|
For those of you who haven't heard of it yet: Artificial Intelligence conquered Go a while ago. So far, the game and it's gosu human players were considered the toughest nut for artificial intelligence to crack. And humans indeed are kind of good at it - They have a bad ass 4'000-year old East Asian tradition, with all their different schools and styles and hundreds of years of paper record of ancient master Games. (1)
The Robots are winning!
They are coming for us next. As mentioned here before, Goolge has set aim to crack Starcraft II. Blizzard is on their side already, there is no stopping them now. (2)
The AI will be playing fair, it only sees what's on the screen, maybe even have an APM limit. But it is coming, and it will be tough. We better burrow those Widdow Mines behind the Ultralisk line and cloak them from our Mothership. (2)
But there is hope. There is a way for our race to fight back, while holding back the attack waves. The best part: It actually does read like straight from the Starcraft missions. The chess scene has already done it years ago, and the Go-scene is currently starting with it.
We will become hybrids. Or in Starcraft-Terms, we will play in Archon Mode with that AI. In chess, it became clear quickly that a Human-AI Archon, or „Centaur“ as they call it (3), will beat both two humans and two AIs for the very best players, and they have their own league now, pushing both human and AI understanding of chess to new limits. Gary Kasparow invented that play style. That approach to dealing with advanced AI is also known as “Race with the Machines” and applies to all Human-AI strategic interaction in life, not just strategy games (4).
We still have some time before the AI arrives, so let's focus on defense. We need to put some effort into this as like, right now, and I hope you are with me. What can we do to be ready for them? Well, I did go through a scientific education in my day life, so let me set out a draft build-order. That's really what this video series is meant to develop - a meta-build order, a discussion on how to systematically improve the highest level of play.
Let me give you some examples of the thinking:
My first hypotheses for the human-AI Achron mode meta is that all teams will end up playing random.
A first primitive model of the Game was enough to develop that theoretical prediction, with only two assumptions: The 1st Axiom is Perfect Play – So I'm saying random has an advantage, once you play all races perfect, because you have less pressure to scout. Elazers 4-Ravenger rush against random without Probe-Scout anyone? The only reason for not choosing random would be imbalance - Enter the 2nd Axiom - Perfect Balance. Every thing else being equal, and assuming god-like Blizzard perfectly nerving the Adept when ever it gets too much of those shady Protos wins, Random will rule the day. The experimental test for Oma's Random Hypothesis will take a long time though.
(1) http://www.nature.com/news/google-ai-algorithm-masters-ancient-game-of-go-1.19234 and https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol - There is plenty of deeper analysis out there (2) https://deepmind.com/blog/deepmind-and-blizzard-release-starcraft-ii-ai-research-environment/ (3) start at: https://en.wikipedia.org/wiki/Advanced_Chess https://en.wikipedia.org/wiki/Advanced_Chess (4) https://www.forbes.com/sites/gregsatell/2013/04/13/racing-with-the-machines/#643d08052997 which refers to https://www.amazon.com/Race-Against-Machine-Accelerating-Productivity-ebook/dp/B005WTR4ZI
Fighting Back Stage 1: Scouting out the AI's Weak Spots Edit
In general terms, what the AI's in Go and Chess are really bad at is Meta. I understand just about enough Go to extract that much information from the Go-Communities analysis of the Pro-Go games against Alpha Go. Don't get me wrong: The AI invented a new move that killed the current meta. Move 37 in the first game, where the human took a 2 hour break before making his move – that's like half of the entire game's thinking time. The moved altered something in the meta of Go. But the humans where faster to start adapting the Meta, and that is the only way then can at least occasionally win a game. Changing number 37 is a bit like adding 6 burrowed Banelings in the mineral line of your opponents 3rd before he takes it. A really small detail in the bigger picture. But it may be enough to break the balance in favor of Zerg when used in an otherwise perfectly executed hydra-ling-bane in ZvP. Once that piece is broken, all the logic about the early-game strategy leading to that point also changes.
