|
On March 28 2016 15:39 JimmyJRaynor wrote:Show nested quote +On March 28 2016 13:58 kingjames01 wrote:On March 28 2016 13:50 JimmyJRaynor wrote:On March 28 2016 13:46 kingjames01 wrote:On March 28 2016 13:39 JimmyJRaynor wrote:
without heuristic functions i think AlphaStar will have trouble "learning" the game.
and with heuristic functions it'll play with the style of dictated by them. That was the prevailing stance before AlphaGo fought Lee Sedol... It didn't hold up to reality though. who is the winner? the best Chess AI with heuristics or Alpha? That's a weird way to defend a point... Are they playing Chess, Go or StarCraft? here is why Alpha will have a tough time without heuristics in a game like Chess or Starcraft https://webdocs.cs.ualberta.ca/~cdavid/pdf/aiide12-combat.pdfyears from now .. who knows.. maybe no heuristics are required. but for now.. Alpha will have a tough time.
Did you even read the paper? You can't just skim the abstract for keywords and post it if the article doesn't actually support your stance.
This paper outlines an alpha-beta pruning method which can be used in small group combats. The only relevant text in the entire article is: These advantages come at a cost, namely increased runtime and/or memory requirements. Therefore, it can be challenging to adapt established search algorithms to large scale realtime decision problems, e.g. video games and robotics.
I'm not sure if I'm missing your real point or how familiar you are with the internal workings of AlphaGo. AlphaGo uses a policy network to decide which moves to consider. This plays the same role as heuristic-based pruning. The difference between AlphaGo's and previous Go playing AIs is that AlphaGo's policy network does not rely on domain knowledge (ie. special rules about Go) explicitly programmed in by the developers. AlphaGo learned how to play Go using an initial seed of sample human games and then it was set to play against itself.
|
Russian Federation1016 Posts
Sniper should be a consultant for Google. Just imagine it: most evil and loathsome zerg helps an AI to defeat humanity.
|
its a big misstake to chose sc2 over bw btw, one is complex strategical game and sc2 is just a pure tactical game dissapointed in google here
|
We should really hype this as a community and help it become a mainstream Machines vs Humans event (for the sake of gaining a bigger fanbase for SCII at least).
Also, if people understand the depth and layers of a complex game such as SCII, people would be more inclined on giving it a try, you know, people like challenges.
|
On March 28 2016 23:44 alukarD wrote:
We should really hype this as a community and help it become a mainstream Machines vs Humans event (for the sake of gaining a bigger fanbase for SCII at least).
Also, if people understand the depth and layers of a complex game such as SCII, people would be more inclined on giving it a try, you know, people like challenges.
This is years away from happening, if it does happen.
|
On March 28 2016 19:32 Mendelfist wrote:Show nested quote +On March 28 2016 19:15 MockHamill wrote:
Starcraft is so much more complex than Go it is absurd. Basically the amount of permutations in the first 3 minutes of Starcraft is larger then any match of Go. As someone who has played chess (no rank), Go (10 kyu) and Starcraft (platinum) at low amateur level I say you are wrong. At the strategic level Go is enormously more complex and deep than both chess and Starcraft. There is absolutely no comparison. Go completely dwarfs any game that I have come in contact with. Measuring complexity in "number of permutations" is not meaningful. How many permutations are there in a game of billiard? Do you think it would be hard for a computer?
You miss the point.
Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable. The dimensionality of the space of all possible moves for starcraft is orders, and orders of magnitude larger than that of Go. For instance, in Go, at every turn you can move one unit. You have N units. Thus each turn, you have N possible choices of which units to move. In Starcraft, if you have N units, there are 2^N possible choices of which units to select. This includes buildings. In Go, the average number of turns is in the tens. In Starcraft there's 24 frames per second, and each frame has a possible set of actions. Go is a 19x19 board, Starcraft is a thousands by thousands board.
