@blackinback the only problem from a hardware perspective is you need a GPU to make it work fast enough. But in order to use the engine you don't need to retrain it each time, so if this is really better than stockfish I'm sure it will get there eventually. Sometimes they publish the trained models after the paper.
@m4rbel understood. My question was indeed more about how available/cheap a fitting GPU is.
On the other hand, one should not get overexcited. The thing scored 61% against stockfish. That's ~70 Elo points higher than SF. Very impressive and a nice show case for deep mind but not a total game changer (from a pure chess strength point of view).
On the other hand, one should not get overexcited. The thing scored 61% against stockfish. That's ~70 Elo points higher than SF. Very impressive and a nice show case for deep mind but not a total game changer (from a pure chess strength point of view).
@blackinback Don't forget
1) There are 3 Black wins in this, and Stockfish was previously "unbeatable" even if the opposing engine had White
2) Supposedly this was after 4 hours (!) of learning chess. You can imagine that this will continue to skyrocket, as -unlike Stockfish- Alpha Zero actually gets stronger and stronger over time (on its own)
3) 64% IS a total game changer from a pure chess strength point of view. It is absolutely MASSIVE. TCEC 2016 was a very convincing win for Stockfish over Houdini, and he scored 54.5%.
1) There are 3 Black wins in this, and Stockfish was previously "unbeatable" even if the opposing engine had White
2) Supposedly this was after 4 hours (!) of learning chess. You can imagine that this will continue to skyrocket, as -unlike Stockfish- Alpha Zero actually gets stronger and stronger over time (on its own)
3) 64% IS a total game changer from a pure chess strength point of view. It is absolutely MASSIVE. TCEC 2016 was a very convincing win for Stockfish over Houdini, and he scored 54.5%.
@IsaVulpes I am not sure that longer learning will yield a higher playing strength. If you check figure 1 in the paper, playing strength seems to level out already after half of their learning time.
Machine learning is obliterating traditional AI.
comment from groups.google.com/forum/?fromgroups=#!topic/fishcooking/pcFRIurN_l4
"They used 4 TPUs (which is like 400-2000 cores of normal cores based on how you count them) + they had stupid TC that favors NN (sf has some time allocation features that gives it elo... Yet is was forced to use exactly 1 minute/move), they played this 400-2000 cores vs only 64 cores sf with 1 GB hash (how often do you set 1 GB hash to 64 cores engine? It makes no sence), Alphazero used precalculated "good" openings vs no-book sf (brainfish someone? 50 elo just from adding book)... Basically it played vs insanely handicapped engine and won by 70-100 elo? Well, it's good, but it's not like "omg it's much better than any engine now". Not even close. You can say that it will improve itself further but elo graph looks like a flat line so I doubt it heavily."
"They used 4 TPUs (which is like 400-2000 cores of normal cores based on how you count them) + they had stupid TC that favors NN (sf has some time allocation features that gives it elo... Yet is was forced to use exactly 1 minute/move), they played this 400-2000 cores vs only 64 cores sf with 1 GB hash (how often do you set 1 GB hash to 64 cores engine? It makes no sence), Alphazero used precalculated "good" openings vs no-book sf (brainfish someone? 50 elo just from adding book)... Basically it played vs insanely handicapped engine and won by 70-100 elo? Well, it's good, but it's not like "omg it's much better than any engine now". Not even close. You can say that it will improve itself further but elo graph looks like a flat line so I doubt it heavily."
They should enter the computer chess championship to settle the argument
Deep mind was only given the rules of chess... and 4 hours.
Already beat top engine.
Already beat top engine.
This topic has been archived and can no longer be replied to.