|
|
Post by matejst on May 8, 2022 17:40:08 GMT
Amanj published Zahak 10.
A few words about the author: so far, he has been responsive to all the usual, reasonable requests -- he implemented multiPV, skill levels, TBs. It was not too much, but he made his engine comfortable enough for everyday use. He tested the skill levels long enough to be sure the implementation was correct/useful enough. I don't know if he will continue developing Zahak, but I will probably switch from Rebel (I still need to test it, though).
BTW, he is not the only one. While authors like Connor Macmonigle (and others) care only about their own experience and rating lists, David Carteau was kind enough to get in touch with me (he knew I was using Orion) and ask for features I would like to see in Orion. I can only be thankful to such kindness, and try to repay them with some testing, reports, suggestions.
|
|
|
|
Post by Ozymandias on May 8, 2022 21:06:43 GMT
I see you use a lot of secondary engines, is that for the purpose you previously stated? (Finding playable moves SF doesn't offer to use OTB)
|
|
|
|
Post by matejst on May 8, 2022 21:17:27 GMT
I see you use a lot of secondary engines, is that for the purpose you previously stated? (Finding playable moves SF doesn't offer to use OTB) In general, it is easier to understand positions with slower engines. Lately I was testing Leela cpu -- unfortunately, on my system, it is too slow, otherwise, on a bit faster system, it could perhaps be the best solution. I can follow the process, often see wrong moves, and understand what to play and how to play. (I have another problem with newer versions of Lc0 crashing at the end of the analysis). SF is simply too fast (and too good, I most often don't have a clue what it is playing).
Edit: A better explanation. It is more important to understand what not to play in a given position, what's bad, than what's good. When I used Orion 0.8, I could see that in a certain position I play with white, a sacrifice on e6 is possible. It is not always sound, but OTB it creates a mess. And the first possibilities Orion gives are errors by black -- after a few mn, at depth 18 or 20, it will found a refutation, but I see the idea, where black can make a mistake and what I should do, and also how I will lose the game if the opponent is too strong. But in such a case, he would beat me anyway.
|
|
|
|
Post by Ozymandias on May 8, 2022 22:06:01 GMT
And doesn't MultiPV help likewise? I get that SF's 1st choice may be estrange, but maybe the 2nd (or the 6th) are still good enough and more in line with what you might be inclined to play. Likewise, the opponent's response might no include a mistake in the main PV, but you can spot the plausible moves too, or is it more difficult that way?
I don't know, with electricity as expensive as it's been for the past year, I consider carefully before turning on the desktop, and also for how long ;-)
|
|
|
|
Post by matejst on May 8, 2022 23:41:05 GMT
And doesn't MultiPV help likewise? I get that SF's 1st choice may be estrange, but maybe the 2nd (or the 6th) are still good enough and more in line with what you might be inclined to play. Likewise, the opponent's response might no include a mistake in the main PV, but you can spot the plausible moves too, or is it more difficult that way? I don't know, with electricity as expensive as it's been for the past year, I consider carefully before turning on the desktop, and also for how long ;-) It is too fast, Ozy. In less than five seconds it has already reached depth 25, 30. And all his lines are top quality. I use it sometimes to check deeper, or to analyze complete openings in Chess Position Trainer. It is also good in technical, simple positions, endgames, when I am sure to understand his plan.It is also good for sparring, since it has UCI_limit_strength. I prefer though the free Fat Fritz, since the difference in Elo is completely irrelevant for me.
In general, I find it difficult to find a good engine for what I need. I like to think by myself about the position, and to check (calculate or enter the moves) the lines I see, until I know "I got it". I also find that I play better sharp positions that force me to work hard, so I like aggressive engines. Right now, I combine Marvin/Rebel, FF2, and sometimes Lc0 (whenever I forget that it crashes on my system). I also used Orion 0.8, and now I will try Zahak 10, to see if it his different.
Electricity is cheap here. We still have gas... My bill is about 25 euros in the summer, 50-80 euros in the winter.
|
|
|
|
Post by Ozymandias on May 9, 2022 5:08:29 GMT
Well, if electricity is not a problem for you, I guess that's as good a way as any other to find out what you're looking for. My argument for SF was that of speed, to reduce operating time, but if it's a disadvantage for you...
in the aggressive engine front, what about the MateFinder versions of SF? Or SashChess, that one has personalities. I guess you already tried them.
|
|
|
|
Post by matejst on May 9, 2022 8:12:58 GMT
I have Shashchess installed, but I did not try it lately, certainly not in these few months since I started working on chess again. I was not sure that the personalities worked as intended, though, since the NNs. Perhaps it is time to try it again.
