EPTStarCraft II: DeepMind Demonstration: Jan 24 - Page 2
| Forum Index > SC2 General |
|
ZergX
France436 Posts
| ||
|
vyzion
308 Posts
| ||
|
KaiserCommander
Mexico290 Posts
On January 23 2019 06:45 vyzion wrote: eventually deepmind and similar AI will be unbeatable by humans. Assuming they'll keep improving her neural network yes. In it's actual form thera are limitations of what and how much can it learn. | ||
|
figq
12519 Posts
| ||
|
alexanderzero
United States659 Posts
| ||
|
ZigguratOfUr
Iraq16955 Posts
On January 23 2019 06:17 Ronski wrote: Didn't they make it pretty clear that its the best chess engine there is atm? Beating the strongest engine at chess means that no human player could ever hope to beat it so at least when it comes to chess I would say its clear that AlphaZero is the best there is. I mean probably? But even when their paper was eventually released, it's still just a bunch of games against an old version of Stockfish in circumstances completely controlled, set up, and chosen to be favourable by the Deepmind team. The newest version of Stockfish can also beat the older version of Stockfish by about the same margin. But arguing who is the best and stuff like that isn't too meaningful in the first place (it isn't of any importance if AlphaZero is the best or the second best)--the important thing is the machine learning research. And with Deepmind controlling everything about their research there's no room for other people to investigate things like whether AlphaZero with the current training would also be able to play Chess960 or adapt to starting with a piece handicap and so on and so forth. It would be very disappointing if AlphaStarcraft came out and crushed Serral, Maru and Stats in showmatches and got shelved never to see the light again, leaving people to wonder about how AlphaStarcraft would react to (for example) playing on an island map, or how it would defend a cannon rush. | ||
|
mierin
United States4946 Posts
On January 23 2019 10:14 alexanderzero wrote: I'm pretty skeptical like everyone else, especially considering the state of their AI when it was shown at Blizzcon just three months ago. Still, it's weird for them to bring on professional commentators and livestream something if they don't have something impressive to show. Right. this is what I think. Like...who cares if deepmind can micro units to grab mineral or gas deposits faster than humans or other AIs? If deepmind really thinks people will be impressed by an AI "solving" a limited minigame...I mean, we can already do that. If the AI can compete from the barebones beginnings that regular human players do, that'd be something. I'm pessimistic but will watch regardless. | ||
|
neutralrobot
Australia1025 Posts
On January 23 2019 10:53 ZigguratOfUr wrote: I mean probably? But even when their paper was eventually released, it's still just a bunch of games against an old version of Stockfish in circumstances completely controlled, set up, and chosen to be favourable by the Deepmind team. The newest version of Stockfish can also beat the older version of Stockfish by about the same margin. But arguing who is the best and stuff like that isn't too meaningful in the first place (it isn't of any importance if AlphaZero is the best or the second best)--the important thing is the machine learning research. And with Deepmind controlling everything about their research there's no room for other people to investigate things like whether AlphaZero with the current training would also be able to play Chess960 or adapt to starting with a piece handicap and so on and so forth. It would be very disappointing if AlphaStarcraft came out and crushed Serral, Maru and Stats in showmatches and got shelved never to see the light again, leaving people to wonder about how AlphaStarcraft would react to (for example) playing on an island map, or how it would defend a cannon rush. Well, actually... They recently played more games vs Stockfish in better conditions and AlphaZero comprehensively destroyed Stockfish. Also, they released the algorithm, which might not be as open as releasing the code or the trained network, but it did mean that the algorithm was implemented in a more open manner in the Leela Chess Zero project, which is now pretty competitive with Stockfish and playing interesting games against it in the TCEC. (https://www.youtube.com/watch?v=UPkcAS2B60s) This is the generalized Alpha Zero algorithm -- can be applied to a variety of games. So if they follow that pattern, maybe with Starcraft they'll shelve their code but release the research, which means it can be replicated. Guess we'll see! Keen to see what they've come up with. You'd think it must be a big leap. Bear in mind that once they had the right algorithm, they could train AlphaZero in a matter of hours and get it to a point where it's the best in the world by a mile. They have an incredible ability to test and implement learning algorithms quickly. Part of what gives them such an edge is their TPU hardware. So once there's been a breakthrough it could go from "how do we do this?" to "HOLY SHIT!" in a very short timeframe. | ||
|
Deleted User 3420
24492 Posts
https://github.com/deepmind/pysc2 | ||
|
KalWarkov
Germany4126 Posts
