While my opponent thinks, I couldn't resist making a chessmap for the Grob, zooming in this part of the A00-A03 ECO codes:
ENHANCE!
I always wondered if this method could be used to rank chess moves, but the biggest box after 4 half-moves is 1.g4 d5 2.Bg2 e5, and it's not even in Stockfish's top 4!
Doing these visualizations has helped me realize the vastness of chess, this box is what is left after just the first move of white, it's a really tiny part of the whole map, and what remains in the box is still immense!
Sorry for the delay, I'm having some hardware issues (might need a new motherboard). I did run Stockfish to 0.77 trillion nodes (depth 55) and it dislikes 8.d4 the least (-1.37). The response to 8.... Rb8 is expected to be 9. Qa6. I will have the move once I fix the hardware issue.
koedem wrote: ↑Sun Jan 26, 2020 9:07 pm
If you're looking to run very long searches, the standard Leela binary will not be very useful since AFAIK they haven't changed the data type of their Q. (at least last time I checked Q was a float which means that for long searches (> 100MN) the eval and with it the search will become stuck due to rounding errors essentially)
For such long searches I still use a relatively old forked version that changes Q to be a double.
From reading the GitHub issue, the problem appears to for nodes visited more than 8 million times - they stop updating. Some people there don't think it's big deal and a patch hasn't been accepted. I don't understand why they insist on testing it by playing matches vs. unpatched version, seems very inefficient. I could try building a version with the patch.
koedem wrote: ↑Sun Jan 26, 2020 9:07 pm
If you're looking to run very long searches, the standard Leela binary will not be very useful since AFAIK they haven't changed the data type of their Q. (at least last time I checked Q was a float which means that for long searches (> 100MN) the eval and with it the search will become stuck due to rounding errors essentially)
For such long searches I still use a relatively old forked version that changes Q to be a double.
From reading the GitHub issue, the problem appears to for nodes visited more than 8 million times - they stop updating. Some people there don't think it's big deal and a patch hasn't been accepted. I don't understand why they insist on testing it by playing matches vs. unpatched version, seems very inefficient. I could try building a version with the patch.
My fork with this change should still work though it is somewhat outdated https://github.com/koedem/lc0
(kudos to MelleKoning for implementing it)
It is not quite as clear cut with 8 million nodes, though around that time "strange" things can start to happen. I don't recall the details but the eval update should be roughly like newEval = oldEval * (oldNodes / (oldNodes + 1)) + Q_of_evaluated_node) / (oldNodes + 1). (so that the eval stays the average of all visited nodes' Qs)
As oldNodes gets in the range of a floats mantissa size the left summand will eventually just be oldEval whereas the right summand will eventually be small enough that adding it doesn't change the eval. (as said, I don't recall how the details were, may be somewhat different formula, but something like that is happening)
Changing Q to be a double is only a 2% or so increase of a nodes size and for long analyses it is kinda necessary but their point I guess is that it doesn't affect match play very much since you rarely hit those node counts there... (I disagree with that sentiment, the main usage of an engine is analysis not match play though considering I wrote 0 lines of code for Leela I am not really in a position to put up demands )
mmt wrote: ↑Tue Jan 28, 2020 1:04 am
Sorry for the delay, I'm having some hardware issues (might need a new motherboard). I did run Stockfish to 0.77 trillion nodes (depth 55) and it dislikes 8.d4 the least (-1.37). The response to 8.... Rb8 is expected to be 9. Qa6. I will have the move once I fix the hardware issue.
To clarify, you have played 8. d4 officially, and will respond to 8...Rb8 when you are able?
I haven't seen him playing anything officially, so I think he can still change his mind about 8.d4 (which would made me very happy. Hey, at this point I'd let him go back and play some other white move anywhere, except for 1.g4 )
From the current position
[d]r2qkb1r/pQ1nnppp/2p5/3p4/6b1/2N5/PP1PPPBP/R1B1K1NR w KQkq - 0 8
The only other viable option besides 8. d4 is 8. d3. But Stockfish finds a good attack quicker than LC0: 8.... Rb8. White is forced to play 9. Qxa7 (9. Qa6 loses after 9.... Nc5 10. Qa3 Nf5). 9.... Nf5
[d]1r1qkb1r/Q2n1ppp/2p5/3p1n2/6b1/2NP4/PP2PPBP/R1B1K1NR w KQk - 1 10
and here Stockfish knows right away that if 10. h3 then 10... Bc5! 11. Qa4 Bxf2+!. LC0 takes much longer.
But after 10. Qa4 Rb4 11. Qd1 Ne5 black puts on a lot of pressure and probably wins.
[d]3qkb1r/5ppp/2p5/3pnn2/1r4b1/2NP4/PP2PPBP/R1BQK1NR w KQk - 5 12
Anyway, 8. d3 means a very dangerous tactical position for white where Stockfish's finds nice sacrifices and is better as expected.
)
)
