Well currently, I'm doing null-move pruning, reverse-futility pruning, and I'm also using history heuristics to improve my move ordering, as of version 6.0.0. So I am doing some pruning. This is in addition to having a TT, killer moves, and MVV-LVA moves to order already.mvanthoor wrote: ↑Thu Sep 16, 2021 10:54 am Maybe it's a combination of factors? In the above position, Rustic's current development version searches 60 million nodes at depth 9. (Including PVS and transposition table.) Your engine searches only 7.4 mln (without PVS) and 5.5 mln (with PVS) nodes. What other enhancements do you have at this point?
Right, that's what's confusing to me. By all other metrics I should be getting an improvement in strength, but from a recent test I ran last night, I'm losing 12 Elo (6.1.0 is the dev version):
Code: Select all
Score of Blunder 6.1.0 vs Blunder 6.0.0: 1552 - 1709 - 1050 [0.482] 4311
... Blunder 6.1.0 playing White: 778 - 846 - 531 [0.484] 2155
... Blunder 6.1.0 playing Black: 774 - 863 - 519 [0.479] 2156
... White vs Black: 1641 - 1620 - 1050 [0.502] 4311
Elo difference: -12.7 +/- 9.0, LOS: 0.3 %, DrawRatio: 24.4 %
SPRT: llr -2.96 (-100.4%), lbound -2.94, ubound 2.94 - H0 was accepted
Finished match
Maybe so Marcel, that's something I'll definitely look into because I'm a little stumped about where to start debugging this. I'm fairly confident my TT isn't buggy, because if I remember correctly, I got about ~160-180 Elo from it in self-play. But I'll look back over it and how it's being used. My other working theory at the moment is that my PVS is running out of time on certain searches, and returns a bad move every now and then, as mentioned here: http://www.talkchess.com/forum3/viewtop ... =7&t=69252mvanthoor wrote: ↑Thu Sep 16, 2021 10:54 am Could it be that the TT is not working correctly, making your PVS re-search more expensive than it needs to be? Seeing that your time to depth drops from just over 3 seconds down to 2.2 for depth 9, it's unlikely; but with a time to depth speedup of almost 30%, the engine -should- have been measurably stronger.