Crazyhouse Computer Championships 2016

[Evert]

Hello people, I have 2 questions :
Is Stockfish-zh still the only multi threaded engine for zh?
Would it be possible to include the latest SF-zh running on 2 cores in this tournament or the next one? I'm really interested to see how much a zh engine gains when running on more cores.

I think Nebiyu can use more cores. Never tried that though.

[Ferdy]

I now uploaded CrazyWa 0.1.2. It differs from version 0.1.1 only in that I retuned one parameter (I tripled a certain contribution to the King Safety). This performed some 3% better against my usual gauntlet of Imortal 1.0, Sunsetter 7g, TJchess 1.1 and CrazyWa 0.0.8. And against CrazyWa 0.1.1 it is nowleading by 3% after 400 games, which in itself is barely significant, but at least makes it quite unlikely that it is weaker. So this is the version I want to enter in the championship.

I could not reproduce the crash you reported on 0.1.1. If nothing was written in dump.txt, it was not due to an assert. I have suffered one crash in many thousands of games (weeks of continuously running 4 games in parallel) here too, which was due to an assert, but I could not reproduce that either. It is likely due to a hash-key collision creating some very rare situation due to an invalid hash move that is not detected, but cannot be properly handled.

Looking at the task manager, it only uses 264 MB though I set it to use 512 MB.

[hgm]

You are right! It starts considering 16M entries (of 16 bytes), and then keeps dividing by 2 until it is less than the number of MB specified in the 'memory' command.

In the next version I will start at a larger power of 2. It should be able to beat KKFchess with 256MB, so for now just keep using this one.

BTW, about your tournament scheme: does in the 3rd losers round the winner of 7 really play the loser of the finals (8) of the same group? Or would you do a cross match between groups here? If you stay in the same group, the number 3 in one group never gets the opportunity to prove whether it is better or worse than the number 2 of the other group. E.g. in the hypothetical case that the number 4 and 5 seeds are quite close in strength, 1-3 much stronger, and 6-8 much weaker, number 4 would get sort of a free ticket to the finals, relying on the number 3 to eliminate 5 in the other group, never having to face it himself. 'With cross-qualifiers' 4-5 and 3-6 you would not have this problem. Both 5 and 6 would get the opportunity to prove that although they had two stronger participants above them in their own group, they are still stronger than the number two af the other (and thus belong to the 4 strongest in total).

[Ferdy]

BTW, about your tournament scheme: does in the 3rd losers round the winner of 7 really play the loser of the finals (8) of the same group?

That is correct.

If you stay in the same group, the number 3 in one group never gets the opportunity to prove whether it is better or worse than the number 2 of the other group. E.g. in the hypothetical case that the number 4 and 5 seeds are quite close in strength, 1-3 much stronger, and 6-8 much weaker, number 4 would get sort of a free ticket to the finals, relying on the number 3 to eliminate 5 in the other group, never having to face it himself. 'With cross-qualifiers' 4-5 and 3-6 you would not have this problem. Both 5 and 6 would get the opportunity to prove that although they had two stronger participants above them in their own group, they are still stronger than the number two af the other (and thus belong to the 4 strongest in total).

What you describe here is possible.

Initially the idea was just to separate the first seed and the 2nd seed, but applying single elimination in each group and the top 1 in each group will advance to the final stage or stage 2, then later decided to give a chance to the 3rd and 4th seeds, so this is extended to double elimination in each group and the top 2 in each group will advance to the final stage or stage 2 (the current scheme). The format was decided knowing the strength gap between these engines at Blitz TC.

[Ferdy]

Result at stage 1, group A, pairing 2:
TC 1h + 10s/move, best in 24 games
CrazyWa - KKFChess : 13 - 0

https://sites.google.com/site/zhassocia ... ndpairings#

Next match will be from group B, pairing 1, NebiyuAlien vs TSCP zh.

[hgm]

What you describe here is possible.

Initially the idea was just to separate the first seed and the 2nd seed, but applying single elimination in each group and the top 1 in each group will advance to the final stage or stage 2, then later decided to give a chance to the 3rd and 4th seeds, so this is extended to double elimination in each group and the top 2 in each group will advance to the final stage or stage 2 (the current scheme). The format was decided knowing the strength gap between these engines at Blitz TC.

