EPTHow can we prepare for fighting the AI? - Page 2
| Forum Index > SC2 General |
|
StarscreamG1
Portugal1653 Posts
| ||
|
Ej_
47656 Posts
On July 12 2017 22:23 StarscreamG1 wrote: No way alpha go will be on SC2, 100 apm, marines and medivacs would be enough. The Deepmind team announced SC2 AI as their new project quite a while ago. | ||
|
Deleted User 3420
24492 Posts
On July 12 2017 23:53 Ej_ wrote: The Deepmind team announced SC2 AI as their new project quite a while ago. where did you see this? I've only ever seen it announced as an "interest" or "something they'd like to conquer in the future" (basically something they spend time on but not an official project like alphago for example) anyways sc2 is a different beast from turn based games chess or go have branching factors that could at least be represented with numbers the branching factors in sc2 are basically infinity machine learning approaches to games like sc2 are going to need completely new and radical methods, and when they've conquered it, it will mean we are close to generalized machine learning. don't be surprised if when machine learning based sc2 AIs are released by say, google - that they are actually very limited in comparison to what many people might expect. they may not be able to adapt or strategize quite like you might think if you use alphago's performance to form your expectations | ||
Poopi
France12907 Posts
| ||
|
BronzeKnee
United States5219 Posts
But many of you assume the AI will have an unfair advantage without actually realizing that it is an unfair advantage. On July 12 2017 19:34 Grumbels wrote: While it seems very difficult to create an AI capable of outthinking humans in SC2, I don't think it will take them 10 years to beat humans. If you have perfect micro there must be an infinite number of one-trick build orders that are unstoppable by conventional means. The biggest limitation a human has is having to use a monitor, speakers, keyboard and mouse to control the game. Build an AI that mechanically controls the game with a mouse, monitor, speakers and a keyboard, and therefore suffers the same limitations as a human (like the mouse slipping a bit more on their mouse pad than expected causing a mis-micro) and the best players will toast the AI for many years to come. The more interesting dynamic is allow humans to see, hear and control the game with their mind and compete with an AI. Then, humans could achieve perfect micro and macro and a ridiculously high APM, and I see no possibility for AI to defeat them with regularity in my lifetime. The mind is too powerful. Remember SC2 is played with a mouse, keyboard, speakers and a monitor. The AI should have to play with those too. Or the mind shouldn't have to. | ||
|
DeepElemBlues
United States5079 Posts
On July 12 2017 19:57 hypercube wrote: AI have beaten humans in heads up texas holdem which has hidden information and randomness. And the real-time element probably favours AI, not humans. At least that was true for chess, where humans resisted far longer in correspondence chess than over the board play. SC2 has far more ambiguity, far more possibilities, in its hidden information and randomness than Texas holdem. Far, far more. AI advantage over humans in SC2 would be mostly tactical - perfect build execution, perfect micro - until the AI built up knowledge of its human opponents and also had the capacity to correctly weigh possibilities based on incomplete information. Which is one of the hardest parts of creating a true AI, giving it that human intuition, the leaps of logic that human brains are capable of making using incomplete information and still reaching an accurate conclusion. | ||
|
dragoon
United States695 Posts
| ||
|
hypercube
Hungary2735 Posts
On July 13 2017 01:25 DeepElemBlues wrote: SC2 has far more ambiguity, far more possibilities, in its hidden information and randomness than Texas holdem. Far, far more. But one of the consequences is that a lot of that complexity isn't explored by humans either. One of the interesting takeaway from the poker match was that humans had trouble adjusting to styles that was very rare among humans or to situations that came up infrequently. E.g. strange slowplays followed by large overbets (with superhuman balance). People are wondering about how the AI will be able to read humans and adjust to mindgames. But humans reading the AI might be an even tougher challange. Which is one of the hardest parts of creating a true AI, giving it that human intuition, the leaps of logic that human brains are capable of making using incomplete information and still reaching an accurate conclusion. Intuition is no divine spark, it relies on experience. The further out you are from your usual experience the more you have to rely on conscious thought, which is far slower. And modern AI do learn from experience. That's what made AlphaGo so successful. Learning first from human play, than self-play, to the point where it amassed orders of magnitudes more experience than pros who studied the game their whole life. If humans still have an advantage it's in our pace of learning. We can learn from fewer examples. But since Deepmind was only restricted in APM, not the number of training games they can use, I don't see how this will serve us. | ||
|
Modesty00
Bulgaria262 Posts
