EPTStarCraft II: DeepMind Demonstration: Jan 24 - Page 3
| Forum Index > SC2 General |
Poopi
France12904 Posts
| ||
|
Loccstana
United States833 Posts
| ||
|
MockHamill
Sweden1798 Posts
| ||
|
deacon.frost
Czech Republic12129 Posts
On January 23 2019 16:47 Loccstana wrote: I hope we will get a Bo31 showmatch between deepmind and avilo! Nah, if the Deepmind is really good let it play Avilo so Avilo doesn't know he's playing it. And let us bet how many cheater calls will be made. Then we can give those money to some charity | ||
|
Lazzarus
Faroe Islands114 Posts
| ||
|
Grumbels
Netherlands7031 Posts
There are some interesting quirks with Leela. For instance, it's not capable of playing endgames efficiently, it seemingly aimlessly moves around, making moves that don't lose the advantage. It doesn't "get to the point". If an SC2 AI is built on the same concept, expect it to not be able to finish off games quickly and take an hour to mine out the entire map and build a fleet of random units to randomly move around the map. Another quirk of the project is that the algorithm uses not just the current move as input, but also the history of moves. This gives it some measure of what part of the board to pay "attention" to. It also means that if you give it a random position as input, without history, that it can't function. As far as I know Leela is useless in solving tactical puzzles and in handicap games without training it first. Leela also typically doesn't understand theory of endgames. It doesn't just play them weirdly, but it also doesn't grasp some almost mathematical ideas such as identifying a class of endgames that are drawn despite material imbalances (opposite color bishops, wrong color bishop). It's apparently also not better at fortress positions, where you have material disadvantages, but your position can't be cracked. There are some known positions like these, and it was hoped that neural networks would be better at them, and would be capable of reasoning that these are a special class of positions that require a different approach. But it doesn't really seem like it. Leela is also probably already better than Stockfish if you have bad hardware and no opening book. You can imagine that if there was a market for SC2 bots, that they could have opening books updated for every patch and which would have a team of people dedicated to keeping track of the meta and adding knowledge of it to the bot. But Deepmind's AI would use self-learning, i.e. only playing itself and developing its own meta. I don't know if that would make it easier or harder to beat as a human. I think the tree-search method for chess is bound to scale better with hardware than a neural network approach, given that chess is theoretically solvable with tree search. But this method would be useless for SC2, unless the AI uses some sort of abstraction of strategy and tries to think ahead. But I don't think you really need to think ahead in SC2 to get decent results. If you just react to your opponent and have perfect, bot-like control, you will win. | ||
|
Grumbels
Netherlands7031 Posts
Also, I heard a pro player say that an engine such as Leela would be less useful than Stockfish in preparing, because the latter is tactically superior, while the former is strategically superior. But humans are already good at strategy, they just need to make use of their tactical ability of engines to check their ideas and openings for tactical flaws. Leela's is unreliable because it doesn't have concrete reasons for preferring one move over the other, just a strategical intuition. Whereas Stockfish can instantly tell you if there are tactical problems with a move and produce a refutation. It might be the case that AlphaZero will remain a novelty for computer chess enthusiasts. Especially since Leela and AlphaZero run on TPU/GPU's, not CPU's, afaik, so if you want to use both locally, you have to invest in both a good graphic card and a good processor. | ||
|
alexanderzero
United States659 Posts
I think the tree-search method for chess is bound to scale better with hardware than a neural network approach, given that chess is theoretically solvable with tree search. This would suggest otherwise: ![[image loading]](https://storage.googleapis.com/deepmind-live-cms/images/AZ-Blog-Fig2-Search-Per-Decision.width-1500.png) Isn't go also theoretically solvable with a search tree? | ||
|
xongnox
540 Posts
Out-microing and out-multitasking everyone by playing 30.000 APMs and 100 screens/second is surely automaton-2000 impressive to watch one or two time, but is not very conclusive for the intelligence part. I guess they done it right and have set limiting factors as parameters (like 250/apms max, max actions per second, max screens per second, human-like time mouse movements, etc, etc. ) | ||
|
Grumbels
Netherlands7031 Posts
