Page 7 of 11

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Fri Feb 01, 2019 6:18 am
by lkaufman
Chessqueen wrote:
Thu Jan 31, 2019 10:20 pm
Chessqueen wrote:
Tue Jan 29, 2019 5:08 pm
lkaufman wrote:
Mon Jan 28, 2019 4:48 pm
With no minor pieces on the board and an average number of pawns (five per side), two rooks are nearly a pawn better than the queen. But on a full board the queen is indeed better than two rooks, as GM Roman Dzindzichashvilli taught me before computers were strong enough to answer the question. On average thruout the game the queen is a bit weaker.
What apply to two GMs does NOT hold true with 2 two powerful programs of equal strength like Komodo vs Komodo or SF10 vs SF10 or Komodo vs SF10. I believe that a Queen can find too many squares to keep the two Rooks busy, but for two Humans GMs is probably a little bit harder to find all the best moves or squares to force a draw. Therefore, there are a lot of pieces situations where what we believe true for two GMs, nowadays very strong programs are proving it to be wrong :shock:

It seems to me that you are confusing the question of "which side is better" with the totally different question of "can one side win with perfect (or very strong) play". With equal pawns (five or less per side) and no particular positional advantage for either side, the two rooks will almost always be superior in the sense that it takes stronger play to hold for the queen than to hold for the rooks. Whether the rooks can win depends on the details of the position. I don't doubt that the queen can draw with perfect play from many such positions, but I also don't doubt that the two rooks can draw more often than not with a pawn less in similar situations. To me, "equal" does not mean "drawable" but it means that neither side has a significantly higher chance to win the game than the other side assuming good but not perfect play (which in most cases is not knowable). Whether "good" means 2000, 2500, 3000, or 3500 level play depends on context.

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Fri Feb 01, 2019 4:02 pm
by MonteCarlo
Dann Corbit wrote:
Fri Feb 01, 2019 4:43 am
Chessqueen wrote:
Fri Feb 01, 2019 4:15 am
I believe that the Alpha Zero team has done a lot of tests with different endgame pieces + pawns configurations. not necessarily thousand of games, but meaningful configurations that helped them to determine a more exact value of each pieces, in which computer has proven different and more exact value than what we humans has given each pieces for so long.
AlphaZero searches 60 thousand positions per second and it trained for 9 hours.
60,000 Nodes / second * 3600 seconds / hour * 9 hours = 1,944,000,000 nodes. About 2 billion chess positions were considered by the neural network.
This is actually a drastic underestimate.

The 60,000 nodes per second number is from the match play against SF, where it played on 4 first-generation TPUs.

To generate self-play games for training, 5000 first-generation TPUs were used for 9 hours. The supplementary materials note that each move in training was based on an 800 node search, which averaged ~40 milliseconds.

That's 45000 TPU-hours, times 3600 seconds in an hour, equals 162 million TPU-seconds. At 800 nodes in .04 seconds, that's (162000000/0.04)*800 or about 3.2 trillion nodes calculated in the training games.

As a check on the plausibility of that math, they also say they played 44 million training games in those 9 hours. At ~ 0.04 seconds per played ply, that would suggest there were 162000000/.04 ply played in training, which divided by 44000000 games would be 92 ply per training game, which is similar to what we see in Leela.

So, if we're talking about all positions seen by search in training including duplicates (lots and lots of 800 node search from the start position and such), it's likely something like 3.2 trillion nodes total for that training run.

:D

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Sat Feb 02, 2019 12:15 pm
by Chessqueen
lkaufman wrote:
Fri Feb 01, 2019 6:18 am
Chessqueen wrote:
Thu Jan 31, 2019 10:20 pm
Chessqueen wrote:
Tue Jan 29, 2019 5:08 pm
lkaufman wrote:
Mon Jan 28, 2019 4:48 pm
With no minor pieces on the board and an average number of pawns (five per side), two rooks are nearly a pawn better than the queen. But on a full board the queen is indeed better than two rooks, as GM Roman Dzindzichashvilli taught me before computers were strong enough to answer the question. On average thruout the game the queen is a bit weaker.
What apply to two GMs does NOT hold true with 2 two powerful programs of equal strength like Komodo vs Komodo or SF10 vs SF10 or Komodo vs SF10. I believe that a Queen can find too many squares to keep the two Rooks busy, but for two Humans GMs is probably a little bit harder to find all the best moves or squares to force a draw. Therefore, there are a lot of pieces situations where what we believe true for two GMs, nowadays very strong programs are proving it to be wrong :shock:

It seems to me that you are confusing the question of "which side is better" with the totally different question of "can one side win with perfect (or very strong) play". With equal pawns (five or less per side) and no particular positional advantage for either side, the two rooks will almost always be superior in the sense that it takes stronger play to hold for the queen than to hold for the rooks. Whether the rooks can win depends on the details of the position. I don't doubt that the queen can draw with perfect play from many such positions, but I also don't doubt that the two rooks can draw more often than not with a pawn less in similar situations. To me, "equal" does not mean "drawable" but it means that neither side has a significantly higher chance to win the game than the other side assuming good but not perfect play (which in most cases is not knowable). Whether "good" means 2000, 2500, 3000, or 3500 level play depends on context.
I see what you mean, and also a lot has to do with what side has the initial move like in this case and if the initial position is advantageous to the Queen side or the two Rooks side to move. Can you show me several positions where there is no advantage to either side no matter which side start and the two Rooks manage to win at the end?


Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 9:34 am
by hgm
Chessqueen wrote:
Fri Feb 01, 2019 5:23 am
In order to get an Idea oh how many more nodes can Alpha Zero calculate than lets say Komodo or SF10 on the fastest hardware or computer that SF10 can find, can you give us an Idea of how many Nodes / per second can Komodo or SF10 calculate on the fastest computer that Komodo can get hook on?
The way you formulate the question suggests you expect the opposite from what is actually the case.

Stockfish and Komodo calculate around 70 million nodes per second on a fast machine. Alpha Zero calculates only 80 thousand nodes per second.

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 2:23 pm
by Chessqueen
hgm wrote:
Mon Feb 04, 2019 9:34 am
Chessqueen wrote:
Fri Feb 01, 2019 5:23 am
In order to get an Idea oh how many more nodes can Alpha Zero calculate than lets say Komodo or SF10 on the fastest hardware or computer that SF10 can find, can you give us an Idea of how many Nodes / per second can Komodo or SF10 calculate on the fastest computer that Komodo can get hook on?
The way you formulate the question suggests you expect the opposite from what is actually the case.

Stockfish and Komodo calculate around 70 million nodes per second on a fast machine. Alpha Zero calculates only 80 thousand nodes per second.
Thanks for making me realize that it is not how fast a machine with Komodo or SF10 can calculate, but according to what I think I read is that Alpha Zero has done thousand and thousands of training hours and analyzed billions of positions under those training and it is capable of retaining those positions in case it comes again in one of its future game. Therefore, LCO will be stronger than Komodo or SF10 very soon.

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 2:34 pm
by Chessqueen
lkaufman wrote:
Mon Jan 28, 2019 4:48 pm
With no minor pieces on the board and an average number of pawns (five per side), two rooks are nearly a pawn better than the queen. But on a full board the queen is indeed better than two rooks, as GM Roman Dzindzichashvilli taught me before computers were strong enough to answer the question. On average thruout the game the queen is a bit weaker.
In this game GM Magnus Carlsen was trying to demonstrate that to GM Aronian, but when the Opponent is as strong or close in rating and they do NOT have to be two programs like Komodo versus SF10, it becomes very very hard to win with two Rooks. :shock:
https://www.youtube.com/watch?v=5jo0wLVAErc

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 3:15 pm
by hgm
Chessqueen wrote:
Mon Feb 04, 2019 2:23 pm
Thanks for making me realize that it is not how fast a machine with Komodo or SF10 can calculate, but according to what I think I read is that Alpha Zero has done thousand and thousands of training hours and analyzed billions of positions under those training and it is capable of retaining those positions in case it comes again in one of its future game. Therefore, LCO will be stronger than Komodo or SF10 very soon.
That is also completely wrong. AlphaZero cannot retain any positions. It just has a far more advanced (but far slower) evaluation than SF10 / Komodo. And not significantly more training games have been used for tuning it than have been used for tuning Komodo or SF10. After 4 hours of training AlphaZero had already reached the level of Stockfish. That was done with many machines in parallel, though. But that is what fishtest does too.