With Go, it is reasonable (though probably still extremely computationally expensive) to set your dimensionality of the data set as all possible moves, then take games as a data set, and just straightforwardly train a neural network to that. I'm sure the Alpha Go team was more sophisticated than that, and took a crazy amount of clever shortcuts. But most of the cleverness there came from finding ways to conserve computational resource, not from trying to think of the deep strategies. How deep the strategies are in Go had virtually nothing to do with the difficulty of their task. Starcraft will be insurmountably harder to train than Go, regardless of whether or not its strategies are deeper (and I agree with you, Go is much deeper in this regard). Like I said, with Go, even though the team was more clever, it is feasible to brute force train a NN on the exact moves. With Starcraft that isn't even remotely feasible. Not with the greatest super computers in the world. The number of possible moves scales exponentially with the number of units you have, and again exponentially with the number of frames the game takes. They'll have to take extremely different and far more clever strategies in training a SC2 AI, and it will be a hell of a lot more difficult.
I think a lot of people here are thinking too much about this as if it were like teaching a human. If that were so, then yes, Go would be a lot harder to teach the strategies to. But the problem of taking the best strategy when you have in mind the set of all strategies you can play is pretty much solved in machine learning already. Any maximum likelihood estimator can make a better decision than pretty much any human, as long as it has the right inputs. Giving it the set of strategies it can use is the real problem, and much, much harder to train for an AI than for a human. And ridiculously harder to train for Starcraft than for Go.
|
You can easily cut SC2's difficulty by giving the computer a non-Korean visa.
|
On March 28 2016 23:50 The Bottle wrote:Show nested quote +On March 28 2016 19:32 Mendelfist wrote:On March 28 2016 19:15 MockHamill wrote:
Starcraft is so much more complex than Go it is absurd. Basically the amount of permutations in the first 3 minutes of Starcraft is larger then any match of Go. As someone who has played chess (no rank), Go (10 kyu) and Starcraft (platinum) at low amateur level I say you are wrong. At the strategic level Go is enormously more complex and deep than both chess and Starcraft. There is absolutely no comparison. Go completely dwarfs any game that I have come in contact with. Measuring complexity in "number of permutations" is not meaningful. How many permutations are there in a game of billiard? Do you think it would be hard for a computer? You miss the point. Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable. The dimensionality of the space of all possible moves for starcraft is orders, and orders of magnitude larger than that of Go. For instance, in Go, at every turn you can move one unit. You have N units. Thus each turn, you have N possible choices of which units to move. In Starcraft, if you have N units, there are 2^N possible choices of which units to select. This includes buildings. In Go, the average number of turns is in the tens. In Starcraft there's 24 frames per second, and each frame has a possible set of actions. Go is a 19x19 board, Starcraft is a thousands by thousands board. With Go, it is reasonable (though probably still extremely computationally expensive) to set your dimensionality of the data set as all possible moves, then take games as a data set, and just straightforwardly train a neural network to that. I'm sure the Alpha Go team was more sophisticated than that, and took a crazy amount of clever shortcuts. But most of the cleverness there came from finding ways to conserve computational resource, not from trying to think of the deep strategies. How deep the strategies are in Go had virtually nothing to do with the difficulty of their task. Starcraft will be insurmountably harder to train than Go, regardless of whether or not its strategies are deeper (and I agree with you, Go is much deeper in this regard). Like I said, with Go, even though the team was more clever, it is feasible to brute force train a NN on the exact moves. With Starcraft that isn't even remotely feasible. Not with the greatest super computers in the world. The number of possible moves scales exponentially with the number of units you have, and again exponentially with the number of frames the game takes. They'll have to take extremely different and far more clever strategies in training a SC2 AI, and it will be a hell of a lot more difficult.