I liked a version of Sugar from 26/01/2020 where Marco implemented a lot of great options, but Marco makes changes too often, there is really no continuity in his work.
|
|
|
|
Post by Ozymandias on May 9, 2022 9:34:14 GMT
I also wondered if the Shashin theory could be applied to the engine, since the change to neural networks, but the author assured me it did. It's my understanding of his reply that it influences the search part of the engine. I could be wrong about his meaning, though, as his English wasn't very reliable.
|
|
|
|
Post by Ozymandias on May 11, 2022 12:16:21 GMT
Ratinglist-testrun of Zahak 10, by Stefan Pohl, finished. Quite behind Rebel 15. |
|
|
|
Post by matejst on May 11, 2022 12:25:24 GMT
Ratinglist-testrun of Zahak 10, by Stefan Pohl, finished. Quite behind Rebel 15. I did not see the rating list, but I tried the engine and the NN did not look good enough. Amanj improved the search, so it is better in tactics, but it seems that he had not discovered what Ed found when training nets: that you have to have bad moves, blunders in your games to really improve the net.
Unfortunately, there will be no Rebel 15 for me. My only hope is to buy now a refurbished laptop with avx2. Meanwhile, I switched to FF2 in CPT, and I will try Shashchess in Aquarium (one cannot modify the engine in CPT, so no Tal option for me).
|
|
|
|
Post by Ozymandias on May 11, 2022 17:09:06 GMT
IIRC, Capablanca - Petrosian was the most interesting setting. Tal gave the most boring evals/moves.
Speaking of AVX2, I have a 5820K laying around. I'm doing absolutely nothing with it, so I can send it to you. Motherboards are expensive, though, so maybe it would do you just as much good.
|
|
|
|
Post by matejst on May 11, 2022 17:26:53 GMT
IIRC, Capablanca - Petrosian was the most interesting setting. Tal gave the most boring evals/moves. Speaking of AVX2, I have a 5820K laying around. I'm doing absolutely nothing with it, so I can send it to you. Motherboards are expensive, though, so maybe it would do you just as much good. Thanks, Ozy. I am touched by your generosity, and I don't know how to express my gratitude. But you know very well that I could not accept. On the other hand, the cost of living in Serbia is not great, and I have some savings.
I hope that we will meet one day, face to face, here or/and in Spain, and share a glass of wine together. Or rather, rakija.
|
|
|
|
Post by Ozymandias on May 14, 2022 18:07:31 GMT
I see you use a lot of secondary engines, is that for the purpose you previously stated? (Finding playable moves SF doesn't offer to use OTB) In general, it is easier to understand positions with slower engines. Lately I was testing Leela cpu -- unfortunately, on my system, it is too slow, otherwise, on a bit faster system, it could perhaps be the best solution. I can follow the process, often see wrong moves, and understand what to play and how to play. (I have another problem with newer versions of Lc0 crashing at the end of the analysis). SF is simply too fast (and too good, I most often don't have a clue what it is playing). Edit: A better explanation. It is more important to understand what not to play in a given position, what's bad, than what's good. When I used Orion 0.8, I could see that in a certain position I play with white, a sacrifice on e6 is possible. It is not always sound, but OTB it creates a mess. And the first possibilities Orion gives are errors by black -- after a few mn, at depth 18 or 20, it will found a refutation, but I see the idea, where black can make a mistake and what I should do, and also how I will lose the game if the opponent is too strong. But in such a case, he would beat me anyway.
How about Uri's idea? Sound like an interesting NN could come out of that training scheme. |
|
|
|
Post by matejst on May 15, 2022 8:44:13 GMT
Did not see it yet, Ozy. Uri has sometimes good ideas, and, sometimes... He is a kind of genuine, sincere troll.
|
|
|
|
Post by Ozymandias on May 15, 2022 9:13:34 GMT
TBH, I've seen him more as the former than as the latter. Even in Ed's forum's old thread (the one before his own special military operation AKA "Nazis aren't Nazis"), he didn't come off to me as heavy-handed.
|
|
