TalkChess.com

Ovyron wrote:Ha! I find refutations for Cerebellum lines on a regular basis. The thing is, these refutations may be so deep, that the new moves may not be playable in anything other than correspondence games, because soon enough the engine will be out of book and 10 minutes per move might not be enough to play the advantageous position right.

The main use of Cerebellum is being able to predict what moves a "Cerebellum slave" will play and manage to come on top by forcing them into one of those holes

Cerebellum light was never intentend to be used in correspondence chess by simply following a line till the leaf node is reached.
It is a subset of Cerebellum full which I released as a demonstration to be used in engine-engine chess. I was interested how this would work and it worked very well.

You can used Cerebellum light for correspondence chess but near the leaf nodes you have to check some variations to see if everything is ok.

The tool for correspondence chess will be Cerebellum full.
There you have all the scores and every variation displayed, which means not only the best move but all moves which where regular played and some new ones found by Cerebellum calculations.

With Cerebellum full you can expand every leaf node with your own analysis tree re-calculated by Cerebellum, this will avoid the errors you mentioned and find possible winning lines too.

And maybe you can send me one of your refutations so that I can check what exactly you are refering too. Cerebellum light changes every week so it would be strange if an error lasts for more than a month.

Btw I wrote in the readme that everyone can always send me feedback about wrong lines, but I never received something from correspondence chess players till now, so I don't think it is much used in that area.

Thomas Zipproth wrote:......................................................

Btw I wrote in the readme that everyone can always send me feedback about wrong lines, but I never received something from correspondence chess players till now, so I don't think it is much used in that area.

Hi Thomas,

what news about the pgn I sent you? Has it been useful?

Thanks,
Rodolfo

Rodolfo Leoni wrote:

Thomas Zipproth wrote:......................................................

Btw I wrote in the readme that everyone can always send me feedback about wrong lines, but I never received something from correspondence chess players till now, so I don't think it is much used in that area.

Hi Thomas,

what news about the pgn I sent you? Has it been useful?

Thanks,
Rodolfo

Hi Rodolfo,

yes thanks it was really useful but I didn't had the time to check all the lines.
I will release a new Cerebellum in one or two days so that you can check the variations.
Of course this would be much easier with Cerebellum full which at the moment only exists on my Computer.

It should be possible to release a complete view of your variations in the demo page of the Brainfish website, I will do this when I have time.

Thomas

Thomas Zipproth wrote:And maybe you can send me one of your refutations so that I can check what exactly you are refering too.

I'm referring to positions where. if you play Cerebellum's white moves against my deeply analyzed black moves, after Brainfish is out of book, it shows an eval of 0.00 for the positions, and after even more deeply analyzed positions, I've managed to squeeze some black advantage.

I'm not interested in having Cerebellum "fixed" for those "problems." What has been happening is that I've managed to get an easy, equal position out of book, and managed to get advantage from the black side, against opponents that evidently have been "Cerebellum slaves."

So even though Cerebellum isn't supposed to be used like this, my opponents have been following it blindly. I like that, and I'd hate for them to use the moves I send to fix the holes against me, so in the opening, the moves that have been taking me days to analyze would be used against me by people that just load the engine and hit "move."

Please note that those "holes" are okay in, say, one minute/move, it's only when you analyze them deeply at "1/day move" or some time controls that the holes appear apparent. Otherwise, Cerebellum is a very interesting entity, that also provides deep lines the other way that refute my own deeply analyzed lines, so in some games going there without Cerebellum's aid would have been very dangerous.

Thomas Zipproth wrote:Maybe I should stop releasing Cerebellum till the professional Version is ready where everyone can expand the book himself which is the correct way to use it for correspondence chess.
This would also avoid such discussions.

Yes please! May I pay you in advance? Please let me know.

-Carl

Uri Blass wrote:

Milos wrote:

pijl wrote:

Jouni wrote:What do You think about this message in SF forum:

"sign up with ICCF and let stockfish make ALL the moves with no intervention on human part - in all probability your rating will not go over 2200"

This statement is probably true, although it may depend a little on the amount of time/processors you have at your disposal

Time control is known, 864000+259200s. Processors if you have at least 16 at your disposal mean very little.
At that TC and with 16 cores Xeon server with 128GB of RAM and Cerebelum limited to 8 moves for example, newest SFdev alone would have ICCF rating of 2600+ and probably not lose a single game.

I totally agree.

I got 2600+ ICCF rating and quit correspondence chess and I did not use top hardware.
I did not follow blindly the engine moves but I do not believe that my human input was worth more than running the engine on significantly better hardware with significant more time.

Using better hardware in most games could change nothing and I believe that at correspondence time control stockfish against itself is a draw in most cases even with time handicap of 10:1 but the same is for human input and in most cases my human input did not change the result.

Maybe today it is harder to get 2600+ because people use stronger hardware so you practically cannot get perfomance better than 2400 when you play against 2200 by this way but I guess that when you improve your rating and play in stronger tournaments you may get performance of 2530 when you play against 2400 and performance 2600 when you play against 2500 and continue in this way to get progress.

