jdart wrote:I am really amazed how far people are jumping to conclusions.
I personally, wasn't assuming that my prediction would happen tomorrow.
Even if it took ten years, I believe my prediction (and Graham's) would be most likely.
LCZero despite being weak currently, is as far as I can see, the biggest change in computer chess since Rybka 1.0 (plz correct me if I'm wrong), so even if it takes a decade, wouldn't it be realistic to predict that one day it - or some entity much like it - will be dominant?
I believe this poll was about the preferences of computer chess enthusiasts, if/when this outcome becomes a reality.
To me that's not jumping to conclusions.
Frankly ten years from now stockfish would be so disgustingly strong that no one would really care about lc zero.
In ten years engines like stockfish won't exist anymore and only engine like alpha zero leela zero will dominate because it is a better way to reach elo.
Self learning is future. Many program has shown that already.
Better way to reach elo is not to see simple tactics.
I totally disagree and I believe stockfish is going to use self learning in the future.
Engines should be superior to humans and not inferior and being inferior not only relative to engines but also relative to strong humans in tactics after million of training games suggest that something is clearly wrong in the way Leela is developed.
An engine should never miss simple tactics that finish the game in a short number of moves and it should do a brute force search to prevent it.
I believe for example that if Leela start training games with 0 blunders of missing mate in 1 or mate in 2 it could be stronger because it could learn from all the correct mates about mating patterns more quickly(and it could be easy to do it because finding if there is mate in 1 or mate in 2 or not with 0 mistakes is not a problem).
I believe that Leela could progress more quickly in case of simply using tablebases in all training games and stopping the games in tablebase positions.
It's not about which one will become dominant and which one will leave the scene. It's the match between two of the strongest entirely different breed of engines fighting for supremacy is what I'm excited about. One is pure CPU while the other is pure GPU. One requires finely tuned sophisticated human evaluation input while the other learns by self play. One is a tactical beast while the other is a positional monster. The contrast of power makes the clash really exciting!
Uri Blass wrote:Better way to reach elo is not to see simple tactics.
I totally disagree and I believe stockfish is going to use self learning in the future.
Engines should be superior to humans and not inferior and being inferior not only relative to engines but also relative to strong humans in tactics after million of training games suggest that something is clearly wrong in the way Leela is developed.
An engine should never miss simple tactics that finish the game in a short number of moves and it should do a brute force search to prevent it.
I believe for example that if Leela start training games with 0 blunders of missing mate in 1 or mate in 2 it could be stronger because it could learn from all the correct mates about mating patterns more quickly(and it could be easy to do it because finding if there is mate in 1 or mate in 2 or not with 0 mistakes is not a problem).
I believe that Leela could progress more quickly in case of simply using tablebases in all training games and stopping the games in tablebase positions.
Clearly there are soon going to be hybrid engines developed that either try to incorporate neural net evaluations into an AlphaBeta engine, or try to assist the tactical ability of an LCZero-type engine by giving it an AlphaBeta expert system for tactics. Which of the two works best, and how, will likely be a main focus of chess engine development over the coming years.
For LCZero, I think it's completely fine to keep it a pure UCT-NN engine without external help, and see how far it can go on that basis. There are loads of ideas to further improve playing strength from there, all of which we can expect to see developed soon.
jkiliani wrote:... For LCZero, I think it's completely fine to keep it a pure UCT-NN engine without external help, and see how far it can go on that basis. There are loads of ideas to further improve playing strength from there, all of which we can expect to see developed soon.
Not just completely fine, but very important, almost necessary.
Otherwise we'll never know how strong an engine such an approach can create.
Uri Blass wrote:Better way to reach elo is not to see simple tactics.
I totally disagree and I believe stockfish is going to use self learning in the future.
Engines should be superior to humans and not inferior and being inferior not only relative to engines but also relative to strong humans in tactics after million of training games suggest that something is clearly wrong in the way Leela is developed.
An engine should never miss simple tactics that finish the game in a short number of moves and it should do a brute force search to prevent it.
I believe for example that if Leela start training games with 0 blunders of missing mate in 1 or mate in 2 it could be stronger because it could learn from all the correct mates about mating patterns more quickly(and it could be easy to do it because finding if there is mate in 1 or mate in 2 or not with 0 mistakes is not a problem).
I believe that Leela could progress more quickly in case of simply using tablebases in all training games and stopping the games in tablebase positions.
Clearly there are soon going to be hybrid engines developed that either try to incorporate neural net evaluations into an AlphaBeta engine, or try to assist the tactical ability of an LCZero-type engine by giving it an AlphaBeta expert system for tactics. Which of the two works best, and how, will likely be a main focus of chess engine development over the coming years.