Though I agree with your basic thoughts, I have to point out that it is NOT POSSIBLE to brute force solve the game of Go. There are 10^761 possible board states. That is a number beyond our capability to imagine. With current technologies it is impossible to store the entire game tree.
|
Starcraft 2 doesn't even deserve to have Starcraft in its name. I'd be surprised if Deepmind didn't get the better of the pros, and it will be a black eye for this game. This game used to be chess like and strategical. Now it's a game with 0 options that lives on the past of BW. Should have strategy stripped from the genre if SC 2 pros lose. Deepmind should challenge pros in Z vs P. I wonder how long it would take Deepmind to learn that every opponent will go phoenix into mass immortals.
A true test of AI here, ladies in gents. Is tac tac toe next? Look out world. My only question is will the AI bm the creators of this boring fraudulent strategy game. I'd like to see Deepmind BM someone for wasting its time.
User was temp banned for this post.
|
On March 29 2016 00:02 kingjames01 wrote:Show nested quote +On March 28 2016 23:50 The Bottle wrote:On March 28 2016 19:32 Mendelfist wrote:On March 28 2016 19:15 MockHamill wrote:
Starcraft is so much more complex than Go it is absurd. Basically the amount of permutations in the first 3 minutes of Starcraft is larger then any match of Go. As someone who has played chess (no rank), Go (10 kyu) and Starcraft (platinum) at low amateur level I say you are wrong. At the strategic level Go is enormously more complex and deep than both chess and Starcraft. There is absolutely no comparison. Go completely dwarfs any game that I have come in contact with. Measuring complexity in "number of permutations" is not meaningful. How many permutations are there in a game of billiard? Do you think it would be hard for a computer? You miss the point. Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable. The dimensionality of the space of all possible moves for starcraft is orders, and orders of magnitude larger than that of Go. For instance, in Go, at every turn you can move one unit. You have N units. Thus each turn, you have N possible choices of which units to move. In Starcraft, if you have N units, there are 2^N possible choices of which units to select. This includes buildings. In Go, the average number of turns is in the tens. In Starcraft there's 24 frames per second, and each frame has a possible set of actions. Go is a 19x19 board, Starcraft is a thousands by thousands board. With Go, it is reasonable (though probably still extremely computationally expensive) to set your dimensionality of the data set as all possible moves, then take games as a data set, and just straightforwardly train a neural network to that. I'm sure the Alpha Go team was more sophisticated than that, and took a crazy amount of clever shortcuts. But most of the cleverness there came from finding ways to conserve computational resource, not from trying to think of the deep strategies. How deep the strategies are in Go had virtually nothing to do with the difficulty of their task. Starcraft will be insurmountably harder to train than Go, regardless of whether or not its strategies are deeper (and I agree with you, Go is much deeper in this regard). Like I said, with Go, even though the team was more clever, it is feasible to brute force train a NN on the exact moves. With Starcraft that isn't even remotely feasible. Not with the greatest super computers in the world. The number of possible moves scales exponentially with the number of units you have, and again exponentially with the number of frames the game takes. They'll have to take extremely different and far more clever strategies in training a SC2 AI, and it will be a hell of a lot more difficult. Though I agree with your basic thoughts, I have to point out that it is NOT POSSIBLE to brute force solve the game of Go. There are 10^761 possible board states. That is a number beyond our capability to imagine. With current technologies it is impossible to store the entire game tree.
The number of board states is irrelevant. It's the number of possible moves you can make, given a board state, that's relevant. That's all you really need to give for a training AI. My issue is the dimensionality of a training data set of moves. The number of dimensions of that set is the number of possible moves. For Go, with a given board state, it's actually not that much. (In the hundreds, I'd say.) For Starcraft, it's absurdly high. Think of the easiest machine learning problem (training a linear model in one dimension). The set of possible parameters is infinite, but there are only two parameters, and thus it's really quick and easy to train.
|
On March 29 2016 00:05 playa wrote:
Starcraft 2 doesn't even deserve to have Starcraft in its name. I'd be surprised if Deepmind didn't get the better of the pros, and it will be a black eye for this game. This game used to be chess like and strategical. Now it's a game with 0 options that lives on the past of BW. Should have strategy stripped from the genre if SC 2 pros lose. Deepmind should challenge pros in Z vs P. I wonder how long it would take Deepmind to learn that every opponent will go phoenix into mass immortals.