On January 23 2019 11:16 neutralrobot wrote: Well, actually... They recently played more games vs Stockfish in better conditions and AlphaZero comprehensively destroyed Stockfish. Also, they released the algorithm, which might not be as open as releasing the code or the trained network, but it did mean that the algorithm was implemented in a more open manner in the Leela Chess Zero project, which is now pretty competitive with Stockfish and playing interesting games against it in the TCEC. (https://www.youtube.com/watch?v=UPkcAS2B60s) This is the generalized Alpha Zero algorithm -- can be applied to a variety of games. So if they follow that pattern, maybe with Starcraft they'll shelve their code but release the research, which means it can be replicated. Guess we'll see! Keen to see what they've come up with. You'd think it must be a big leap. Bear in mind that once they had the right algorithm, they could train AlphaZero in a matter of hours and get it to a point where it's the best in the world by a mile. They have an incredible ability to test and implement learning algorithms quickly. Part of what gives them such an edge is their TPU hardware. So once there's been a breakthrough it could go from "how do we do this?" to "HOLY SHIT!" in a very short timeframe. 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? | ||
|
ZigguratOfUr
Iraq16955 Posts
On January 23 2019 11:16 neutralrobot wrote: Well, actually... They recently played more games vs Stockfish in better conditions and AlphaZero comprehensively destroyed Stockfish. Also, they released the algorithm, which might not be as open as releasing the code or the trained network, but it did mean that the algorithm was implemented in a more open manner in the Leela Chess Zero project, which is now pretty competitive with Stockfish and playing interesting games against it in the TCEC. (https://www.youtube.com/watch?v=UPkcAS2B60s) This is the generalized Alpha Zero algorithm -- can be applied to a variety of games. So if they follow that pattern, maybe with Starcraft they'll shelve their code but release the research, which means it can be replicated. Guess we'll see! Keen to see what they've come up with. You'd think it must be a big leap. Bear in mind that once they had the right algorithm, they could train AlphaZero in a matter of hours and get it to a point where it's the best in the world by a mile. They have an incredible ability to test and implement learning algorithms quickly. Part of what gives them such an edge is their TPU hardware. So once there's been a breakthrough it could go from "how do we do this?" to "HOLY SHIT!" in a very short timeframe. They played a newer version of Stockfish in somewhat better conditions, and released some implementation details (far from releasing the entire algorithm). I certainly hope they'll be much more open this time. | ||
|
imCHIEN
14 Posts
vs Maru to see how AI deals with his creative vs Serral to see how AI deals with a strong late game opponent. | ||
Excalibur_Z
United States12281 Posts
On January 23 2019 11:45 travis wrote: The event could be a way of issuing an open challenge to independent programmers and other professional groups, deepmind did release their API after all. https://github.com/deepmind/pysc2 That's more or less what I'm expecting as well. They'll probably have some showcases like how the AI handles different types of tasks like micro, adapting build order based on scouting information, which parts of the map it prefers to hold, how it learns different types of maps, and so on. The rest will probably be left up to third parties to develop further. I'm certainly not expecting a full exhibition series, especially not against top-level competition. If that is what they present, though, I'll be absolutely floored. That said, if it is a surprise exhibition, then it means the AI has already been developing skills against real players on the ladder ahead of this event, in which case it would be amusing to speculate who exactly that is. | ||
|
blunderfulguy
United States1415 Posts
| ||
|
MrMischelito
347 Posts
| ||
|
MrMischelito
347 Posts
On January 23 2019 13:12 Excalibur_Z wrote: That's more or less what I'm expecting as well. They'll probably have some showcases like how the AI handles different types of tasks like micro, adapting build order based on scouting information, which parts of the map it prefers to hold, how it learns different types of maps, and so on. The rest will probably be left up to third parties to develop further. I'm certainly not expecting a full exhibition series, especially not against top-level competition. If that is what they present, though, I'll be absolutely floored. That said, if it is a surprise exhibition, then it means the AI has already been developing skills against real players on the ladder ahead of this event, in which case it would be amusing to speculate who exactly that is. Isn't botting something that is completely against Blizzards Terms and Conditions? I wonder how many accounts they got banned from battle.net already... | ||
|
alexanderzero
United States659 Posts
On January 23 2019 15:29 MrMischelito wrote: Isn't botting something that is completely against Blizzards Terms and Conditions? I wonder how many accounts they got banned from battle.net already... Blizzard can always make an exception if they want to. | ||
|
Akio
Finland1838 Posts
| ||
|
Akio
Finland1838 Posts
Also I wonder what race it would play, maybe Terran? The production mechanics are the "easiest" so that's my guess. | ||
|
FFW_Rude
France10201 Posts
On January 23 2019 03:22 Charoisaur wrote: Probably Deepmind vs Serral showmatch FlaSh quit SC2 a long time ago man... | ||
| ||