And that point is where we humans still prevail: Adapting the Meta. If I really understand enough Go-lingo, I believe that this is exactly how the only human win in the 4-1 series with Lee was achieved: He change up something in the early game to a new move that was considered impossible long ago. But then somehow made it work, by using a variation of the infamous 37. This weakness makes sense when looking at the technical way the AIs the game at the moment: They need a very, very large library of games to get going with their reinforcement learning. We are still ahead when it comes to creating a small number of viable new game styles in that library as soon a new change pops up. That is because we have a good theory of how the game works, while the AI is only experimenting based on re-enforcement learning. Or can they develop theories? I don't even know what that would mean… Anyway, for now that is our only chance to keep winning. Or at least to have a continued role in AI-Human Archon games. So what can we practically do to achieve that?
We need to speed up on improving the Meta.
Please excuse me being all sciency about it and let me briefly define a bit more formally what is commonly referred to as the “Meta”. The terms is generally used to describe the expected reactions of your opponent and outcomes of strategic confrontations. It can also be called the viable decision tree. For example, you can say: “Battlecruisers are currently not in the meta in TvT”, which means you won't see battlecruisers in like 95% of Grand-Master game winners in TvT. The meta is composed of a set of build-orders for each race, which are all connected with a set of “reasonable” responses. To make it work out formally, I need to use a slightly bigger definition of “build order” including certain non-building “meta-moves” like purposefully engaging in a non-efficient trade, or base-trading in the definition to keep it more broad.
But how can we get more systematic about Meta Progress'
That is easier said than done… Aren't we all trying all the time to get ahead in the Meta? I'd even claim that that is what makes Starcraft so great and let's us stick with the same fucking game for decades...? We are doing it, yes, and we are even quite advanced at it, so we have a language to talk about, as used in this forum. While I'm sure there is a lot of talk about exactly the kind of weird move like 6 burrowed banelings in your opponents third mineral line before he takes it, we do not take records or systematize how we understand Starcraft in a unified theory. We don't have a real “Theory of Starcraft” that lists all the relevant factors. Chess and Go both have such theoretical frames, they are used for example to create the scoring scheme of a mid-game situation, and are really helpful to figure out new crazy moves. Or, the be more precise: There is a pretty advanced theory out there in the community that we somehow share so we can talk about these things and know a lot, but we never formalized that theory. The AI's will learn all that about the current Meta from our previous games and start cornering us due to their better execution. So how can we as a whole human community improve our meta?
Developing the Science of Starcraft
One way to formalize that talk would be to develop a code that allows us to talk faster. We have an informal code that is really sophisticated already – like in when I make a statement like: “Can you believe it? Stats just won after seven minutes in the finals against Gumiho, when he did the Void Ray & Basetrade move that Strange had used in that “Strange vs. Polt” game”. Most readers in this forum will have understood right away which move in which game I'm referring to, even if it's already more than a year old. I would argue that what Strange did in Strange vs. Polt was a potentially Meta altering event. There are many other such games, so I'm trying to formalize what he said here. A more recent one would be “Neebs Archon Drop”, but I'll stick to Strange vs. Polt as an example for now. The more subtly we can improve the Meta, the better. Defending Archon Drops with Roaches at perfect execution is something so clearly possible that the AI would probably find that by just experimenting a bit beyond a fully standardized, perfect execution ZvP game.
We already have the machine-readable formats of replays and game stats, and our informal language to theorize about possible SC2 is pretty good so far. Saying “Stats just won the GSL against Gumhio by initiating a base trade against Terran with Void Rays at 6 minutes” works pretty good. Probably that game would also end up the Hall of Fame of crazy games, so that would make it “Stats won in (Stats;Gumiho; 6:00;BC17; GF), because he did…what? Maybe “BaseTrade in PvT with Void-Rays at 6 minutes” – codified that would make it “Strategy in MatchUp in BuildOrderDeviation from Meta”. One thing we might want to do in order to improve our understanding is draw up a Meta-Decsion Tree, with verbal reasoning why certain things will or won't work at perfect execution. But that's only one part of doing science:
Deliberately Experimenting
But as all sciences, Starcraft will for now be an experimental driven science. We have no way of telling theoretically if (BT; PvT;VoidR; 600) is opening up a new avenue for P to get a larger % in wins or not. I doubt it can, please excuse the silly example. Currently, our way to find out is competitive play of pros, who play at our best approximation to perfect execution.
