Laskos wrote:Yes, I checked and re-checked my SF Scorpio, it seems to work as intended. The conclusion would by that Syzygy WDL and Scorpio are not helping SF to convert root TB wins.
Unsurprising is unsurprising.
The rudimentary "progress" scoring function that can be found in the egbb probing code does not replace DTM/Z information.
I found that out in 2012 while implementing the Scorpio base probing in Houdini 3 - in my opinion table base support without DTM/Z tables is a waste of opportunity.
syzygy wrote:It would be interesting to see the result for SF+syzygy with only the WDL tables (just leave the DTZ tables out of the path, assuming you have them in a separate directory).
I used an EPD file with 3-4-5 men wins, having a bit harder positions, wins in at least 30-40 moves. Shredder 12 with Nalimovs is the perfect player here:
10'' + 0.1''
Program Score
1 Shredder 12 Nalimov : 2911.0/5000
2 SF 04.02 Syzygy : 500.0/1000
3 SF 04.02 NO TB : 422.5/1000
4 SF 04.02 Syzygy WDL : 419.0/1000
5 Toga Scorpio : 415.5/1000
6 Toga NO TB : 332.0/1000
Syzygy WDL doesn't help SF at all.
Yes, this is because your starting positions are already TB positions. WDL probing should still help in real games, when the root position has more pieces and is not a TB position. The point is that WDL allows perfect pruning in the search tree.
With starting positions that are TB positions, WDL is useless. No surprise. In temrms of utility in real game play, WDL+DTZ is obviously best, and my guess is that WDL still has some value (but not as much as WDL+DTZ).
Theory and practice sometimes clash. And when that happens, theory loses. Every single time.
syzygy wrote:It would be interesting to see the result for SF+syzygy with only the WDL tables (just leave the DTZ tables out of the path, assuming you have them in a separate directory).
I used an EPD file with 3-4-5 men wins, having a bit harder positions, wins in at least 30-40 moves. Shredder 12 with Nalimovs is the perfect player here:
10'' + 0.1''
Program Score
1 Shredder 12 Nalimov : 2911.0/5000
2 SF 04.02 Syzygy : 500.0/1000
3 SF 04.02 NO TB : 422.5/1000
4 SF 04.02 Syzygy WDL : 419.0/1000
5 Toga Scorpio : 415.5/1000
6 Toga NO TB : 332.0/1000
Syzygy WDL doesn't help SF at all.
Yes, this is because your starting positions are already TB positions. WDL probing should still help in real games, when the root position has more pieces and is not a TB position. The point is that WDL allows perfect pruning in the search tree.
With starting positions that are TB positions, WDL is useless. No surprise. In temrms of utility in real game play, WDL+DTZ is obviously best, and my guess is that WDL still has some value (but not as much as WDL+DTZ).
Yes, with 6-7 men, both 5-men SF Scorpio and 5-men SF Syzygy WDL showed improvement over SF NO TB, look to my previous post in this thread. Though, SF with the full Syzygy (WDL+DTZ) is still the best.
I can try out Scorpio, but I will need a Linux compile. I have tried to compile it on my computer but ran into problems that I have not solved yet.
Yes, I checked and re-checked my SF Scorpio, it seems to work as intended. The conclusion would by that Syzygy WDL and Scorpio are not helping SF to convert root TB wins. Maybe it would be better to switch them off when a TB position is reached.
Did you and Adam try the latest version on http://abrok.eu/stockfish_syzygy/ ? I don't see how WDL probing, the way it is done, could not be at least of some help with converting TB wins (with the exception of KBNK-like endgames).
Here is the result with the latest version, using hard 4-5 men TB wins file. No significant change:
I can try out Scorpio, but I will need a Linux compile. I have tried to compile it on my computer but ran into problems that I have not solved yet.
Ok I have a better option for Stockfish. Its EGBBs were not loaded in RAM because they used to take about 2 minutes to load but now they take only 1 second!! That should be better than any cache.
Only thing I can say from looking at it is that Stockfish has a lot of internal node recognizers that it can't be helped when the root starts from the root. The positions are too easy for the default engine without TBs. It is like an online TB generator. For the EGBB version to compete here all have to be loaded in RAM to avoid any slow down. Positions where default Stockfish sees a KNOWN_WIN >= 70 on display screen _should_ be removed.
Clearly Toga needed help and scored by 60 points more so there is no questions weeker engines without extensive internal recognizers benefit from it. Scorpio should also improve by similar margins. I will see if I can compile an x64 compile and send you both Scorpio and Stockfish.
Last edited by Daniel Shawul on Thu Feb 06, 2014 12:52 pm, edited 1 time in total.
I can try out Scorpio, but I will need a Linux compile. I have tried to compile it on my computer but ran into problems that I have not solved yet.
Yes, I checked and re-checked my SF Scorpio, it seems to work as intended. The conclusion would by that Syzygy WDL and Scorpio are not helping SF to convert root TB wins. Maybe it would be better to switch them off when a TB position is reached.