On January 23 2019 18:07 alexanderzero wrote: This would suggest otherwise: Isn't go also theoretically solvable with a search tree? AlphaZero claimed that their approach scaled well, iirc they had extremely good hardware for the recent rematch. On the other hand, there is some reason to doubt their work, since they might be more familiar setting up their own engine versus setting up Stockfish. Leela seems to do worse than Stockfish on good hardware / longer time controls. But I'm not sure, since e.g. people complain about hardware set-ups for computer chess tournaments all the time, since now it's the case that you have a prominent engine that requires a different set-up. There are also different ways of comparing hardware, e.g. price or energy consumption. Go is comparable to chess, but is has significantly more possibilities per move than chess. There existed engines using the chess-like tree search for Go, but they were pretty bad because they get lost in all the variations. The neural network approach works much better there. Chess is interesting since both approaches seem fairly equal, so you can investigate scaling more meaningfully. And there's no real point in comparing humans to AlphaZero, since humans are much worse. edit: just a point about terminology, it's misleading to say that Stockfish uses tree search while AlphaZero uses neural networks. Because AlphaZero also uses (MC) tree search and Stockfish uses an evaluation function with weights tuned with machine learning tools. Given the obvious weaknesses that Leela possesses (and presumably AlphaZero too), the future best chess engine is probably somewhere in the middle between current SF and AZ. | ||
|
gpanda.sc2
20 Posts
On January 23 2019 12:50 imCHIEN wrote: vs $O$ to see how AI deals with cheese vs Maru to see how AI deals with his creative vs Serral to see how AI deals with a strong late game opponent. vs TY to see all the above at one time. | ||
|
DreamOen
Spain1400 Posts
But making it look like human problem solving and winning due to strategy and not insane sharp micro/multitask would be a really different thing. | ||
|
neutralrobot
Australia1025 Posts
On January 23 2019 11:53 KalWarkov wrote: until alpha zero beats stockfish in TCEC finals, i will never call alpha zero the strongest engine. everything is controlled by google. no table base, no opening books - which sf isnt trained for. and still, it isn't live games vs sf11dev. and who knows if they released all games or are just cherry picking? Well, I mean, it's always possible that they're presenting some kind of falsehood about the 100 game match vs Stockfish recently where Alpha Zero took no losses, but... why? Why would they flatly lie about the results of that match? Honestly I don't think they even care much about proving themselves in the domain of chess -- it was just part of a proof of concept about generalizing the AlphaGo algorithm to be applicable to other games. What do they gain by lying about this? Like, if you want to say that there should be a public tournament with different conditions before it's definitive, I can respect that, but the cherry picking idea seems pretty far-fetched to me, particularly considering the growth of Leela this year. On January 23 2019 17:40 Grumbels wrote: AlphaZero becoming the strongest engine in a matter of hours is a bit deceiving, given that it still required fifty million games of practice and computing a new version of the network every 25k games. It was estimated to take months for the Leela project (open source imitation of AZ), which is distributed on hundreds of computers. Google just has really powerful hardware. There are some interesting quirks with Leela. For instance, it's not capable of playing endgames efficiently, it seemingly aimlessly moves around, making moves that don't lose the advantage. It doesn't "get to the point". If an SC2 AI is built on the same concept, expect it to not be able to finish off games quickly and take an hour to mine out the entire map and build a fleet of random units to randomly move around the map. Another quirk of the project is that the algorithm uses not just the current move as input, but also the history of moves. This gives it some measure of what part of the board to pay "attention" to. It also means that if you give it a random position as input, without history, that it can't function. As far as I know Leela is useless in solving tactical puzzles and in handicap games without training it first. Leela also typically doesn't understand theory of endgames. It doesn't just play them weirdly, but it also doesn't grasp some almost mathematical ideas such as identifying a class of endgames that are drawn despite material imbalances (opposite color bishops, wrong color bishop). It's apparently also not better at fortress positions, where you have material disadvantages, but your position can't be cracked. There are some known positions like these, and it was hoped that neural networks would be better at them, and would be capable of reasoning that these are a special class of positions that require a different approach. But it doesn't really seem like it. Leela is also probably already better than Stockfish if you have bad hardware and no opening book. You can imagine that if there was a market for SC2 bots, that they could have opening books updated for every patch