But the AI will have perfect execution and will be able to trial-and-err much faster than we can. What can we do?
We need to search the SC Meta for more strange, but viable games like (Strange; Polt; 15:00; DH16; RO32) in a smart and efficient way. And you know who probably would like to see more games like that? The casters and the viewers. I didn't check the numbers, but Strange vs. Polt pretty much was the most watched Starcraft game ever. Unfortunately games like that are rare, and moves like Strange's VoidRay Basetrade againt Terran even rarer. Pro-gamers cannot rely on the scientific axiom of “perfect execution” that will become a reality as soon as the AIs hit, so they can't just do something they never tried under high pressure to win a lot of money. We already see a lot more experimentation among the Pro's non-price games, which is what made todays Meta so sophisticated. But could there be a way to push our Pro's to focus less on execution and more on Meta? They will have too, as soon as the Robo-Hybrids enter the ladder and pure executing skill will lose value first. So let me just propose the first Starcraft research policy:
A crowd-sourced award for the coolest game.
As the crowd apparently really, really liked Polt vs. Strange, and the casters and you-tubers even made actually more money because of that particular game, we could launch a second reward to the biggest tournaments. One that is rewards the player who made the most unorthodox move work in all the matches. You can vote based on how much money you put in & everybody who chips in a minimum amount gets all the replays.… That, and the crowd, should be all it takes to get one, or maybe two, very important things in Starcraft II:
1) Better chances against the robots
2) More Epic Games
This is it for now from Oma Morkie's Starcraft Science Blog, but there is already my first policy recommendation.
|
Give up, AlphaSC2 already won the last VSL and is on its way to conquer the Starleagues. Resistance is futile.
|
We won't win, I agree with you there. But can a Human-AI Archon beat a pure AI?
|
Even in chess, the Human-AI Centaurs are still winning...
|
In SC2, any advanced AI is gonna mop the floor with human players. SC2 is very, very different from Chess and Go though, so I am still not sure that the AlphaGo team has much of an advantage from conquering Go. Chess and Go have the peculiar property that a single move can influence the game very unpredictably far down the road and that it also isn't straightforward, what is the best move at any given position and which move is "winning". SC2 (or BW), as much as people want to pretend that they are a "strategic chess", they are, in fact, mostly games of local tactical skill. You can see immediately if a fight helped you or harmed you, you can have tangible short-term goals that are undoubtedly positive for you etc. This makes it actually less suitable for a "play and learn" AI approach which was so successful with Go, because such a strategy would waste incredible amount of resources of going through useless space of moves that can be eliminated based on short-term judgement. However if an AI capable of short-term judgement is constructed and those steps are put together in a manner of a (very simple) overall Go/chess approach, such an AI would be unbeatable. It's not just about the micro, but about the fact that such AI should always be able to judge the outcome of a fight and never take losing fights. How do you beat someone who never loses armies? I think an optimal AI game of SC2 looks extremely different from what we are used to see from human players, it's basically a clusterfuck of small armies trying to outposition each other for a local advantage or surround that makes the smaller army unable to retreat - or the AI will discover some unbeatable timings and that will be all, that's also a possibility.
|
On July 12 2017 16:34 OmaMorkie wrote:
That is because we have a good theory of how the game works, while the AI is only experimenting based on re-enforcement learning. Or can they develop theories? I don't even know what that would mean… Anyway, for now that is our only chance to keep winning.
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.
|
[QUOTE]On July 12 2017 17:03 opisska wrote:
Depends a bit on whether APM limits are introduced. Without APM limits, I agree with you, it will be a reaper-micro freakshow. But with strict APM limits, things could get a lot more interesting.
While the micro decisions have less influence, build orders, base-setup etc. are still really complex. Not sure if it can compare to Go (again, under the assumption of perfect execution, think of two Pro's playing Archon at Very Slow game speed).
One way or another, improving on the Meta seems like the only role for humans in competitive Starcraft.
I guess I should also change the title - It's really not about beating the AI (no chance), but about teaming up with it and become better than pure AI. What Kasparow did for Chess, but for SC.
|
[QUOTE]On July 12 2017 17:05 Mendelfist wrote:
[QUOTE]On July 12 2017 16:34 OmaMorkie wrote:
I would argue that Chess is the better reference of where the game is going long term. Human / AI Hybrids still win there, and I believe they will soon start showing up in Go.