A true test of AI here, ladies in gents. Is tac tac toe next? Look out world. My only question is will the AI bm the creators of this boring fraudulent strategy game. I'd like to see Deepmind BM someone for wasting its time.
I agree completely with you! The same happened with chess and now go, these games aren't worthy of the title "strategy game" when an AI can beat humans.
I heard the first go players are already retiring and starting to play bw because they read the bw forums of tl, they came to the conclusion (based on the intelligent posts there) that bw is the epitome of strategic/tactical complexity.
Preach it!
|
On March 29 2016 00:05 The Bottle wrote:Show nested quote +On March 29 2016 00:02 kingjames01 wrote:On March 28 2016 23:50 The Bottle wrote:On March 28 2016 19:32 Mendelfist wrote:On March 28 2016 19:15 MockHamill wrote:
Starcraft is so much more complex than Go it is absurd. Basically the amount of permutations in the first 3 minutes of Starcraft is larger then any match of Go. As someone who has played chess (no rank), Go (10 kyu) and Starcraft (platinum) at low amateur level I say you are wrong. At the strategic level Go is enormously more complex and deep than both chess and Starcraft. There is absolutely no comparison. Go completely dwarfs any game that I have come in contact with. Measuring complexity in "number of permutations" is not meaningful. How many permutations are there in a game of billiard? Do you think it would be hard for a computer? You miss the point. Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable. The dimensionality of the space of all possible moves for starcraft is orders, and orders of magnitude larger than that of Go. For instance, in Go, at every turn you can move one unit. You have N units. Thus each turn, you have N possible choices of which units to move. In Starcraft, if you have N units, there are 2^N possible choices of which units to select. This includes buildings. In Go, the average number of turns is in the tens. In Starcraft there's 24 frames per second, and each frame has a possible set of actions. Go is a 19x19 board, Starcraft is a thousands by thousands board. With Go, it is reasonable (though probably still extremely computationally expensive) to set your dimensionality of the data set as all possible moves, then take games as a data set, and just straightforwardly train a neural network to that. I'm sure the Alpha Go team was more sophisticated than that, and took a crazy amount of clever shortcuts. But most of the cleverness there came from finding ways to conserve computational resource, not from trying to think of the deep strategies. How deep the strategies are in Go had virtually nothing to do with the difficulty of their task. Starcraft will be insurmountably harder to train than Go, regardless of whether or not its strategies are deeper (and I agree with you, Go is much deeper in this regard). Like I said, with Go, even though the team was more clever, it is feasible to brute force train a NN on the exact moves. With Starcraft that isn't even remotely feasible. Not with the greatest super computers in the world. The number of possible moves scales exponentially with the number of units you have, and again exponentially with the number of frames the game takes. They'll have to take extremely different and far more clever strategies in training a SC2 AI, and it will be a hell of a lot more difficult. Though I agree with your basic thoughts, I have to point out that it is NOT POSSIBLE to brute force solve the game of Go. There are 10^761 possible board states. That is a number beyond our capability to imagine. With current technologies it is impossible to store the entire game tree. The number of board states is irrelevant. It's the number of possible moves you can make, given a board state, that's relevant. That's all you really need to give for a training AI. My issue is the dimensionality of a training data set of moves. The number of dimensions of that set is the number of possible moves. For Go, with a given board state, it's actually not that much. (In the hundreds, I'd say.) For Starcraft, it's absurdly high. Think of the easiest machine learning problem (training a linear model in one dimension). The set of possible parameters is infinite, but there are only two parameters, and thus it's really quick and easy to train.
Upon re-reading your previous post, I see that you were in fact discussing possible moves, not possible board states.
|
On March 28 2016 23:50 The Bottle wrote:
Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable.
The problem is only more complex if you count every possible game state as a "possible move". That's a ridiculous approach to the problem, in the same way as it's ridiculous trying to do it in billiard, which you didn't comment. I have no notion that Starcraft is an easy problem for an AI, or even easier than Go.
