jp wrote: ↑Thu Apr 09, 2020 6:45 am
Is it too late to change the match conditions?!
Yes, too late. Now I'm 3.5 out of 4. Komodo beat my son 3 to 1, and he is stronger than I am now, in fact he just tied for first in an online event with GMs in it. And it was 10.5 to 2.5 vs FM Gilden. No obvious problems. Maybe it's my vast experience playing Komodo with handicaps. Usually I do poorly at faster time controls, but 15' + 10" is enough time for me to play well. Still very strange.
I didn't see anything anywhere except the TC. Maybe get chess.com to boost the hardware.
I assume it's playing with no book. I would prepare it a wicked opening book.
jp wrote: ↑Thu Apr 09, 2020 6:45 am
Is it too late to change the match conditions?!
Yes, too late. Now I'm 3.5 out of 4. Komodo beat my son 3 to 1, and he is stronger than I am now, in fact he just tied for first in an online event with GMs in it. And it was 10.5 to 2.5 vs FM Gilden. No obvious problems. Maybe it's my vast experience playing Komodo with handicaps. Usually I do poorly at faster time controls, but 15' + 10" is enough time for me to play well. Still very strange.
I didn't see anything anywhere except the TC. Maybe get chess.com to boost the hardware.
I assume it's playing with no book. I would prepare it a wicked opening book.
The hardware is almost the best money can buy, at least for Komodo MCTS. I don't know what would be better than a 32 core threadripper for it; even a 64 core threadripper would probably be a tad worse as it is slower and Komodo MCTS has trouble using much more than 32 cores. I drew my final two games so the score of my match under identical conditions was 3 wins for me and 3 draws. I scored 2 wins and 2 draws vs MCTS, 1 win and 1 draw vs. normal K, so same percentage. My subjective opinion is that the MCTS games were harder for me. Yes, there is an opening book, but I wouldn't call it "wicked". At knight odds all openings are awful for White; for example with b1 missing, 1e4? is already a mistake due to 1...d5. Morphy usually did open 1.e4 but I think there was generally some kind of understanding or gentlemen's agreement not to respond 1...d5, as only one opponent did that, and he won. I just try to avoid those openings that are especially bad, and to insure enough variety that we're unlikely to have repeat games. I even gave Smerdon a copy of the book; he agrees it's not much help for him to prepare due to great variety.
Which engine type handles piece odds better? The standard AB or Neural based? AFAIK AB engines on piece odds don't mind exchanging queens. Do you think Neural engine is better in handling a losing position?
Jhoravi wrote: ↑Fri Apr 10, 2020 8:28 am
Which engine type handles piece odds better? The standard AB or Neural based? AFAIK AB engines on piece odds don't mind exchanging queens. Do you think Neural engine is better in handling a losing position?
The early versions of Lc0 (the 10,000 series) were very good at giving handicaps like knight odds, I think it was 11248 that won a blitz match from GM Naroditsky giving about knight odds on average. But after that they started resigning training games down a piece, so all versions after that series have been absolutely awful at this. They just blunder additional pieces, they consider all positions down a piece or more resignable. As for AB engines, you can't generalize; with high enough Contempt they will avoid trading queens when down a piece unless it's clearly advantageous or necessary. But Komodo is better than Stockfish at this, I think because the intrinsic Contempt built into SF has the effect of making it trade when losing. Our Komodo NN is pretty good at giving odds, but it's not quite ready for prime time yet.
If it's not to much to ask, can you provide us with just one sample game of Komodo NN even if it's still not really for prime time? Don't care if it's verus Human or Computer.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
lkaufman wrote: ↑Fri Apr 10, 2020 7:16 am
I even gave Smerdon a copy of the book; he agrees it's not much help for him to prepare due to great variety.
This is too generous!
lkaufman wrote: ↑Fri Apr 10, 2020 8:40 am
The early versions of Lc0 (the 10,000 series) were very good at giving handicaps like knight odds, I think it was 11248 that won a blitz match from GM Naroditsky giving about knight odds on average. But after that they started resigning training games down a piece, so all versions after that series have been absolutely awful at this. They just blunder additional pieces, they consider all positions down a piece or more resignable.
It may also be partly because later Leela versions play a bit less "incorrectly", i.e. unsoundly. For odds matches, it may help to have a more poker-like bluffing hustler playing.
lkaufman wrote: ↑Fri Apr 10, 2020 8:40 am
The early versions of Lc0 (the 10,000 series) were very good at giving handicaps like knight odds, I think it was 11248 that won a blitz match from GM Naroditsky giving about knight odds on average. But after that they started resigning training games down a piece, so all versions after that series have been absolutely awful at this. They just blunder additional pieces, they consider all positions down a piece or more resignable. As for AB engines, you can't generalize; with high enough Contempt they will avoid trading queens when down a piece unless it's clearly advantageous or necessary. But Komodo is better than Stockfish at this, I think because the intrinsic Contempt built into SF has the effect of making it trade when losing. Our Komodo NN is pretty good at giving odds, but it's not quite ready for prime time yet.
Now I get the mystery why I often see Leela doesn't promote to queen. It doesn't know what to do when advantage is too much.
jp wrote: ↑Fri Apr 10, 2020 4:12 pm
It may also be partly because later Leela versions play a bit less "incorrectly", i.e. unsoundly.
Nah, giving away pieces for free because you view all options as certain losses is just a pathological side-effect. It's definitely not a way of playing less incorrectly
AdminX wrote: ↑Fri Apr 10, 2020 12:02 pm
If it's not to much to ask, can you provide us with just one sample game of Komodo NN even if it's still not really for prime time? Don't care if it's verus Human or Computer.
The last match with FM Larry Gilden at knight odds was with the NN version. The five games were all posted here on talkchess.
MonteCarlo wrote: ↑Fri Apr 10, 2020 5:23 pm
Nah, giving away pieces for free because you view all options as certain losses is just a pathological side-effect. It's definitely not a way of playing less incorrectly
I'm not sure later Leela nets give away pieces more than earlier nets. Is there evidence they do?
Then again, I'm not sure later Leela nets don't hustle and fake as much as earlier nets either.