I agree, beating an AI without aid is impossible, but I would take a bet that (at least for a few more years), Human-AI Archons will beat pure AI.
|
The idea of collectively focusing more on Meta is meant to improve our play style for the future of AI-Human Archon tournaments with APM limit... Which I expect to be absolutely awesome to play and watch.
|
I was not talking about APM. Even with limited actions, the ability of the AI to make calculated decisions correctly at each point will eventually completely turn the gamr upside down. I don't think there will be much room for any "thinking" left. How many actual strategic decisions are there during one game? How many situations where one action isn't strictly better? I'd say just a handful.
|
France12888 Posts
On July 12 2017 18:19 opisska wrote:
I was not talking about APM. Even with limited actions, the ability of the AI to make calculated decisions correctly at each point will eventually completely turn the gamr upside down. I don't think there will be much room for any "thinking" left. How many actual strategic decisions are there during one game? How many situations where one action isn't strictly better? I'd say just a handful.
Most decisions depends on your ability to micro (at least for terran), so I don't think there will ever be a fair way to make AI play against humans, the game has not that much strategic depth afaik, but maybe I am underestimating our ability to make inferences...
|
On July 12 2017 16:41 OmaMorkie wrote:
Even in chess, the Human-AI Centaurs are still winning...
You can call this "winning", but what happens in reality is that the AI does almost all of the work, only when it evaluates two moves equally does human judgment come into play to break the tie. The same situation in SC2 it would be like a hybrid where the human player makes high level build order decisions while the AI actually plays the game, including micro / macro / tactics. Basically, it's totally humiliating for humans to participate in as a competitive activity, which is why centaur chess is not very well respected.
And theoretically, one can say that AI + something > AI. That is to say, as long as humans have some marginallly useful input it will improve the AI, but this will be true even if this input is essentially negligible.
|
On July 12 2017 16:34 OmaMorkie wrote:
A really small detail in the bigger picture. But it may be enough to break the balance in favor of Zerg when used in an otherwise perfectly executed hydra-ling-bane in ZvP.
What? ZvP is already in favour fo zerg w/o borrowed banelings. for god's sake...
|
On July 12 2017 17:03 opisska wrote:
It's not just about the micro, but about the fact that such AI should always be able to judge the outcome of a fight and never take losing fights. How do you beat someone who never loses armies? I think an optimal AI game of SC2 looks extremely different from what we are used to see from human players, it's basically a clusterfuck of small armies trying to outposition each other for a local advantage or surround that makes the smaller army unable to retreat - or the AI will discover some unbeatable timings and that will be all, that's also a possibility.
This used to be a contentious topic of discussion in BW vs SC2 threads, the degree to which the game encourages you to engage. The fear was that unlimited selection, improved pathfinding (responsiveness, movement) and clumping of units, in a world where both sides had powerful splash attacks, would prohibitively discourage engagements. One wrong step would mean your entire army would evaporate and you couldn't take that risk, and because of unlimited selection you could easily direct your entire army to avoid engagements. Supposedly, in BW you could more easily withdraw from an engagement without severe losses (as your army is more spaced out), but you were also more likely to get drawn into an engagement and have some sort of skirmish because controlling your army was so hard.
I can't evaluate whether SC2 suffers from this, I guess there are also countervailing forces to negate this dynamic so it's not a clear cut case. One thing that happens is that players with some reason to will push for engagements, and it's not always possible to avoid them. And pro players are good enough to spread out their army, be careful and not overcommit. So this dynamic might not be that obvious.
What you can say is that the AI should be very confident in its micro (as it will be near perfect), so it should always push for engagements, even if it seems to human strategically ill-advised. This will immediately break the meta and give the AI a likely insurmountable strategic advantage. On the other hand, in AI vs AI games you might see the exact opposite, both sides will never engage because defensive set-ups are so powerful etc. It's hard to predict, and it depends also on how they "grow" the AI. It it only plays itself it might severely overestimate its opponent's ability to micro. I think it will be interesting to witness this development, both in relation to AI behavior and within the context of these fundamental aspects of SC2 design. The AI gives a fresh perspective and a potentially higher caliber of play, so it might break a lot of assumptions.
|
4713 Posts
I think its quite premature to evaluate the potential success of a AlphaSC2 given that SC2 and GO are completely different games with different rules and parameters.