This is a different problem, in the same way that Go is a different problem than chess, and that's no doubt why DeepMind think it's interesting. Starcraft is much more similar to real world problems than Go. The number of possible moves are for all practical purposes infinite, and there is no point in putting a number on it or saying that it has anything to do with complexity or how hard it is. It's a meaningless number. If we doubled the precision for all game state variables, the number of possible game states would increase dramatically. Does that mean that the problem becomes much harder? No. It's exactly the same problem. The infinite state space makes it belong to a different class of problems, an unsolved one. I'll not argue that.
|
On March 29 2016 01:12 Mendelfist wrote:Show nested quote +On March 28 2016 23:50 The Bottle wrote:
Number of permutations is indeed important, in fact it's the most important factor when it comes to making a self learning AI. You still have to actually use the inputs of a game in order to train the gaming algorithm. The data set is way larger, and the task of purely training a neural network on this is highly intractable. The problem is only more complex if you count every possible game state as a "possible move". That's a ridiculous approach to the problem, in the same way as it's ridiculous trying to do it in billiard, which you didn't comment. I have no notion that Starcraft is an easy problem for an AI, or even easier than Go. This is a different problem, in the same way that Go is a different problem than chess, and that's no doubt why DeepMind think it's interesting. Starcraft is much more similar to real world problems than Go. The number of possible moves are for all practical purposes infinite, and there is no point in putting a number on it or saying that it has anything to do with complexity or how hard it is. It's a meaningless number. If we doubled the precision for all game state variables, the number of possible game states would increase dramatically. Does that mean that the problem becomes much harder? No. It's exactly the same problem. The infinite state space makes it belong to a different class of problems, an unsolved one. I'l not argue that.
I'm not talking about the number of possible game states. That's not important for a machine learning algorithm. Number of possible moves you can make is important. That is, the way you encode a particular action. This is essential for making a training data set for your algorithm to learn.
Your billiard example doesn't work here, because there's no self-learning AI algorithm for billiard, at least not that I know of. (If there is, let me know.) The ones that exist simply do an on-the-spot calculation for each move. And while it is intractable to consider all possible moves, it takes a clever shortcut (one example is taking a large sample of N moves, properly separated in the angle-power space by use of k-means clustering, and picking the best one). It's scripting, similar to what Starcraft AI does now. The other problem with the billiard example is that each move can be completely encoded by three continuous variables (position along the rectangle, angle of shot, and power). Even though they're continuous, and have an infinite set of values, encoding this into a data set of moves for a given game is extremely easy. It's a Nx3 matrix of values, where N is the number of turns and 3 is the dimensions of moves. (Although given the nature of billiard, I'm not sure self learning would be any better than scripting; probably worse, in fact.) There's no analogous simple reduction like that in Starcraft.
The reason self learning is so much harder than scripting, is because you have to be so much more meticulous about giving your algorithm the necessary degrees of freedom in the former case. Because you can't feed it any particular strategy, it has to be able to infer those strategies from knowing the moves it can make, and using historic data to calculate which combination of moves maximizes the probability of success. They will have to find clever ways to transform the game input data in order to remove redundancies, and coarsen the scale of discrete moves. I'm sure they did something like that with Go already, but it will be substantially harder for SC2. You say it's just a different problem, sure. But a much, much harder problem, one I'm not quite sure they'll solve, even knowing their success with Go.
|
Why is everyone discussing APM, shouldn't EAPM be a more suitable metric. I doubt the AI will spam keys so even if you cap APM it will still outmatch a human easy. Because 500 APM from a human could be 200 EAPM while 500 APM from an AI could be actually be 500 EAPM and thus the AI are allowed double the actual actions than a human.
|
1001 YEARS KESPAJAIL22272 Posts
One aspect that people also misunderstand in comparing Go and SC2 is that of input/output.
