HUGE ANNOUNCEMENT -- "LAS TORRES" NEW ENGINE THAT OBLITERATES STOCKFISH

[DrEinstein]

Wow! Two such long posts for a non-existing engine...

[Uri Blass]

Note that there were some bad opening like 1.g4 or 1.f3 but basically stockfish did not win many equal positions in the first match.

I will repeat the same experiment when the side with 100:1 nodes advantage is black and it is going to take many hours.

I used nodes and not time control so everybody can reproduce the games.
hash size is 2048 mbytes.
For tablebases I did not download the 6 piece tablebases and have only the 5 piece tablebases and my guess is that tablebases are not going to change much because I read that the advantage from tablebases for the latest version of stockfish is smaller.

It is possible to try to find equivalent normal time control to nodes in the meaning of playing strength for your computer.

[Uri Blass]

Wow! Two such long posts for a non-existing engine...

Unfortunately the size of the games with details about evaluation and nodes per move was too big for one post so I needed to post the games with 2 posts.

[Sopel]

Is this forum being moderated at all?

[dangi12012]

Note that there were some bad opening like 1.g4 or 1.f3 but basically stockfish did not win many equal positions in the first match.

I will repeat the same experiment when the side with 100:1 nodes advantage is black and it is going to take many hours.

I used nodes and not time control so everybody can reproduce the games.
hash size is 2048 mbytes.
For tablebases I did not download the 6 piece tablebases and have only the 5 piece tablebases and my guess is that tablebases are not going to change much because I read that the advantage from tablebases for the latest version of stockfish is smaller.

It is possible to try to find equivalent normal time control to nodes in the meaning of playing strength for your computer.

Yes your claim was Black vs White has a high draw rate with 100:1 node advantage. You proved that 50M White vs 500k Black has 14/0 win rate which is expected.
Looking forward to your results.

And now I will tell you that you forgot one thing: The TT Table is shared among both engines. It is a memory mapped file that resides at a specific location. (last i read SF sourcecode). You can verify that by analysing a move for 10min. Then restarting analysis - you will see that exe is restarted. But the evaluation picks off where you stopped it. I dont know if the TT is compatible among different build versions but running both engines at the same time will have them both see the same hashset. (1.5 years ago maybe they changed it?)

If you run SF vs SF you will get a high draw rate. What you need to do Uri Blass - is to run one or both in seperate Virtual Machines.

[amanjpro]

Note that there were some bad opening like 1.g4 or 1.f3 but basically stockfish did not win many equal positions in the first match.

I will repeat the same experiment when the side with 100:1 nodes advantage is black and it is going to take many hours.

I used nodes and not time control so everybody can reproduce the games.
hash size is 2048 mbytes.
For tablebases I did not download the 6 piece tablebases and have only the 5 piece tablebases and my guess is that tablebases are not going to change much because I read that the advantage from tablebases for the latest version of stockfish is smaller.

It is possible to try to find equivalent normal time control to nodes in the meaning of playing strength for your computer.

Yes your claim was Black vs White has a high draw rate with 100:1 node advantage. You proved that 50M White vs 500k Black has 14/0 win rate which is expected.
Looking forward to your results.

And now I will tell you that you forgot one thing: The TT Table is shared among both engines. It is a memory mapped file that resides at a specific location. (last i read SF sourcecode). You can verify that by analysing a move for 10min. Then restarting analysis - you will see that exe is restarted. But the evaluation picks off where you stopped it. I dont know if the TT is compatible among different build versions but running both engines at the same time will have them both see the same hashset. (1.5 years ago maybe they changed it?)

If you run SF vs SF you will get a high draw rate. What you need to do Uri Blass - is to run one or both in seperate Virtual Machines.

What are you doing here? this is about EMANUELL'S GREAT FFffff--ART!! let it stink

and your claim regarding SF TT is wrong

[dangi12012]

and your claim regarding SF TT is wrong

Run it 10 minutes on a position. Kill the executable. Run it again. You will see that there is a persistent TT still used by eval(). Which will instantly start where you left off.

[Mergi]

Run it 10 minutes on a position. Kill the executable. Run it again. You will see that there is a persistent TT still used by eval(). Which will instantly start where you left off.

Not sure where you saw TT being persistent, but that is absolutely wrong and i've never seen any engine actually do that. And if they do, it's wrong.

