The exes should be almost exactly the same, and strength was the same in my overnight test. (Actually was a bit better, but nps bit slower, assume both were random noise.)
I don't have any older computers to test on, but if 2.4 worked for you, then hopefully the "slow64-sse" should now work fine.
(I also fixed an unrelated check evade move ordering bug where it sometimes accessed a static array slightly out of bounds, which was confusing to track down because the initial issue I noticed was sometimes commenting out completely unused code would change node count.)
Carlos777 wrote: ↑Sun Jan 17, 2021 10:44 pm
Unfortunately, it does not work either.
I've just now updated the zip file download again (slow64-sse.exe should be a newer modified time).
I found the max 32-bit instruction is SSE4 also, while max 16-bit is SSE2. So it seems likely this is the remaining issue, but if it still doesn't work I'll probably have to wait and see if I can find a way to test on an older machine.
Frank Quisinsky wrote: ↑Sun Jan 17, 2021 10:16 pm
I added your v2.5 in my still running FCP Tourney-2021 and changed to v2.5 from today after round 4 out of 50.
Now after 160 games place 3 / 41 with 116 / 160 points, 1.5 points more as Ethereal.
Seems to be the same strength as Ethereal, but I haven't enough games.
Thanks for updating it, glad to see Slow 2.5 in 3rd place so far. It does beat Ethereal 12.75 on my own computer in 2-moves book, 15s+.15 2000 game matches (by various small amounts of elo.) But never sure if those results that will generalize, especially with neural nets. I assume will be some mixed results depending on list and conditions.
jonkr wrote: ↑Sun Jan 17, 2021 11:56 pm
I've just now updated the zip file download again (slow64-sse.exe should be a newer modified time).
I found the max 32-bit instruction is SSE4 also, while max 16-bit is SSE2. So it seems likely this is the remaining issue, but if it still doesn't work I'll probably have to wait and see if I can find a way to test on an older machine.
It scores +60 elo to 2.5 in bullet self-play games. (2-moves book, 8-moves book, herz book)
The main feature is a new network structure detailed in this forum thread, which combined with consistently running more training games overnight when my computer is free has resulted in a good increase in strength.
While I was pretty excited to see clear improvement, I still have a long Todo list. I'm hoping to eventually release at least a couple more version before calling it good. Some of the weaknesses I noticed
It is becoming very slow with the larger evaluation. I just measure Elo so this is fine by me, but I think on some other computers especially older ones the current structure probably isn't the best. (I may never get around to trying to sort this out, but maybe can make some general optimization and try testing on an older computer.)
FRC is still very bad compared to standard. I think this structure will be better for FRC, but I will need to start training FRC at some point and make sure it doesn't hurt standard.
The own book just doesn't cut it at this level. I tried running more book gen, but my book gen doesn't remove losing lines, and there are a lot of the lines are left from around the 2800 elo range where a greater variety is playable.
The net still doesn't seem big enough to detect things like totally trapped pieces. I was surprised how often I saw this in losing games against the strongest opponents like SF11.
I've stubbornly never even looked at NNUE except getting a general sense of it during the initial hype, so again the only net compatible with 2.6 is the one that ships with it. I'm not sure if I will add support, given the progress of NNUE engines especially stockfish, I can only assume using the nnue-ecosystem would quickly improve over my current efforts. Then I'd be left with the dilemma of whether or not to do the experiments that seem interesting to me (it's nice not to be too strong... there's a much higher chance of clear success trying something new. To be clear I think SF has at this point changed or replaced most of the original NNUE from what I read? But I'm unlikely to improve current best on my own.)
Despite these caveats, Slow's general positional board sense looks clearly improved to me and the +60 self-play suggests that the improvement is real.
It scores +60 elo to 2.5 in bullet self-play games. (2-moves book, 8-moves book, herz book)
The main feature is a new network structure detailed in this forum thread, which combined with consistently running more training games overnight when my computer is free has resulted in a good increase in strength.
While I was pretty excited to see clear improvement, I still have a long Todo list. I'm hoping to eventually release at least a couple more version before calling it good. Some of the weaknesses I noticed
It is becoming very slow with the larger evaluation. I just measure Elo so this is fine by me, but I think on some other computers especially older ones the current structure probably isn't the best. (I may never get around to trying to sort this out, but maybe can make some general optimization and try testing on an older computer.)
FRC is still very bad compared to standard. I think this structure will be better for FRC, but I will need to start training FRC at some point and make sure it doesn't hurt standard.
The own book just doesn't cut it at this level. I tried running more book gen, but my book gen doesn't remove losing lines, and there are a lot of the lines are left from around the 2800 elo range where a greater variety is playable.
The net still doesn't seem big enough to detect things like totally trapped pieces. I was surprised how often I saw this in losing games against the strongest opponents like SF11.
I've stubbornly never even looked at NNUE except getting a general sense of it during the initial hype, so again the only net compatible with 2.6 is the one that ships with it. I'm not sure if I will add support, given the progress of NNUE engines especially stockfish, I can only assume using the nnue-ecosystem would quickly improve over my current efforts. Then I'd be left with the dilemma of whether or not to do the experiments that seem interesting to me (it's nice not to be too strong... there's a much higher chance of clear success trying something new. To be clear I think SF has at this point changed or replaced most of the original NNUE from what I read? But I'm unlikely to improve current best on my own.)
Despite these caveats, Slow's general positional board sense looks clearly improved to me and the +60 self-play suggests that the improvement is real.
Very impressive progress!
... it's nice not to be too strong... there's a much higher chance of clear success trying something new.