For Go, it is relatively simple. The input is the game board at a given game state, which is all the current pieces in their positions. The output is a position on the board that corresponds to the move done once per turn. The cycle goes back and forth until the game is resolved. This means that the AI and the human have exactly the same input/output, and the game is 100% strategy.
For SC2, will that same stipulation hold? Will they require the AI to use the same input/output as a human being? I believe they should, otherwise it's too big an advantage for the AI and it will no longer be a test of strategy, but rather the efficiency of input/output that they are using.
First, the input. There's the screen, minimap, control groups, unit selection, command card, supplies. That's a lot of information that the AI must learn to interpret. They aren't going to program the AI what these things mean and how to use them, the AI has to learn that on its own. There's also the limitation that the AI can only look at one sector of the map at any given time. It has to learn to how to do that as well.
For output, it's just as difficult. How would they program the AI so that its outputs mimic that of a human being? I don't even know how they'd do that. Limiting APM is one way, but that's an artificial cap and I don't know if that's going to be effective. This is a problem they have to figure out and I doubt anyone here has an answer. The way they output has to be as similar as possible. Even then a machine will still have the advantage of 100% accuracy (perfect force fields, perfectly timed casting, perfect micro given APM restrictions).
The AI has to learn this input/output system on its own, from scratch. There is a lot more to consider compared to Go. This obstacle alone would be an amazing one to solve.
This is an important part of the discussion because the purpose of this test is to show advances in AI learning through strategy games. Unless the game is as close to 100% strategy as possible like Go, it's pointless to play.
|
On March 28 2016 19:32 Mendelfist wrote:Show nested quote +On March 28 2016 19:15 MockHamill wrote:
Starcraft is so much more complex than Go it is absurd. Basically the amount of permutations in the first 3 minutes of Starcraft is larger then any match of Go. As someone who has played chess (no rank), Go (10 kyu) and Starcraft (platinum) at low amateur level I say you are wrong. At the strategic level Go is enormously more complex and deep than both chess and Starcraft. There is absolutely no comparison. Go completely dwarfs any game that I have come in contact with. Measuring complexity in "number of permutations" is not meaningful. How many permutations are there in a game of billiard? Do you think it would be hard for a computer?
Thank you for this post. It's so annoying to see so many people saying that Go is easy.
The reason why Go was such an interesting game for AI to tackle is precisely that it has so many possible moves that backtracking sucks too much.
|
On March 29 2016 02:08 lichter wrote:
One aspect that people also misunderstand in comparing Go and SC2 is that of input/output.
For Go, it is relatively simple. The input is the game board at a given game state, which is all the current pieces in their positions. The output is a position on the board that corresponds to the move done once per turn. The cycle goes back and forth until the game is resolved. This means that the AI and the human have exactly the same input/output, and the game is 100% strategy.
For SC2, will that same stipulation hold? Will they require the AI to use the same input/output as a human being? I believe they should, otherwise it's too big an advantage for the AI and it will no longer be a test of strategy, but rather the efficiency of input/output that they are using.
First, the input. There's the screen, minimap, control groups, unit selection, command card, supplies. That's a lot of information that the AI must learn to interpret. They aren't going to program the AI what these things mean and how to use them, the AI has to learn that on its own. There's also the limitation that the AI can only look at one sector of the map at any given time. It has to learn to how to do that as well.
For output, it's just as difficult. How would they program the AI so that its outputs mimic that of a human being? I don't even know how they'd do that. Limiting APM is one way, but that's an artificial cap and I don't know if that's going to be effective. This is a problem they have to figure out and I doubt anyone here has an answer. The way they output has to be as similar as possible. Even then a machine will still have the advantage of 100% accuracy (perfect force fields, perfectly timed casting, perfect micro given APM restrictions).
The AI has to learn this input/output system on its own, from scratch. There is a lot more to consider compared to Go. This obstacle alone would be an amazing one to solve.
This is an important part of the discussion because the purpose of this test is to show advances in AI learning through strategy games. Unless the game is as close to 100% strategy as possible like Go, it's pointless to play.