The thing is can the machine outsmart and predict human in strategy, build order and early game plan. Can a machine make a massive drop while setting a trap on other side of the map? Can machine change game plan and readjust accoriding to new information that receives about opposition? Can hide his intention? Cancel building? Make a pylon in secret area? Those and more of this type of questino are importhant. A.I did this many years ago with chess. Not only showed perfect tactical vision, traps, sacrifices, incredible defense and deep calculation predicting future, (the easy part) but was able to spot deep strategic ideas and plans. And i think this eventually will happen. The Brute force of CPU is already enough. We don't need faster CPU. The SC is limited in terms of possibilities. Chess and GO had infinity positions and still the machine won not by seeing everything because it can't (CPU not fast enough even all super computers put togheter). It will do it with starcraft too, but it might take longer, because is completely different type of game with inconpete information but defintely possible. It will be expontential growth. For example 3 month AI practice will be playing like crap. 1 year later still bad. 3 years later still making silly mistakes and playing worse than C- broodwar player. but 2 weeks more and is beyond super human level out of nowhere. Just like they expected Alpha Go to beat best GO player in 2025.. but it happened in 2016 huge surprise. So we just have to wait and will happen. AI to be smart as bug - a lot of years, to be smart as dog - many many years, to reach chimpase level - super long. To go from chimpase to ultra smarter many times more than all humans brain put togheter - just a year. Technical Singularity is close, but that does not mean it will happen, but is getting closer. Is like building a puzzle with 100 pieces. It's extremely hard to make sense which piece have to go with which piece at beginning or even at middle. So one by one, with test you link them togheter. At the end you have just 5 pieces left. So much work and so much time has passed. But still nothing - is incomplete. Then BUM.. last 5 pieces are easier than first 80 and from nothing to complete picture is way shorter time. | ||
|
sabas123
Netherlands3122 Posts
On July 13 2017 01:23 BronzeKnee wrote: It is going to be a long time before AI conquers the human mind. I don't think I will be alive to see it, and even the youngest among us probably won't. But many of you assume the AI will have an unfair advantage without actually realizing that it is an unfair advantage. Remember SC2 is played with a mouse, keyboard, speakers and a monitor. The AI should have to play with those too. Or the mind shouldn't have to. I have no idea what your trying to imply, but the AI won't use anything physical, but will have access to a virtual version of those. | ||
|
Simberto
Germany11735 Posts
On August 06 2017 18:58 sabas123 wrote: I have no idea what your trying to imply, but the AI won't use anything physical, but will have access to a virtual version of those. And he implies that that is an advantage. The AI can basically control the game "with its mind", while for a human, there are a bunch of hoops to jump through. I actually hope that we will eventually get to control games "with our minds". UI design is basically just that, trying to make controlling games feels more natural and not require extra thoughts. For me this is often one of the problems with games, especially strategy games that require a lot of different inputs quickly. I know what needs to be done, but i just can't get that information from my mind to the units on screen quickly enough. Of course, more practice would make this easier and work better, but there is a big hurdle in between my mind and the game, and an AI does not have this problem. | ||
|
lestye
United States4186 Posts
On July 13 2017 01:04 travis wrote: where did you see this? I've only ever seen it announced as an "interest" or "something they'd like to conquer in the future" (basically something they spend time on but not an official project like alphago for example) anyways sc2 is a different beast from turn based games chess or go have branching factors that could at least be represented with numbers the branching factors in sc2 are basically infinity machine learning approaches to games like sc2 are going to need completely new and radical methods, and when they've conquered it, it will mean we are close to generalized machine learning. don't be surprised if when machine learning based sc2 AIs are released by say, google - that they are actually very limited in comparison to what many people might expect. they may not be able to adapt or strategize quite like you might think if you use alphago's performance to form your expectations This was announced last Blizzcon. They even had a panel on it and everything. http://us.battle.net/sc2/en/blog/20352109/blizzcon-recap-deepmind-and-starcraft-ii-deep-learning-11-5-2016 On July 12 2017 17:05 Mendelfist wrote: Sorry, not true. AlphaGo learns by self play, not just by observing human games. This means that it can develop new strategies and meta by itself. The latest iteration of AlphaGo won 30-0 against top pros on a Go-server, and 3-0 against the no 1 world champion Ke Jie. It's play style especially in the opening is new and all over the world professionals are trying to copy it and experiment with these new ideas. In fact, the AI that is teaching us humans, not the other way around. As a fan of cyberpunk, that sounds so awesome! | ||
|
mishimaBeef
Canada2259 Posts
| ||
|
Grumbels
Netherlands7031 Posts
