mmt Vs. Ovyron (G4 D5 BG2)

[Uri Blass]

Your claim that you'd play exactly the same moves just 10 times faster is flat out false

Nope. I can emulate that precisely by, say, analyzing a position for 10 minutes instead of 1. Then I get to see what a x10 times faster computer would show, wouldn't I? I'd just get that in 1 minute instead of 10, that's how it works.

And I've never found an use for the extra info, it's like I wasted those extra 9 minutes, so I better stick with 1.

I'm playing people without clocks like mmt, I could use x10 times the time I take to play without problems, and then we'd see how I'd play with faster hardware. Except we'd see nothing better.

, and even taking it as hyperbole, it would mean that your much-vaunted analysis method scales like trash.

I take as much time as I need. The problem is "self-refutation". There comes a point in analysis where you have found the best move (optimal, most efficient, or whatever), and then you should make it. More analysis is just going to muddy the waters, either by making second best "catch up" to the same score so you're no longer sure about what was best anymore, or by finding a very strong defense by the opponent that makes your mainline score a hit so you're not sure if it holds up. Then you switch your move, and perform worse.

I have performed better by analyzing less because of this! I discovered this by accident when I had so many games going on the ICCF+LSS+FICGS that I actually didn't have enough time to do what I do, so I had to settle for lower quality analysis, or so, I thought. But I won more games than with regular analysis!

What was happening was less self-refutation: In the past I'd have found an opponent's defense that refuted my line, so I'd have gone with something different instead. Now I didn't have time to refute my line, so I played it anyway, my opponent didn't find the refutation either, and I'd win!

This is how I got my ICCF IM norm with ease with a point to spare (I actually let an opponent escape with a draw in a position I was winning, but not because of my hardware, because of my apathy), and why I kept starting new games elsewhere while that was going. My method seems good enough to hold people on better hardware and beat people with worse analysis methods, and improving it leads to self-refutation and fewer wins, so I've maxed out what I can do.

I believe that unassisted stockfish with no opening book and your hardware can easily get the ICCF IM title today(assuming you give it 12 hours per move) because almost nobody at this low level give this time to engines in correspondence games.

I guess most opponents in your tournament that you got the IM norm do not give engines more than few minutes per move or maybe even do not use stockfish but some inferior engine like Fritz.

I did not check it and maybe I am wrong but here is the relevant tournament that you got the IM norm and first place.

https://www.iccf.com/event?id=74015

I did not analyze seriously the games but it seems that in the game you won against the player from portugal he made a clerical error by 17....Nfd7 instead of 17...Ned7 and in the game against the player from venezuela he blundered by 19...f5 that miss a simple tactics that even I can see OTB and maybe the opponent replaced order of moves.

[Ovyron]

I believe that unassisted stockfish with no opening book and your hardware can easily get the ICCF IM title today(assuming you give it 12 hours per move) because almost nobody at this low level give this time to engines in correspondence games.

I take nowhere near 12 hours per move on those games, so being generous and saying that I take 1 hour per move, that's a x12 speed up that I'm able to give my hardware with my analysis methods.

Unassisted engines just use resources really poorly when it comes to analysis, that's why using methods that give you the move it'd take them 12 hours to find in 1 hour is no feat, and if people don't have "time to waste" and they need faster hardware for this, perhaps corr chess isn't for them and they'd be better using their money in something else.

[mmt]

Is there any simple way to add some randomness to SF 11's move choices in self-play? Otherwise, I guess the best way to play out 1. g4 would be to have LC0 with randomness create an opening book.

[Ovyron]

I've experimented with stuff like that since the times of Rybka Randomizer and can say that you'll end with a huge bulk of useless games where either the engine will have missed the only drawing mode every time or the only winning move every time, and then the percentages that you get are useless and unreliable.