and which would have a team of people dedicated to keeping track of the meta and adding knowledge of it to the bot. But Deepmind's AI would use self-learning, i.e. only playing itself and developing its own meta. I don't know if that would make it easier or harder to beat as a human. I think the tree-search method for chess is bound to scale better with hardware than a neural network approach, given that chess is theoretically solvable with tree search. But this method would be useless for SC2, unless the AI uses some sort of abstraction of strategy and tries to think ahead. But I don't think you really need to think ahead in SC2 to get decent results. If you just react to your opponent and have perfect, bot-like control, you will win. Yeah, there are some quirks about Leela's play like the ones you mentioned. It's kinda hilarious watching Leela take forever to mate with Queen and King vs King, for example. But in most contexts, when both engines agree that the game is completely decided, they call it. Maybe Fantasy would make a new AI play for 2+ hours under totally lost conditions, but hopefully there would be a gg called before then in most cases. The talk of openings and the translation to SC2 is interesting to think about. AlphaZero seemed to keep going back to a relatively small handful of openings (I seem to remember it kept using the Berlin defense?) when left to its own devices as opposed to starting from a book position. But SC2 openings seem like they have to account for a lot more variables. Would a deep RL algorithm for SC2 play differently when optimizing for series vs single maps? Would it develop opening strategies that are more or less water-tight no matter what the context? Also, Would it show some of AlphaZero/Leela's brilliance for understanding positional compensation and imbalanced material? I guess we might find out about all this stuff soon. | ||
|
mishimaBeef
Canada2259 Posts
| ||
|
ZigguratOfUr
Iraq16955 Posts
On January 23 2019 17:17 Lazzarus wrote: So this is another AI playing SCII? https://twitter.com/ENCE_Serral/status/1087742590357774336 Yes, but those are 'regular' AIs coded up by someone (and with 100k APM for crazy micro tricks). | ||
|
Ronski
Finland266 Posts
On January 24 2019 00:57 neutralrobot wrote: Well, I mean, it's always possible that they're presenting some kind of falsehood about the 100 game match vs Stockfish recently where Alpha Zero took no losses, but... why? Why would they flatly lie about the results of that match? Honestly I don't think they even care much about proving themselves in the domain of chess -- it was just part of a proof of concept about generalizing the AlphaGo algorithm to be applicable to other games. What do they gain by lying about this? Like, if you want to say that there should be a public tournament with different conditions before it's definitive, I can respect that, but the cherry picking idea seems pretty far-fetched to me, particularly considering the growth of Leela this year. Yeah, there are some quirks about Leela's play like the ones you mentioned. It's kinda hilarious watching Leela take forever to mate with Queen and King vs King, for example. But in most contexts, when both engines agree that the game is completely decided, they call it. Maybe Fantasy would make a new AI play for 2+ hours under totally lost conditions, but hopefully there would be a gg called before then in most cases. The talk of openings and the translation to SC2 is interesting to think about. AlphaZero seemed to keep going back to a relatively small handful of openings (I seem to remember it kept using the Berlin defense?) when left to its own devices as opposed to starting from a book position. But SC2 openings seem like they have to account for a lot more variables. Would a deep RL algorithm for SC2 play differently when optimizing for series vs single maps? Would it develop opening strategies that are more or less water-tight no matter what the context? Also, Would it show some of AlphaZero/Leela's brilliance for understanding positional compensation and imbalanced material? I guess we might find out about all this stuff soon. The latest match where Stockfish and AlphaZero played 1000 games Stockfish was using its opening books and did manage to win a decent amount of games with white pieces. Alphazero still won the match overall but Stockfish did take games on a somewhat consistent rate. | ||
|
waiting2Bbanned
United States154 Posts
On January 24 2019 00:31 gpanda.sc2 wrote: vs TY to see all the above at one time. TY's cheese is repetitive and boring. sOs' is not bad, but neither can hold a candle to Has' dairy farm. | ||
|
Zreg
9 Posts
I wouldnt want to play against it! | ||
|
Rodya
546 Posts
| ||
|
Cyro
United Kingdom20318 Posts
On January 24 2019 02:18 Rodya wrote: Is there something about neural nets that make this interesting? I mean wont we just see insane tank dropship abuse? Even with deep learning computers are really bad at some stuff and really good at other stuff, it would be amazing to see one able to take on a pro in a variety of situations | ||
| ||