On August 06 2017 17:39 Modesty00 wrote: Ofcourse AI will be better in terms ot makro and micro. Mechanical task are easy for a machine , humans dont want to compete in that area and you dont have to. Once game reaches late middle game and position is equal the machine will take over easy. The thing is can the machine outsmart and predict human in strategy, build order and early game plan. Can a machine make a massive drop while setting a trap on other side of the map? Can machine change game plan and readjust accoriding to new information that receives about opposition? Can hide his intention? Cancel building? Make a pylon in secret area? Those and more of this type of questino are importhant. A.I did this many years ago with chess. Not only showed perfect tactical vision, traps, sacrifices, incredible defense and deep calculation predicting future, (the easy part) but was able to spot deep strategic ideas and plans. And i think this eventually will happen. The Brute force of CPU is already enough. We don't need faster CPU. The SC is limited in terms of possibilities. Chess and GO had infinity positions and still the machine won not by seeing everything because it can't (CPU not fast enough even all super computers put togheter). It will do it with starcraft too, but it might take longer, because is completely different type of game with inconpete information but defintely possible. It will be expontential growth. For example 3 month AI practice will be playing like crap. 1 year later still bad. 3 years later still making silly mistakes and playing worse than C- broodwar player. but 2 weeks more and is beyond super human level out of nowhere. Just like they expected Alpha Go to beat best GO player in 2025.. but it happened in 2016 huge surprise. So we just have to wait and will happen. AI to be smart as bug - a lot of years, to be smart as dog - many many years, to reach chimpase level - super long. To go from chimpase to ultra smarter many times more than all humans brain put togheter - just a year. Technical Singularity is close, but that does not mean it will happen, but is getting closer. Is like building a puzzle with 100 pieces. It's extremely hard to make sense which piece have to go with which piece at beginning or even at middle. So one by one, with test you link them togheter. At the end you have just 5 pieces left. So much work and so much time has passed. But still nothing - is incomplete. Then BUM.. last 5 pieces are easier than first 80 and from nothing to complete picture is way shorter time. I think that's a good point. If you look at the Encephalization quotient (roughly brain to body mass ratio) then humans outpace other great apes by only a factor of three. Assuming that the main difference between humans and others is in brain size, which translates to computing power, that means that if you are capable of simulating a chimpansee then simulating human intelligence is only a question of linking another supercomputer. This is obviously a rather simplified of looking at it, since brain differences between mammals are qualitative as well as quantitative, but it is useful. If you compare Starcraft players at the top level, then there are primarily quantitative differences. All players are capable of doing everything, but some are better at it. The leap from awful to godlike execution for an AI is fairly trivial, the more difficult question is whether you can bring an AI to care about scouting or micro or build orders etc. at all. Once it is capable of competing with a human on any level, then it's only a few more months of practice or another upgrade in hardware for it to vastly outpace any human. On August 06 2017 19:26 Simberto wrote: And he implies that that is an advantage. The AI can basically control the game "with its mind", while for a human, there are a bunch of hoops to jump through. I actually hope that we will eventually get to control games "with our minds". UI design is basically just that, trying to make controlling games feels more natural and not require extra thoughts. For me this is often one of the problems with games, especially strategy games that require a lot of different inputs quickly. I know what needs to be done, but i just can't get that information from my mind to the units on screen quickly enough. Of course, more practice would make this easier and work better, but there is a big hurdle in between my mind and the game, and an AI does not have this problem. This seems fun for FPS games. Everyone always has perfect headshots because that's how they envisioned it mentally. | ||
|
Slydie
1929 Posts
On August 07 2017 18:02 Grumbels wrote: I think that's a good point. If you look at the Encephalization quotient (roughly brain to body mass ratio) then humans outpace other great apes by only a factor of three. Assuming that the main difference between humans and others is in brain size, which translates to computing power, that means that if you are capable of simulating a chimpansee then simulating human intelligence is only a question of linking another supercomputer. This is obviously a rather simplified of looking at it, since brain differences between mammals are qualitative as well as quantitative, but it is useful. If you compare Starcraft players at the top level, then there are primarily quantitative differences. All players are capable of doing everything, but some are better at it. The leap from awful to godlike execution for an AI is fairly trivial, the more difficult question is whether you can bring an AI to care about scouting or micro or build orders etc. at all. Once it is capable of competing with a human on any level, then it's only a few more months of practice or another upgrade in hardware for it to vastly outpace any human. This seems fun for FPS games. Everyone always has perfect headshots because that's how they envisioned it mentally. Once a good enough AI id made, programming it to use a human interface should not be too hard. It would hotkey around like a mofo, not to mention, it could micro pretty well on the minimap. It is not a question of IF, but rather when. I can imagine the horrors a perfected AI could do with 2 medivacs and 10 marines, using pickups, individual stims, splits and targetting to hold off an insane amount of ling/bane controlled by a human. A good startingpoint would be studying a million replays, to learn the metagame. Computers are stupid, so starting with "most probable scenario" would probably mean worker-rushing, because it would calculate it probable that the opponent would move his own workers randomly around the map, as there are more ways to do that than to mine... Good scouting should be a very hard thing to program, but I can imagine an AI sending halucinated fenixes and lings all over the place, and an AI would calculate extremely accurately based on what it can observe, production, mining, expansions etc. | ||