I got that idea, because the way you have set up the seeding now, where the number 3 seeds of each group (5 and 5 of the total) encounter the number 2 seed of the group (and 4 vs 1) in the second round, when everything goes as expected (i.e. 2 knocks out 3, then is knocked out by 1, while 3 knocks out 4 in the second round of the losers branch), 2 would have to play 3 a second time. This seems just a waste of time. So it seemed more interesting to swap opponents with the other group, to get new info.

For the semi-finals you can then use the pairing rule (in case 3 of the same group inadvertanly qualify) that the winners of each group will not be paired with each other, and that the winner from the group where two others qualified, will preferably play against the one he did not play before. As the number 1 seed was paired with the number 4 seed of the group in the second round, it is likely that he can now play the number 3 seed that unexpectedly beat the number 2 seed of the other group.

[Ferdy]

Result at stage 1, group B, pairing 1:
TC 1h + 10s/move, best in 24 games
NebiyuAlien - TSCP zh : 13 - 1

[hgm]

I uploaded a new CrazyWa (version 0.1.3). It should now be able to handle hash tables upto 1GB. (I have never tried more than 64MB myself, though.)

I am embarassed to confess that I discovered a major bug in CrazyWa's LMR implementation. Now that this is fixed, CrazyWa seems to have again gained significant strength. I didn't have much time to test, but it scored 57.5% against 0.1.2, 11% higher than 0.1.2 against Imortal 1.0, 2% higher against Sunsetter 7g, and 1% higher against TJchess 1.1 (60%, 42.5% and 38%, respectively), in 100 games each. But when I changed the minimum LMR'ed depth from 2 to 3, it looks quite different after ~75 games: 49% against Imortal1.0 and Sunsetter 7g, 53% against TJchess 1.1, and 46% against the 0.1.3 version with minimum depth 2.

This is all barely above error bars, but since the results of the depth-3 version seem more balanced, I decided to go for that. If this is not just a fluke, and it really performs better against TJchess 1.1, I hope it will also be better against TJchess 1.3, which is CrazyWa's next opponent in the championship.

[Ferdy]

I uploaded a new CrazyWa (version 0.1.3). It should now be able to handle hash tables upto 1GB. (I have never tried more than 64MB myself, though.)

I am embarassed to confess that I discovered a major bug in CrazyWa's LMR implementation. Now that this is fixed, CrazyWa seems to have again gained significant strength. I didn't have much time to test, but it scored 57.5% against 0.1.2, 11% higher than 0.1.2 against Imortal 1.0, 2% higher against Sunsetter 7g, and 1% higher against TJchess 1.1 (60%, 42.5% and 38%, respectively), in 100 games each. But when I changed the minimum LMR'ed depth from 2 to 3, it looks quite different after ~75 games: 49% against Imortal1.0 and Sunsetter 7g, 53% against TJchess 1.1, and 46% against the 0.1.3 version with minimum depth 2.

This is all barely above error bars, but since the results of the depth-3 version seem more balanced, I decided to go for that. If this is not just a fluke, and it really performs better against TJchess 1.1, I hope it will also be better against TJchess 1.3, which is CrazyWa's next opponent in the championship.

Thanks for the update.
If CrazyWa is able to score 45% or more vs Sunsetter 9 at blitz TC on 50 games, I will give it a chance to enter the stage 2 by matching CrazyWa (assuming it could not make it to the top 2 in stage 1, group A) vs the top 2 engine in group B after stage 1 is done. They will play at std TC best in 24 games.

So far a short blitz gauntlet of CrazyWa 0.1.3 is the following. This is still way better than version 0.1.2.

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   0 CrazyWa 0.1.3 32bit          -267     142      48   17.7%    2.1%
   1 TJchess 1.3 64bit             338     nan      24   87.5%    0.0%
   2 Sunsetter 9 32bit             211     192      24   77.1%    4.2%

48 of 48 games finished.

Added Calendar for match estimated date.
https://sites.google.com/site/zhassociation/calendar

[hgm]

Note that you would not need to make any special provision for this when you cross-pair the numbes 2 and 3 of the groups. Then (if everything would work out according to the blitz ranking) CrazyWa would be paired with Sunsetter anyway, instead of having to play TJchess a second time.