interesting from a research point of view anyway (although it's pretty impressive from a mechanical perspective Algorithms are primarily mathematical, so they would be able to use those with some primarily practical tweaking. The codebase already made, and the BW AI community to collaborate with, that is lost by going to SC2. Well, I presume that having both Blizzard and Google support a new SC2 AI community would cause some drive for the current BW AI community to switch. Note that most research groups working on BW AI are not particularly attached to BW. They just like the tools that the BW community has available. If Blizzard and Google make better tools available (and interesting tournaments), then research groups will switch faster than stalkers can blink.
And they might just draw more people in in general. If they do, I'm all for it, since another API won't hurt anyone. But Blizzard's track record of peaceful transitions is spotty, and there will still be a new community to rebuild, and a codebase too. It will take years before the community is as productive as what BW has now. So DeepMind takes a decidedly less productive approach here.
Though frankly, I think to a large extent Blizzard missed the boat by waiting this long to even announce that they are making the API. If this came out on SC2 release it would have caught on for sure. Now, it's a giant maybe. The community has waned since the release and that will not make a justification for the transition be as easy as it would be when the hype train was in full swing.
|
Russian Federation4235 Posts
On November 06 2016 14:10 Lexender wrote:Show nested quote +On November 06 2016 08:03 BluzMan wrote:
EDIT: I actually think that the best advantage of a potential AI is individual unit control and AoE-immune spreads. StarCraft 2 default AI (the one that players trigger by rightclicking) is VERY BAD in aoe-related pathfinding, so that alone might be enough to make all other advantages irrelevant. They are not going to do that tho, they aren't going to spend a ton of money in a super AI only to make it be a automaton 1000 bot that only makes perfectly microed marines.
Geez, you only read and quoted one half of the post. AlphaGo is based partly on unsupervised learning, basically, it invents algorithms by itself. If it ever finds a game mechanic loophole that gives it unholy winrates without really "understanding" the game, it stops learning, because it completely loses reference (you have no idea what to improve when you win all the time). The only thing it can do then is play with itself, but if it's basis isn't strong enough by that time, it will keep discovering "features" that are not relevant to real games. Just like those poor sods that discovered "tech to carriers and make 12 of them" to be a good strategy because of total lack of competition.
Allowing very high APM might be a concern because it might uncover something that completely breaks the game (I personally think it's AoE). But it's not known if such things exist, and an AI would need all advantages it has to hope to beat a human.
|
United Kingdom13775 Posts
If AI get hax speed performance then humans get multicommand hacks and automated micro bots. If the computers only win because they are not input-constrained by having physical bodies that have to be moved to provide input into the game then the entire problem has been trivialized.
|
On November 08 2016 07:01 LegalLord wrote:
If AI get hax speed performance then humans get multicommand hacks and automated micro bots. If the computers only win because they are not input-constrained by having physical bodies that have to be moved to provide input into the game then the entire problem has been trivialized.
I think everyone agrees that there needs to be a very low limit to APM. I'd even argue that extremely low APM (like 180) and a small "between actions" limit that is similar to the time it takes humans to touch two different keys in succession on a keyboard would be much more interesting because it would require the computer to "outthink" humans instead of "outdoing" us. Also, maybe having the program not "remember" how many resources it has unless it spends one of those actions to "look" at its money (and things like that) would be helpful too.
|
On November 08 2016 09:03 mierin wrote:Show nested quote +On November 08 2016 07:01 LegalLord wrote:
If AI get hax speed performance then humans get multicommand hacks and automated micro bots. If the computers only win because they are not input-constrained by having physical bodies that have to be moved to provide input into the game then the entire problem has been trivialized. I think everyone agrees that there needs to be a very low limit to APM. I'd even argue that extremely low APM (like 180) and a small "between actions" limit that is similar to the time it takes humans to touch two different keys in succession on a keyboard would be much more interesting because it would require the computer to "outthink" humans instead of "outdoing" us. Also, maybe having the program not "remember" how many resources it has unless it spends one of those actions to "look" at its money (and things like that) would be helpful too.
They mentioned a lot of things like that during the panel, actually about resources they said they may just pass it the numbers. It was also mentioned that the restrictions on the AI will also be entirely based on the game interface people use, so selecting things will involve selecting single units as well as boxing over them (with ctrl+click and control groups implied I think). I'm interested in how will they deal with reacting to things happening on the minimap. I bet APM constraints may change on how it all develops.
After all the project is about creating a self-learning AI that would simulate the process of how human thinks, not just creating a bot that plays the game as best as possible - although that is the goal, but it's a goal set for AI, not researchers.