|
LDaVinci
France130 Posts
Same errors. We saw Inno yesterday build a second armory and then canceling, cause it was a mistake. If even those ro players with 12 hours of play per days are making those mistakes, it means that it would be "unfair" to remove them completely from the AI play. Maybe you don't have to go that far in the mistakes made by the computer, but you should have a small percentage of mistakes still. | ||
|
aerlinss
3 Posts
From what it sees, the amout of drones, the amout of minerals mined on every single patches, it will exactly know which of all the possible every played openings are possible, and which are not. It will know if there is an scv/probe out on the map building a proxy, or not. It will see your gas minded, ur units and buildings built and will instantly know what s the earliest possible time when a cloaked banshee or dark templar can arrive at his base. It will detect certain openings and know exactly how many zerglings he has to build to stop this push, he will not build a single zerling too much. You don t even need a really smart AI, all you need is enough data and a programm able to read this data fast enough. The more data you have, the less smart the AI can be, it doesn t need to understand things, it just needs to copy things that worked out in the past, and since AI will have perfect micro and macro, this will be more than enough. | ||
|
LDaVinci
France130 Posts
On August 07 2017 21:30 aerlinss wrote: The AI enters ur base with a scout. From what it sees, the amout of drones, the amout of minerals mined on every single patches, it will exactly know which of all the possible every played openings are possible, and which are not. It will know if there is an scv/probe out on the map building a proxy, or not. It will see your gas minded, ur units and buildings built and will instantly know what s the earliest possible time when a cloaked banshee or dark templar can arrive at his base. It will detect certain openings and know exactly how many zerglings he has to build to stop this push, he will not build a single zerling too much. You don t even need a really smart AI, all you need is enough data and a programm able to read this data fast enough. The more data you have, the less smart the AI can be, it doesn t need to understand things, it just needs to copy things that worked out in the past, and since AI will have perfect micro and macro, this will be more than enough. Then that is not AI, and that is very far from what is developping with deepmind. And though this might be enough for chess, It was not enough for GO. I guess it wouldn't work with Starcraft also. | ||
|
Slydie
1929 Posts
On August 07 2017 21:30 aerlinss wrote: The AI enters ur base with a scout. From what it sees, the amout of drones, the amout of minerals mined on every single patches, it will exactly know which of all the possible every played openings are possible, and which are not. It will know if there is an scv/probe out on the map building a proxy, or not. It will see your gas minded, ur units and buildings built and will instantly know what s the earliest possible time when a cloaked banshee or dark templar can arrive at his base. It will detect certain openings and know exactly how many zerglings he has to build to stop this push, he will not build a single zerling too much. You don t even need a really smart AI, all you need is enough data and a programm able to read this data fast enough. The more data you have, the less smart the AI can be, it doesn t need to understand things, it just needs to copy things that worked out in the past, and since AI will have perfect micro and macro, this will be more than enough. There are far too many ways to play SC2 to make the AI account for everything, even for a supercomputer. One of the most difficult, and important, things to program would be teaching it not to get lost in useless calculations, like the different ways of moving the starting workers randomly around the map or every possible location of a proxy. The first step would probably be some kind of timing-attack, made to exploit its advantages over humans, and making it safe to as much cheese as possible. The AI should love stimmed marines, medivacs, blink stalkers and mutalisks. I can't imagine ling/bane with perfect superhuman micro, but that sounds scary as well. | ||
|
Grumbels
Netherlands7031 Posts
On August 07 2017 22:05 LDaVinci wrote: Then that is not AI, and that is very far from what is developping with deepmind. And though this might be enough for chess, It was not enough for GO. I guess it wouldn't work with Starcraft also. To add to this, AlphaGo was already very strong even without calculation. Using only its facility for pattern recognition it already outpaced most players. I can't recall the exact ratings. I think SC2 is the sort of game where this sort of ponderous deliberation, where you peer into the forest of future possibilities to navigate a winning path, is almost totally useless. The AI only needs to make okayish decisions based on general principles to win. After all, progamers after the opening tend to play very reactively too, they base their decisions on what seems instinctively correct. And it is very hard for them to make decisive mistakes because they can always retreat or build another base etc. That is why the advice given to aspiring youngsters is not to improve their decisionmaking, but to develop good fundamentals. That said, I recall an experiment I did last year playing WC3 a lot, where I tried to actually think about what I was doing and use strategic tools, instead of playing mindlessly as usual. For instance, I would invest in hardcounter units not part of the meta, or I would scout more than usual and try obnoxious base sniping strats and such. And my winrate went up! Still... | ||
| ||