Maybe you are confusing it with syzygy tablebases? But those are for endgame only.

There's many different ways to show that stockfish TT isn't persistent. For example, you can look at the main function, there's a call to Search::clear:
https://github.com/official-stockfish/S ... in.cpp#L46

which calls TT.clear:
https://github.com/official-stockfish/S ... h.cpp#L182

and TT.clear zeros out the entire transposition table:
https://github.com/official-stockfish/S ... tt.cpp#L86

[Uri Blass]

Note that there were some bad opening like 1.g4 or 1.f3 but basically stockfish did not win many equal positions in the first match.

I will repeat the same experiment when the side with 100:1 nodes advantage is black and it is going to take many hours.

I used nodes and not time control so everybody can reproduce the games.
hash size is 2048 mbytes.
For tablebases I did not download the 6 piece tablebases and have only the 5 piece tablebases and my guess is that tablebases are not going to change much because I read that the advantage from tablebases for the latest version of stockfish is smaller.

It is possible to try to find equivalent normal time control to nodes in the meaning of playing strength for your computer.

Yes your claim was Black vs White has a high draw rate with 100:1 node advantage. You proved that 50M White vs 500k Black has 14/0 win rate which is expected.
Looking forward to your results.

And now I will tell you that you forgot one thing: The TT Table is shared among both engines. It is a memory mapped file that resides at a specific location. (last i read SF sourcecode). You can verify that by analysing a move for 10min. Then restarting analysis - you will see that exe is restarted. But the evaluation picks off where you stopped it. I dont know if the TT is compatible among different build versions but running both engines at the same time will have them both see the same hashset. (1.5 years ago maybe they changed it?)

If you run SF vs SF you will get a high draw rate. What you need to do Uri Blass - is to run one or both in seperate Virtual Machines.

I use ponder off games and I believe one stockfish cannot learn from another one.
I checked and found that I can reproduce the game as human-computer games simply by copying the moves of one side in a different interface.

Note that white won 6 games and there were 14 draws in the match so it means majority of draws.
I started also opposite colors and so far black is leading 3-0 and 5 draws in that match.

When I tried 5000 nodes against 500 nodes I got 39.5-0.5 result for 5000 nodes so I believe with more nodes it is going to be more draws.

[dangi12012]

Note that there were some bad opening like 1.g4 or 1.f3 but basically stockfish did not win many equal positions in the first match.

I will repeat the same experiment when the side with 100:1 nodes advantage is black and it is going to take many hours.

I used nodes and not time control so everybody can reproduce the games.
hash size is 2048 mbytes.
For tablebases I did not download the 6 piece tablebases and have only the 5 piece tablebases and my guess is that tablebases are not going to change much because I read that the advantage from tablebases for the latest version of stockfish is smaller.

It is possible to try to find equivalent normal time control to nodes in the meaning of playing strength for your computer.

Yes your claim was Black vs White has a high draw rate with 100:1 node advantage. You proved that 50M White vs 500k Black has 14/0 win rate which is expected.
Looking forward to your results.

And now I will tell you that you forgot one thing: The TT Table is shared among both engines. It is a memory mapped file that resides at a specific location. (last i read SF sourcecode). You can verify that by analysing a move for 10min. Then restarting analysis - you will see that exe is restarted. But the evaluation picks off where you stopped it. I dont know if the TT is compatible among different build versions but running both engines at the same time will have them both see the same hashset. (1.5 years ago maybe they changed it?)

If you run SF vs SF you will get a high draw rate. What you need to do Uri Blass - is to run one or both in seperate Virtual Machines.

I use ponder off games and I believe one stockfish cannot learn from another one.
I am going to check if I can reproduce the game as human-computer games simply by copying the moves of one side.

Note that white won 6 games and there were 14 draws in the match so it means majority of draws.
I started also opposite colors and so far black is leading 3-0 and 5 draws in that match.

When I tried 5000 nodes against 500 nodes I got 39.5-0.5 result for 5000 nodes so I believe with more nodes it is going to be more draws.

How many cores do you have? Maybe I can do it faster it with my Ryzen 5950x. I get 50MNodes/s on average. Also I can run up to 58GB hashtable. If hashtable is full the quality of moves also degrades a little bit.
Just send me a list of command line arguments and I will run it on SF Latest.