Did you and Adam try the latest version on http://abrok.eu/stockfish_syzygy/ ? I don't see how WDL probing, the way it is done, could not be at least of some help with converting TB wins (with the exception of KBNK-like endgames).
Yes, I downloaded the latest Linux version from there (internal Stockfish name 050214).
I realize this morning that my positions are not as similar to Kai's as I thought. They are also 3-4-5 men TB wins. My file from which these came from is sorted by distance to mate. I forgot that the distance is in plies, not moves. So, the positions I used for the results above vary from 21 to 49 plies to mate.
Barring any more screw ups on my part, I believe I can create a suitable early endgame test suit.
syzygy wrote:It would be interesting to see the result for SF+syzygy with only the WDL tables (just leave the DTZ tables out of the path, assuming you have them in a separate directory).
I used an EPD file with 3-4-5 men wins, having a bit harder positions, wins in at least 30-40 moves. Shredder 12 with Nalimovs is the perfect player here:
10'' + 0.1''
Program Score
1 Shredder 12 Nalimov : 2911.0/5000
2 SF 04.02 Syzygy : 500.0/1000
3 SF 04.02 NO TB : 422.5/1000
4 SF 04.02 Syzygy WDL : 419.0/1000
5 Toga Scorpio : 415.5/1000
6 Toga NO TB : 332.0/1000
Syzygy WDL doesn't help SF at all.
Yes, this is because your starting positions are already TB positions. WDL probing should still help in real games, when the root position has more pieces and is not a TB position. The point is that WDL allows perfect pruning in the search tree.
With starting positions that are TB positions, WDL is useless. No surprise. In temrms of utility in real game play, WDL+DTZ is obviously best, and my guess is that WDL still has some value (but not as much as WDL+DTZ).
Your guess is wrong. Kai already made a test and found Nalimov + Shredder bitbases to be equal to other bitbases alone. DTM/DTZ is an absolute waste of time when it comes to Elo. One may think Nalimov is the best because it scored best in this test, but that absolutely means nothing as the starting postions are selected to favour it. Stockfish has so much internal node recognizers that it is a TB generator when the position starts from the root with all the search depth it got. Toga benefited a lot (about 55 positons ) using bitbases alone...
I can try out Scorpio, but I will need a Linux compile. I have tried to compile it on my computer but ran into problems that I have not solved yet.
Yes, I checked and re-checked my SF Scorpio, it seems to work as intended. The conclusion would by that Syzygy WDL and Scorpio are not helping SF to convert root TB wins. Maybe it would be better to switch them off when a TB position is reached.
Did you and Adam try the latest version on http://abrok.eu/stockfish_syzygy/ ? I don't see how WDL probing, the way it is done, could not be at least of some help with converting TB wins (with the exception of KBNK-like endgames).
Yes, I downloaded the latest Linux version from there (internal Stockfish name 050214).
I realize this morning that my positions are not as similar to Kai's as I thought. They are also 3-4-5 men TB wins. My file from which these came from is sorted by distance to mate. I forgot that the distance is in plies, not moves. So, the positions I used for the results above vary from 21 to 49 plies to mate.
Barring any more screw ups on my part, I believe I can create a suitable early endgame test suit.
It maybe very hard to find positions where Stockfish search can not see a KNOWN_WIN from the root. It just has a lot of internal node recognizers for the common cases. The best measurement would be to find a weaker engine since we still need perfect play by DTM players.
I can try out Scorpio, but I will need a Linux compile. I have tried to compile it on my computer but ran into problems that I have not solved yet.
Ok I have a better option for Stockfish. Its EGBBs were not loaded in RAM because they used to take about 2 minutes to load but now they take only 1 second!! That should be better than any cache.
Only thing I can say from looking at it is that Stockfish has a lot of internal node recognizers that it can't be helped when the root starts from the root. The positions are too easy for the default engine without TBs. It is like an online TB generator. For the EGBB version to compete here all have to be loaded in RAM to avoid any slow down. Positions where default Stockfish sees a KNOWN_WIN >= 70 on display screen _should_ be removed.
Clearly Toga needed help and scored by 60 points more so there is no questions weeker engines without extensive internal recognizers. Scorpio must also improve by the same. I will see if I can compile an x64 compile and send you Scorpio and Stockfish.
I will see if I can create a better set of positions than the ones I used.
Adam Hair wrote:
Yes, I downloaded the latest Linux version from there (internal Stockfish name 050214).
I realize this morning that my positions are not as similar to Kai's as I thought. They are also 3-4-5 men TB wins. My file from which these came from is sorted by distance to mate. I forgot that the distance is in plies, not moves. So, the positions I used for the results above vary from 21 to 49 plies to mate.
Barring any more screw ups on my part, I believe I can create a suitable early endgame test suit.
I really need to wake up before posting. The distance is moves to mate, not plies. So they were as long as I originally thought. But still too easy for Stockfish.