jorose wrote: ↑Fri May 31, 2019 6:11 pmI am assuming this is on 1 core at 4/40 time control? How is the speed compared to vs 0.5? I assume there were no losses on time? What GUI are you using?
Yes, it is on 1 core and the time control is 40/2 (adjusted to the reference machine).
The speed was abot 836 knps for 0.5 and 788 knps for 0.6.
There have been no time losses. I am using the Shredder 13 GUI.
Could you test the "unofficial" 0.5.5b to see if the behavior is different? That version has a large part of the (expected) strength improvement over 0.5 but it doesn't a lot of refactoring which could have had unexpected adverse effects. Also it drew a game against Komodo in TCEC so hopefully that is a sign that version is not defect
One point to note is I think 0.5 single core got lucky on the 40/4 list and is probably about 50 points higher than it should be, but 30% against 2980 is roughly where I would expect Winter 0.4a to be...
jorose wrote: ↑Fri May 31, 2019 6:38 pm
Could you test the "unofficial" 0.5.5b to see if the behavior is different? That version has a large part of the (expected) strength improvement over 0.5 but it doesn't a lot of refactoring which could have had unexpected adverse effects. Also it drew a game against Komodo in TCEC so hopefully that is a sign that version is not defect
One point to note is I think 0.5 single core got lucky on the 40/4 list and is probably about 50 points higher than it should be, but 30% against 2980 is roughly where I would expect Winter 0.4a to be...
I'm going to try 0.5.5b as soon as I'll have some time. Meanwhile, 0.6 is at almost 35 % after 88 games.
jorose wrote: ↑Fri May 31, 2019 6:38 pm
Could you test the "unofficial" 0.5.5b to see if the behavior is different? That version has a large part of the (expected) strength improvement over 0.5 but it doesn't a lot of refactoring which could have had unexpected adverse effects. Also it drew a game against Komodo in TCEC so hopefully that is a sign that version is not defect
One point to note is I think 0.5 single core got lucky on the 40/4 list and is probably about 50 points higher than it should be, but 30% against 2980 is roughly where I would expect Winter 0.4a to be...
I'm going to try 0.5.5b as soon as I'll have some time. Meanwhile, 0.6 is at almost 35 % after 88 games.
EDIT: I ran 0.5.5b with go depth 15 and its nps was 850.
jorose wrote: ↑Fri May 31, 2019 6:38 pm
Could you test the "unofficial" 0.5.5b to see if the behavior is different? That version has a large part of the (expected) strength improvement over 0.5 but it doesn't a lot of refactoring which could have had unexpected adverse effects. Also it drew a game against Komodo in TCEC so hopefully that is a sign that version is not defect
One point to note is I think 0.5 single core got lucky on the 40/4 list and is probably about 50 points higher than it should be, but 30% against 2980 is roughly where I would expect Winter 0.4a to be...
I'm going to try 0.5.5b as soon as I'll have some time. Meanwhile, 0.6 is at almost 35 % after 88 games.
EDIT: I ran 0.5.5b with go depth 15 and its nps was 850.
Could be that something in my refactoring (which added a lot of constexprs, can be a bad thing) resulted in a bit slower code when compiled with mingw, combined with a previous overperformance on 1 core for v0.5 (which would be consistent with my expectations for that version) combined with a bad start for v0.6.
If Winter 0.5 is 2880 (10 Elo weaker than on the 40/40 list, which would be consistent with history and within the margin of error the list says for the 4/40 rating) and Winter is 50 above that, then it would be roughly 2930 and expected to score around 43% against 2980 average opposition. My estimates are usually based on lower bounds from self play, but that could easily be overestimating things still, so maybe it is closer to 2900 in which case we are within margin or error of the expected performance at this TC.
jorose wrote: ↑Fri May 31, 2019 6:11 pmI am assuming this is on 1 core at 4/40 time control? How is the speed compared to vs 0.5? I assume there were no losses on time? What GUI are you using?
Yes, it is on 1 core and the time control is 40/2 (adjusted to the reference machine).
The speed was abot 836 knps for 0.5 and 788 knps for 0.6.
There have been no time losses. I am using the Shredder 13 GUI.
I got home half an hour ago and started running some tests. All I can say for sure is that at least the speed unde wine is attrocious compared to the speed I get under Linux.
I ran a gauntlet against previous Winter versions at roughly CCRL 4/40 TC (though in cutechess the moves and time seem to be flipped, so the TC command was 40/240) and the results against the previous releases were roughly as expected. If there is an issue I am thinking it is either compiler or architecture specific.
Rank Name Elo +/- Games Score Draws
0 Winter v0.6 130 9 3000 67.9% 49.3%
1 Winter v0.5 -60 13 1000 41.5% 62.8%
2 Winter v0.4 -143 14 1000 30.6% 52.5%
3 Winter v0.3 -198 19 1000 24.2% 32.6%
0.5.5b should be weaker than 0.6, but much closer to v0.6 than to v0.5.