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 4:08 pm
by Chessqueen
hgm wrote:
Mon Feb 04, 2019 3:15 pm
Chessqueen wrote:
Mon Feb 04, 2019 2:23 pm
Thanks for making me realize that it is not how fast a machine with Komodo or SF10 can calculate, but according to what I think I read is that Alpha Zero has done thousand and thousands of training hours and analyzed billions of positions under those training and it is capable of retaining those positions in case it comes again in one of its future game. Therefore, LCO will be stronger than Komodo or SF10 very soon.
That is also completely wrong. AlphaZero cannot retain any positions. It just has a far more advanced (but far slower) evaluation than SF10 / Komodo. And not significantly more training games have been used for tuning it than have been used for tuning Komodo or SF10. After 4 hours of training AlphaZero had already reached the level of Stockfish. That was done with many machines in parallel, though. But that is what fishtest does too.
I do NOT believe that the Evaluation is much advanced than what Komodo uise, why does it train so much then, if is not going to use that knowledge later ?

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 4:37 pm
by hgm
Why do you say it trains 'so much'? It doesn't train more than Stockfish, does it? Especially if you take into account that it is about 1000 times slower than Stockfish, in terms of positions per second it sees during training. (Or perhaps only 100 times, if Stockfish is not running on a high-end machine.) IIRC the AlphaZero training took 4 hours on 5000 TPUs (which is 1250 machines like the one it was playing on.) That would have been 5000 hours on a single such machine. But taking the factor 100 in nps into account, that would correspond to 50 hours of Stockfish tuning. Stockfish development has taken orders of magnitude more than 50 hours tuning time.

And of course Stockfish or Komodo did not start with zero knowledge. It was hard-coded in them which were good and which were bad captures, or that it makes sense to refute moves from the same position in the same way. AlphaZero initially had no clue that a Queen was worth more than a Pawn. It didn't even know whether a Queen was an asset or a liability. A substantial part of the training was used just to acquire the knowledge that is hardcoded in every conventional chess engine.

Re: How powerful is the Queen compared to 2 Rooks ?

Posted: Mon Feb 04, 2019 8:59 pm
by Chessqueen
hgm wrote:
Mon Feb 04, 2019 4:37 pm
Why do you say it trains 'so much'? It doesn't train more than Stockfish, does it? Especially if you take into account that it is about 1000 times slower than Stockfish, in terms of positions per second it sees during training. (Or perhaps only 100 times, if Stockfish is not running on a high-end machine.) IIRC the AlphaZero training took 4 hours on 5000 TPUs (which is 1250 machines like the one it was playing on.) That would have been 5000 hours on a single such machine. But taking the factor 100 in nps into account, that would correspond to 50 hours of Stockfish tuning. Stockfish development has taken orders of magnitude more than 50 hours tuning time.

And of course Stockfish or Komodo did not start with zero knowledge. It was hard-coded in them which were good and which were bad captures, or that it makes sense to refute moves from the same position in the same way. AlphaZero initially had no clue that a Queen was worth more than a Pawn. It didn't even know whether a Queen was an asset or a liability. A substantial part of the training was used just to acquire the knowledge that is hardcoded in every conventional chess engine.
I understand that Alpha Zero does not have extensive knowledge Like Komodo, or SF10, but I refused to believe that it had Zero knowledge about the worth of the pieces and it only had knowledge about in what direction each piece move, otherwise in the first few in the first few games it probably trade its Queen for a pawn, or a rook or any other pieces that are much inferior, and the same goes for the Rook, that during the first few games the Rook was traded for a Bishop, Knight or even pawn. The only way that I would believe that it had absolutely zero knowledge of the worth of each piece is if I see the first few games.