For LCZero, I think it's completely fine to keep it a pure UCT-NN engine without external help, and see how far it can go on that basis. There are loads of ideas to further improve playing strength from there, all of which we can expect to see developed soon.
Sure, you can do supervised learning with Stockfish games, with probably faster and better Elo-wise results. You could incorporate learning with A/B expert system and could implement A/B in the search itself. With again probably better Elo-wise results.
But.
I do not want another Stockfish-like NN engine, which might even surpass pretty soon Stockfish itself. As of now, playing with playouts=1 LC0, it really gives me the impression of a "drunken genius". It's fun, original and new. This "from scratch" learning with poor tactical abilities is maybe essential for me from purely theoretical and even esthetical view. If LC0 was much more tactically aware, it might have not played such crazy games. Strength-wise, the achievement after less than 10 million self-games is remarkable. The scaling with time control, which is due to UCT-NN, is more human-like than standard A/B engine-like. Maybe even above human-like. LC0 today on a good GPU in tournament time control conditions is already top GM 2800+ FIDE Elo level. That is a stupendous achievement in 2 months. Keep it going that way.
jkiliani wrote:Clearly there are soon going to be hybrid engines developed that either try to incorporate neural net evaluations into an AlphaBeta engine, or try to assist the tactical ability of an LCZero-type engine by giving it an AlphaBeta expert system for tactics. Which of the two works best, and how, will likely be a main focus of chess engine development over the coming years.
For LCZero, I think it's completely fine to keep it a pure UCT-NN engine without external help, and see how far it can go on that basis. There are loads of ideas to further improve playing strength from there, all of which we can expect to see developed soon.
Sure, you can do supervised learning with Stockfish games, with probably faster and better Elo-wise results. You could incorporate learning with A/B expert system and could implement A/B in the search itself. With again probably better Elo-wise results.
But.
I do not want another Stockfish-like NN engine, which might even surpass pretty soon Stockfish itself. As of now, playing with playouts=1 LC0, it really gives me the impression of a "drunken genius". It's fun, original and new. This "from scratch" learning with poor tactical abilities is maybe essential for me from purely theoretical and even esthetical view. If LC0 was much more tactically aware, it might have not played such crazy games. Strength-wise, the achievement after less than 10 million self-games is remarkable. The scaling with time control, which is due to UCT-NN, is more human-like than standard A/B engine-like. Maybe even above human-like. LC0 today on a good GPU in tournament time control conditions is already top GM 2800+ FIDE Elo level. That is a stupendous achievement in 2 months. Keep it going that way.
I agree with you completely, just had this impression that we have a number of classical engine devs here who object to Lc0 based on its issues with tactics, maybe because they feel its success invalidates their work somehow? I just pointed out the obvious huge potential to combine this new method with previous ones, even Deepmind wrote in their paper: "It is likely that some of these techniques could further improve the performance of AlphaZero; however, we have focused on a pure self-play reinforcement learning approach and leave these extensions for future research."
Whether such augmentation is in fact desirable for the purpose of sparring with humans is questionable as you say. I'd not be surprised at all if chess grandmasters are already now starting to use Leela for analysing their games, alongside Stockfish, even with Leela currently still much weaker. Professional Go players are in fact widely using Leela Zero already: https://github.com/gcp/leela-zero/issues/1261
jkiliani wrote:...
Whether such augmentation is in fact desirable for the purpose of sparring with humans is questionable as you say. I'd not be surprised at all if chess grandmasters are already now starting to use Leela for analysing their games, alongside Stockfish, even with Leela currently still much weaker. Professional Go players are in fact widely using Leela Zero already: https://github.com/gcp/leela-zero/issues/1261
I'd be surprised if chess GMs are using Leela now. Chess GMs are not all that into chess programs in general, so 1 or 2 current ones is more than enough for them. The Go players have had Leela Zero longer & it's stronger relative to other Go programs than the chess version.
Augmentation is undesirable if we want to know how far the unaugmented program can go.
The DeepMind words are just the usual words that appear in every engineering paper.
jp wrote:^ Actually, does anyone know how Leela (Go) Zero compares with Alpha Zero (Go)?
While there's no way to make a direct comparison since Deepmind retired Alphago, I think it's safe to say that Leela Zero still has to improve quite a bit to reach AlphaZero (Go) level. It's currently using a 192x15 net with no signs of a stall so far. Once we're using a 256x20 network and progress is stalling, then it's likely that LZ will be roughly on the same level as AlphaZero (Go). However, I think it IS likely that LZ is already stronger than Alphago Lee from two years ago by now, on decent hardware at least.