For one Go is a game with complete information and turns while SC2 is a game a game with incomplete information set in real time.
It is true that if the game was about pure mechanics humans would probably be crushed, however the aim of the Alpha crew is to make a AI that can out-think a human.
And the out-thinking part is much, much tougher when the information you have is incomplete. Thus the AI needs to learn the specific tells of a certain strategy, it needs to learn how to scout for said tells, when to scout and how to potentially identify miss-information from elaborate cheeses.
After the scouting is out of the way the AI also needs to be able to learn how to develop counter-strategies and tactics of its own, how to evaluate its losses in the grand scheme of its strategy and how to evaluate its position vs that of the human.
After the above are ironed out the AI needs to learn how to execute its plan.
I'm not saying its not possible, but I think it'll take the AI team a solid 5 years at the minimum to reach the level of refinement needed to beat humans in SC2.
Heck it might possibly take it even 10.
|
On July 12 2017 19:28 Destructicon wrote:
I think its quite premature to evaluate the potential success of a AlphaSC2 given that SC2 and GO are completely different games with different rules and parameters.
For one Go is a game with complete information and turns while SC2 is a game a game with incomplete information set in real time.
It is true that if the game was about pure mechanics humans would probably be crushed, however the aim of the Alpha crew is to make a AI that can out-think a human.
And the out-thinking part is much, much tougher when the information you have is incomplete. Thus the AI needs to learn the specific tells of a certain strategy, it needs to learn how to scout for said tells, when to scout and how to potentially identify miss-information from elaborate cheeses.
After the scouting is out of the way the AI also needs to be able to learn how to develop counter-strategies and tactics of its own, how to evaluate its losses in the grand scheme of its strategy and how to evaluate its position vs that of the human.
After the above are ironed out the AI needs to learn how to execute its plan.
I'm not saying its not possible, but I think it'll take the AI team a solid 5 years at the minimum to reach the level of refinement needed to beat humans in SC2.
Heck it might possibly take it even 10.
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. I think if you compare chess and go with starcraft, then even if starcraft is the more complex game, and is the game where humans might have the greatest advantage strategically, it is also the game where humans might be weakest tactically.
|
|
|
On July 12 2017 19:28 Destructicon wrote:
I think its quite premature to evaluate the potential success of a AlphaSC2 given that SC2 and GO are completely different games with different rules and parameters.
For one Go is a game with complete information and turns while SC2 is a game a game with incomplete information set in real time.
It is true that if the game was about pure mechanics humans would probably be crushed, however the aim of the Alpha crew is to make a AI that can out-think a human.
And the out-thinking part is much, much tougher when the information you have is incomplete. Thus the AI needs to learn the specific tells of a certain strategy, it needs to learn how to scout for said tells, when to scout and how to potentially identify miss-information from elaborate cheeses.
After the scouting is out of the way the AI also needs to be able to learn how to develop counter-strategies and tactics of its own, how to evaluate its losses in the grand scheme of its strategy and how to evaluate its position vs that of the human.
After the above are ironed out the AI needs to learn how to execute its plan.
I'm not saying its not possible, but I think it'll take the AI team a solid 5 years at the minimum to reach the level of refinement needed to beat humans in SC2.
Heck it might possibly take it even 10.
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.
|
On July 12 2017 16:38 Ej_ wrote:
Give up, AlphaSC2 already won the last VSL and is on its way to conquer the Starleagues. Resistance is futile.
ha ha ha !
|
[QUOTE]On July 12 2017 19:28 Destructicon wrote:
Yes, I agree - and the AI will even have to learn things like anticipating that an active tech lab on a Starport may be fake and be ready for a Raven instead of a cloak Banshee. I.e. it will need to learn how to play mind games, which makes it so exciting.
I would argue that hybrid SC2 would still be a hell lot of fun exactly for the complexities pointed out by Destructicon, with the AI taking care of Micro and Build-Order, while the human makes all the cheesy decisions, mind-games, adapts the play style to new maps and so on.
|
|
|
|
|
|
EPT