This isn't about 1.g4 in the specific, people have tried this to create strong opening books in the past, and all of them are full of holes because of the exponential nature of chess. The best attempt I've seen is Brainfish Cerebellum and it didn't take much effort for Rebel's author to build a book that beat it, so regardless of what book you'd end with I'd expect another book would be able to refute its black attacks to a draw or its white defenses to a loss and we're back to square one.

I believe the only way to settle this is convincing person to person that 1.g4 loses, until one proves us wrong. If zullil is the only remaining person he tells us what line he thinks draws 1.g4 and we bust it. But this is a problem with no procedure that we can use to prove it one way or the other.

[mmt]

I'm not looking to prove it as it's impossible currently. I'm just interested to see what happens at different time controls and with different engines. As was mentioned before, training LC0 with the position after 1. g4 would be interesting. But I don't know if I'm motivated enough to configure everything and learn all the ins-and-outs of LC0.

[jp]

I've experimented with stuff like that since the times of Rybka Randomizer and can say that you'll end with a huge bulk of useless games where either the engine will have missed the only drawing mode every time or the only winning move every time, and then the percentages that you get are useless and unreliable.

This is the same problem with CC games, or any games. I don't think we should call them "useless". They are after all data, of lower or higher quality, depending on the quality of the players. But we won't know if they miss "the only drawing mode" or "the only winning move".

Is there any simple way to add some randomness to SF 11's move choices in self-play?

I've also wanted to know. It feels like it should be easy to hack, but the trick is to do it without compromising SF's strength. At the moment, I suppose the best thing to do is to use as many threads as possible.

[mmt]

I think what might be a realistic goal is to find a way to draw against SF 11 and LC0 playing black normally at some time controls. If we find a way to do it (maybe by using longer time controls as white and trying more varieties), we can increase black's time controls up to long controls to see if finds a win. If we cannot do it, it will become be probably more likely that 1. g4 is a win for black. The good part about LC0 here is that it plays a little different using different networks.

[Uri Blass]

I believe that unassisted stockfish with no opening book and your hardware can easily get the ICCF IM title today(assuming you give it 12 hours per move) because almost nobody at this low level give this time to engines in correspondence games.

I take nowhere near 12 hours per move on those games, so being generous and saying that I take 1 hour per move, that's a x12 speed up that I'm able to give my hardware with my analysis methods.

Unassisted engines just use resources really poorly when it comes to analysis, that's why using methods that give you the move it'd take them 12 hours to find in 1 hour is no feat, and if people don't have "time to waste" and they need faster hardware for this, perhaps corr chess isn't for them and they'd be better using their money in something else.

Note that I did not say that 1 hour per move with unassisted engine is not enough for IM title that you did not get(you got only IM norm) but only that 12 hours per move are enough to get easily the IM title so it is not clear that your methods are better than unassisted engines.

When I played correspondence chess in ICCF I believe that I used average time near 12 hours per move without the best hardware and got the GM title
and I do not believe my methods had a big advantage relative to using unassisted engine with no book.

[jp]

4r1k1/3nrpp1/2pq2n1/3p2Np/1P1P3P/3QPP2/2N2K2/6RR b - - 2 2

Komodo 13.2, depth 27, top three moves: -1.30 (... Nb6), -1.18, -1.17.

[Ovyron]

12 hours per move are enough to get easily the IM title so it is not clear that your methods are better than unassisted engines.

I'm not comparing myself to what I could do against a group of opponents vs. what unassisted engine at 12 hours per move could do against them, but about what I could do against a 12 hours per move unassisted engine in a game against it. It wouldn't be able to surprise me with a move (so losing is out of the question) and depending on opening I could lead it to a position where I predict it'd go into a losing variation (which is the point of being able to predict its moves.)

This would remain valid even if you give more time to the engine or faster hardware to the engine (which is equivalent) and that's why I claim I'm on a higher level than zullil if he went and relayed his unassisted engine moves to me in a game. To catch me he'd need centaur skills (and that's why a game against Zenmastur would be more interesting: he has the hardware and the skills, so he's probably at a higher level already, but is it enough?)