New Kaissa based on the 12.52 Elo gain Stockfish

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
User avatar
Eelco de Groot
Posts: 4277
Joined: Sun Mar 12, 2006 1:40 am
Location: Groningen

New Kaissa based on the 12.52 Elo gain Stockfish

Post by Eelco de Groot » Wed Aug 07, 2019 3:54 pm

The big search tuning patch was tested at VLTC
ELO: 12.52 +-2.6 (95%) LOS: 100.0%
Total: 19291 W: 3119 L: 2424 D: 13748

20000 @ 180+1.8 th 1 VLTC to see how this big param tune scales after LTC (since tune was done at ltc). See discussion at ... /pull/2260 Low tp as I don't want to slow down the ltc tests that are currently running too much.
For single thread, I don't think there was anything like this in single thread commits, there was Joona's patch that was maybe even showing a larger gain but that was multithreading. And at VLTC! It is not known yet exactly what is causing the gain :)

Applied the tuning patch to Kaissa so now code is up to date again with Stockfish. Available soon in Rybka forum because it allows a bit larger upload. Just an analysis tool, I don't think the present code changes would do very well in games. I have not really tested it in any way so consider it more as a beta. I have not changed contempt but these days myself I put it at zero, that does not change the dynamic contempt but makes at least the values jump around not so much when you change side to move.

No testing but I think I can see that the branching factor is larger, more iterations in the same amount of time. It has to be related to more LMR etc, but not all pruning/reducing was increased, so I don't know what is the biggest factor. Also, LMR is different in Kaissa so I don't know if Stockfish also has a larger Branching factor, probably though.
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan

Post Reply