I'm super excited thinking about how FPV of the AI will look like when significant progress is made. It has to be a regular stream on twitch at some point, haha.
|
You know what would be fucking cool? If Blizzard opened up a specific ladder for humans to play the AI. Of course the wait time would be incredibly long, unless Google invests an entire server farm to run the AI so it can play multiple instances.
|
United Kingdom13775 Posts
On November 08 2016 09:24 aQuaSC wrote:Show nested quote +On November 08 2016 09:03 mierin wrote:On November 08 2016 07:01 LegalLord wrote:
If AI get hax speed performance then humans get multicommand hacks and automated micro bots. If the computers only win because they are not input-constrained by having physical bodies that have to be moved to provide input into the game then the entire problem has been trivialized. I think everyone agrees that there needs to be a very low limit to APM. I'd even argue that extremely low APM (like 180) and a small "between actions" limit that is similar to the time it takes humans to touch two different keys in succession on a keyboard would be much more interesting because it would require the computer to "outthink" humans instead of "outdoing" us. Also, maybe having the program not "remember" how many resources it has unless it spends one of those actions to "look" at its money (and things like that) would be helpful too. They mentioned a lot of things like that during the panel, actually about resources they said they may just pass it the numbers. It was also mentioned that the restrictions on the AI will also be entirely based on the game interface people use, so selecting things will involve selecting single units as well as boxing over them (with ctrl+click and control groups implied I think). I'm interested in how will they deal with reacting to things happening on the minimap. I bet APM constraints may change on how it all develops. After all the project is about creating a self-learning AI that would simulate the process of how human thinks, not just creating a bot that plays the game as best as possible - although that is the goal, but it's a goal set for AI, not researchers. I'm super excited thinking about how FPV of the AI will look like when significant progress is made. It has to be a regular stream on twitch at some point, haha.
You can look at the stream I linked to see what an AI FPV looks like for BW. It would be like that, except SC2 instead.
|
United States2186 Posts
It makes sense for Google to start with sc2, though whether they actually had a full understanding of the differences between games is highly uncertain. Blizzard's attitude here is of course obvious and has nothing to do with the spectator quality or strategic aspect. This is already a ridiculously difficult project, so why make it harder for themselves on taking on a game with much more strategic depth (and likely quality of opponents) first. The question is whether they will move onto BW afterwards for the next level of challenge.
When they entered Go it wasn't as if they challenged the best player in the world right away (for that matter they never did). They took it cautiously and challenged a second-tier player, and only challenged Lee Sedol after AlphaGo was proven and had plenty of time for feedback. If Lee Sedol got to play against the initial version of AlphaGo, he likely would have done very well. Then again, Go never had a dice roll aspect to it heh.
|
On November 08 2016 04:01 imp42 wrote:I will give you guys one more reason for BW: BW is very close to being completely reverse engineered and open sourced. This is a screenshot the developer just released from his own BW engine: http://i.imgur.com/rdEcBYM.jpgThis is obviously big news and will deserve its own thread in due time. Take it as a teaser for now I am not personally involved, just the messenger. The genius behind the project is an active BW Bot developer.
Not sure what the point of discussing why not BW is when it's clear that SC2 is the way they're going and that they've already put a lot of work into it both from the Blizzard side and Google side. One of the deepmind devs who spoke at the panel Oriol Vinyals was even one of the lead devs on the Berkeley Overmind bw bot so I'm sure he knows the pros and cons of BW vs SC2 and has a very good idea of what they want to achieve.
|
The amount of discussion here of APM and how to make the showmatch between man and machine fair and all of that is sort of baffling. This shouldn't need to be said, but I'll say it anyways: Deepmind's goal is not to make the perfect Starcraft bot. Their goal is to find new algorithms and techniques to apply to the quest for Artificial General Intelligence - that is, a robot that can fight you in Starcraft, do your taxes and discuss your marital issues without changing the underlying code itself.
This means that worries of it being too mechanically powerful for the showmatch are silly because, again, that's not their goal. They don't need to put millions of dollars and thousands of manhours into making an AI that can microbot a pro to death. Their task is to find a clever way of beating people since that helps their central goal. And, yes, this also means BW vs SC2 talk is silly because they aren't trying to prove that they can create the best Starcraft bot ever. The game is purely a testing ground to see how effective they can make the AI in thinking and planning and executing. There are a lot of cool AI projects for Brood War, yes, but they are not going to base their AI off those because, again, they're going for general. Neural Network AI for Brood War, while super cool, are very very narrow in what they can do. They play Starcraft. Deepmind wants to do a lot more than play Starcraft and other games.
|
United Kingdom13775 Posts
On November 09 2016 04:57 Derpmallow wrote:
The amount of discussion here of APM and how to make the showmatch between man and machine fair and all of that is sort of baffling. This shouldn't need to be said, but I'll say it anyways: Deepmind's goal is not to make the perfect Starcraft bot. Their goal is to find new algorithms and techniques to apply to the quest for Artificial General Intelligence - that is, a robot that can fight you in Starcraft, do your taxes and discuss your marital issues without changing the underlying code itself.
This means that worries of it being too mechanically powerful for the showmatch are silly because, again, that's not their goal. They don't need to put millions of dollars and thousands of manhours into making an AI that can microbot a pro to death. Their task is to find a clever way of beating people since that helps their central goal. And, yes, this also means BW vs SC2 talk is silly because they aren't trying to prove that they can create the best Starcraft bot ever. The game is purely a testing ground to see how effective they can make the AI in thinking and planning and executing. There are a lot of cool AI projects for Brood War, yes, but they are not going to base their AI off those because, again, they're going for general. Neural Network AI for Brood War, while super cool, are very very narrow in what they can do. They play Starcraft. Deepmind wants to do a lot more than play Starcraft and other games.
The entire issue of speed is one of trivializing the problem. It's like building a football/soccer robot that is just a giant automated tank that will crush its human opposition. Yes it will win, but it didn't do so by virtue of its AI but rather by virtue of an interface that gives it an unfair advantage.
|
|
|
|
|
|
EPT