I would bet they won't even try to solve that. It's besides the point. They will probably just use the API and have the AI access unit positioning directly.
|
1001 YEARS KESPAJAIL22272 Posts
On March 29 2016 02:10 rockslave wrote:Show nested quote +On March 29 2016 02:08 lichter wrote:
One aspect that people also misunderstand in comparing Go and SC2 is that of input/output.
For Go, it is relatively simple. The input is the game board at a given game state, which is all the current pieces in their positions. The output is a position on the board that corresponds to the move done once per turn. The cycle goes back and forth until the game is resolved. This means that the AI and the human have exactly the same input/output, and the game is 100% strategy.
For SC2, will that same stipulation hold? Will they require the AI to use the same input/output as a human being? I believe they should, otherwise it's too big an advantage for the AI and it will no longer be a test of strategy, but rather the efficiency of input/output that they are using.
First, the input. There's the screen, minimap, control groups, unit selection, command card, supplies. That's a lot of information that the AI must learn to interpret. They aren't going to program the AI what these things mean and how to use them, the AI has to learn that on its own. There's also the limitation that the AI can only look at one sector of the map at any given time. It has to learn to how to do that as well.
For output, it's just as difficult. How would they program the AI so that its outputs mimic that of a human being? I don't even know how they'd do that. Limiting APM is one way, but that's an artificial cap and I don't know if that's going to be effective. This is a problem they have to figure out and I doubt anyone here has an answer. The way they output has to be as similar as possible. Even then a machine will still have the advantage of 100% accuracy (perfect force fields, perfectly timed casting, perfect micro given APM restrictions).
The AI has to learn this input/output system on its own, from scratch. There is a lot more to consider compared to Go. This obstacle alone would be an amazing one to solve.
This is an important part of the discussion because the purpose of this test is to show advances in AI learning through strategy games. Unless the game is as close to 100% strategy as possible like Go, it's pointless to play. I would bet they won't even try to solve that. It's besides the point. They will probably just use the API and have the AI access unit positioning directly.
I feel like that would be disingenuous to the purpose of this project. The amazing thing about Go is it's 100% a battle of the mind because input/output is the same regardless of computer or human. Unless they do that for SC2 as well, it won't be as meaningful as the breakthrough with Go.
|
On March 29 2016 01:33 The Bottle wrote:
I'm not talking about the number of possible game states. That's not important for a machine learning algorithm. Number of possible moves you can make is important. That is, the way you encode a particular action. This is essential for making a training data set for your algorithm to learn.
Why do you assume that the best way to solve this problem is to throw every single game state variable or pixel at a neural net and then hope that it somehow works out?
Your billiard example doesn't work here, because there's no self-learning AI algorithm for billiard, at least not that I know of.
Then imagine one. One that learns for example by self play. Do you really think that it would have a hard time finding "the right moves" just because they are infinite in number? Edit: And the question is not if it can be better than a script, which can be near perfect. The question is if you think it's a hard problem.
They will have to find clever ways to transform the game input data in order to remove redundancies, and coarsen the scale of discrete moves. I'm sure they did something like that with Go already, but it will be substantially harder for SC2. You say it's just a different problem, sure. But a much, much harder problem, one I'm not quite sure they'll solve, even knowing their success with Go.
Yes, THIS is the problem, and once you have done this the "number of possible moves" in the original problem is irrelevant. That only tells you that you have THIS problem on your hands and that you can't solve it by ordinary search algorithms. You have to find a way to reduce the original problem by levels of abstraction. I don't know if there is a way for an AI find these abstractions by itself. Maybe that's how Alphago works, In any case, I'm not in the least convinced that it's as hard as you are trying to make it sound. In the simplest form you can have an ordinary scripted bot that asks an AI for advice. "Attack now", "build sentries", "expand there" etc. Or you could throw every pixel at it, like you want. I don't think that would work. Or something in between. How about that?
|
|
|
|
|
|
EPT
