I cannot see how Syzygy WDL probe does anything. Even with 249 table base hits, the engine avoids going into the eventual drawn position to keep the false advantage of the extra material score of a pawn. ELO tests also indicate no advantage, or a strength loss with the WDL probe. If the WDL probe is turned off then fewer nodes are searched, along with a proportional increase in speed. What am I missing?
In games with tablebase adjudication, only the WDL probe has any effect on strength, as only WDL tables are used in search. Generally the effect is small (in the order of 10-20 elo).
Note that there is a standard for including the TB hits in the Thinking Output, rather than printing it as a separate remark (which the GUI will ignore if it doesn't crash it): just write it as the 5th number on the line, and separate it from the PV behind it with a TAB character. Then WinBoard / XBoard could print it in the Engine Output, when you open the TB hits column.
konsolas wrote: ↑Tue Jul 02, 2019 11:50 am
In games with tablebase adjudication, only the WDL probe has any effect on strength, as only WDL tables are used in search. Generally the effect is small (in the order of 10-20 elo).
Not by my findings. The use of root DTZ reported about a 20+ elo increase for a 6-piece egtb. This was attributed to benefit from time loss by opponent engines, and the occasional saved game from complicated endings. However, the use of WDL drained this away to a big overall loss.
I cannot see how Syzygy WDL probe does anything. Even with 249 table base hits, the engine avoids going into the eventual drawn position to keep the false advantage of the extra material score of a pawn. ELO tests also indicate no advantage, or a strength loss with the WDL probe. If the WDL probe is turned off then fewer nodes are searched, along with a proportional increase in speed. What am I missing?
There are seven pieces on the board. You have 6-piece tables only. So the correct evaluation of the position needs to be determined by search. The WDL tables provide exact evaluations for any leaf positions that have six or fewer pieces. So search instantly determines that 95...Nxc6 is 0.00. But, as far as the engine knows, the position might be a win for Black. So it searches as usual to find a move that might be better than Nxc6. After a short time, it determines, with the help of the WDL tables, the correct values of all 13 moves in the root position. Here's Stockfish-dev with MultiPV=13:
… and how does Stockfish do with syzygy WDL turned off, but DTZ turned on? I do not have a build environment for SF, but somebody could comment the WDL section out in SF, re-build and compare the probes under the same position. Also, compare the final elo test under game conditions.