dragon 4.5

[MichelG]

I fixed the memory leak issue, and played some matches against scan again.

Dragon 4.6.1 vs scan 2.0: +12 -9 =452 (50.2%, +2 elo)
Dragon 4.6.1 vs scan 3.0: +1 -38 =547 (46.8%, -15 elo)

(Not quite a scientific test; i was using the computer with the matches running, and they don't have the same number of games)

Played with 6 cores/12 threads, 1 min/game. Scan settings: tt-size = 26, bb-size = 6, Dragon settings: 4 GB endgame, large tables.

It seems that scan 3.0 is quite a large improvement over v2!

[Krzysztof Grzelak]

I understand that the current version is now 4.6.1. I also have such a question Michel. Is " large tables " the same you have made or is it new?

[BertTuyt]

Based upon these results Dragon 4.6.1 vs scan 3.0: +1 -38 =547 (46.8%, -15 elo), my ELO calculation yields 22?
So not sure if I have done something wrong.

Also interesting to see that these results are based upon a 1 min/game, whereas tournaments (like Olympics) are based upon 20 min/game.
Last but not least as 8-core (and even 16-core Threadripper) are afordable, than we really reached a situation that most results are within the draw black hole.

So when we continue we are only able to measure engine difference with a 1 sec/game match.....

Bert

[MichelG]

Based upon these results Dragon 4.6.1 vs scan 3.0: +1 -38 =547 (46.8%, -15 elo), my ELO calculation yields 22?
So not sure if I have done something wrong.

Also interesting to see that these results are based upon a 1 min/game, whereas tournaments (like Olympics) are based upon 20 min/game.
Last but not least as 8-core (and even 16-core Threadripper) are afordable, than we really reached a situation that most results are within the draw black hole.

So when we continue we are only able to measure engine difference with a 1 sec/game match.....

Bert

You are right ofcourse, i looked up the wrong value in the table.

I don't think 8/16 core will have that much effect, as alpha beta doesn't scale well with the number of cores (i guess i will find out when i get a many core machine).

One advantage of 16 core machines is that playing 'training' games for machine learning does scale nearly 100% with the number of cores. In any case, the black hole in inevitable.

Michel

[BertTuyt]

Michel, which processor are you thinking about to buy in the (near) future.
Also DB Generation scales quite good with the number of cores.

Bert

[MichelG]

Michel, which processor are you thinking about to buy in the (near) future.
Also DB Generation scales quite good with the number of cores.

10+ i think, probably intel. I would like to know the performance of Ryzen though on the draughts engine and some other projects.

DB generation is mostly limited by the overall memory latency; all you do is randomly access the ram you have.

I don't see much progress in the endgame databases; full 5x5 is still very far away for normal draughts, and even that is pretty useless as far as strength is concerned.

Michel

[Luzimar]

Dragon 4.6 is now available as beta-test from:

https://mdgsoft.home.xs4all.nl/draughts ... v4.6.0.msi

- bugfixes and improvements in the engine everywhere
- engine equal to 2017 tournament minus some big tables
- experimental feature: dragon can now import/copy&paste jpg or png images of positions. For example most fmjd diagrams.

Michel

Hi Michel another position the Bank of Finals points to victory More Can not win

Hi Michel another position the Bank of Finals points to victory More Can not win

[MichelG]

Hi Luzimar,

This is actually a very hard position. While the position is a win for black, it takes 30 moves to actually win it. The program would have to search 59 ply to find the actual moves.

The last move black played (7-1) was the optimal move, i am not sure it is unable to win.

The only sure way to find the path to the win is to generate the 7 piece distance-to-win databases. However this takes a long time and with 1.5 TB, it takes a lot of diskspace.

I have these databases, but there is not way to upload them.

Michel

[Luzimar]

Hi Luzimar,

This is actually a very hard position. While the position is a win for black, it takes 30 moves to actually win it. The program would have to search 59 ply to find the actual moves.

The last move black played (7-1) was the optimal move, i am not sure it is unable to win.

The only sure way to find the path to the win is to generate the 7 piece distance-to-win databases. However this takes a long time and with 1.5 TB, it takes a lot of diskspace.

I have these databases, but there is not way to upload them.

Michel

Hi Michel Ed has created 2 bases of mtc finals and a total wdl of 90 gigs and loaded into the mega in several parts and now the kingsrow kills any end of 8 thanks for your grateful reply

[Krzysztof Grzelak]

Hi Michel Ed has created 2 bases of mtc finals and a total wdl of 90 gigs and loaded into the mega in several parts and now the kingsrow kills any end of 8 thanks for your grateful reply

For the first time I hear Ed's base is 90 GB.

[Sidiki]

Hi Michel Ed has created 2 bases of mtc finals and a total wdl of 90 gigs and loaded into the mega in several parts and now the kingsrow kills any end of 8 thanks for your grateful reply

For the first time I hear Ed's base is 90 GB.

Read carefully "2 bases of MTC finals and a total WDL of 90 gigs"

[Krzysztof Grzelak]

Read carefully "2 bases of MTC finals and a total WDL of 90 gigs"

I read carefully, so 56+28 is not 90.

[Luzimar]

Read carefully "2 bases of MTC finals and a total WDL of 90 gigs"

I read carefully, so 56+28 is not 90.

Hello everyone here is the exact size of the two Kingsrow End Bases

[Luzimar]

Read carefully "2 bases of MTC finals and a total WDL of 90 gigs"

I read carefully, so 56+28 is not 90.

Hello everyone here is the exact size of the two Kingsrow End Bases 83.8 gb

[Krzysztof Grzelak]

Luzimar this way I am thinking where from you have 394 files. I have 374 files. For accuracy folder MTC has 138 files a folder WLD has 236 files. Whether at you this way he is also.You gave all the content of the program. Now I just understood.