The numbers that I wrote are not based on calculation and only demonstrate the idea.

But as you mention, the key is not to blindly follow the engine moves and once you put that own the table I agree as well - then 2600 is possible. But you will have to be pretty darn good. The original question, maybe you would have a sense for this, let stockfish play on decent hardware by today's standard, not TCEC hardware, but maybe 1/4 of TCEC hardware (10/12 M nps) that an average person could afford easily , NO human intervention, and they will start at 1800, where do you think they will end up on the rating scale on ICCF. I would value your opinion more than anyone else since you have played ICCF at a very high level for a number of years.

MikeB wrote:I would value your opinion more than anyone else since you have played ICCF at a very high level for a number of years.

I wouldn't find Uri's opinion reliable for these discussions, for two reasons:

1. The Dunning-Kruger Effect. Uri Blass might be so smart when it comes to interacting with the positions when analyzing with engines that he greatly underestimates his own effect on those. He might only have a subtle influence in the games he play, deviating from an engine's choice only 1 or two times per game, and assume other people would have done the same, and that engine in faster hardware would have chosen his line with more time as well, while in reality, few people would have chosen his moves, and the positions turn out to be critical so his "subtle" changes have an enourmous impact on his games, and aren't replicable by other people or hardware alone.

He has had opponents that played weaker moves than an unassisted engine, and assumes they'd have played better without it, and so, on average, weak human interactions are worse than not interacting at all, but strong interaction in critical positions might have been what made him get to the place he reached.

2. Uri is a successful person, don't trust his advice.

MikeB wrote:

Uri Blass wrote:

Milos wrote:

pijl wrote:

Jouni wrote:What do You think about this message in SF forum:

"sign up with ICCF and let stockfish make ALL the moves with no intervention on human part - in all probability your rating will not go over 2200"

This statement is probably true, although it may depend a little on the amount of time/processors you have at your disposal

Time control is known, 864000+259200s. Processors if you have at least 16 at your disposal mean very little.
At that TC and with 16 cores Xeon server with 128GB of RAM and Cerebelum limited to 8 moves for example, newest SFdev alone would have ICCF rating of 2600+ and probably not lose a single game.

I totally agree.

I got 2600+ ICCF rating and quit correspondence chess and I did not use top hardware.
I did not follow blindly the engine moves but I do not believe that my human input was worth more than running the engine on significantly better hardware with significant more time.

Using better hardware in most games could change nothing and I believe that at correspondence time control stockfish against itself is a draw in most cases even with time handicap of 10:1 but the same is for human input and in most cases my human input did not change the result.

Maybe today it is harder to get 2600+ because people use stronger hardware so you practically cannot get perfomance better than 2400 when you play against 2200 by this way but I guess that when you improve your rating and play in stronger tournaments you may get performance of 2530 when you play against 2400 and performance 2600 when you play against 2500 and continue in this way to get progress.

The numbers that I wrote are not based on calculation and only demonstrate the idea.

But as you mention, the key is not to blindly follow the engine moves and once you put that own the table I agree as well - then 2600 is possible. But you will have to be pretty darn good. The original question, maybe you would have a sense for this, let stockfish play on decent hardware by today's standard, not TCEC hardware, but maybe 1/4 of TCEC hardware (10/12 M nps) that an average person could afford easily , NO human intervention, and they will start at 1800, where do you think they will end up on the rating scale on ICCF. I would value your opinion more than anyone else since you have played ICCF at a very high level for a number of years.

I believe that I could get the same result in correspondence chess in at least 80% of the games by blindly following the advice of the engine.

In part of the games I only chose the engine move because my analysis did not show something better.

It is possible to do a simple exercise to prove 2450-2550 are probably weaker than chess programs of their time in 2008-2010(or at least chess programs could avoid the losing mistake in case of using them for more time)

Analyze games that one side lost by stockfish in order to find the move that was probably the losing mistake(meaning a big change in stockfish's eval before the move and after the move).

After you find the losing mistake take the top chess program of the time that the loser play the move and give it 24 hours to analyze.

I believe that in most cases it is going to avoid the mistake.

I did not do it and it may be interesting to do it.

You can choose for example one of the following tournaments

https://www.iccf.com/event?id=17522
https://www.iccf.com/event?id=17523

https://www.iccf.com/event?id=16587

Thomas Zipproth wrote:

Rodolfo Leoni wrote:

Thomas Zipproth wrote:......................................................

Btw I wrote in the readme that everyone can always send me feedback about wrong lines, but I never received something from correspondence chess players till now, so I don't think it is much used in that area.

Hi Thomas,

what news about the pgn I sent you? Has it been useful?

Thanks,
Rodolfo

Hi Rodolfo,

yes thanks it was really useful but I didn't had the time to check all the lines.
I will release a new Cerebellum in one or two days so that you can check the variations.
Of course this would be much easier with Cerebellum full which at the moment only exists on my Computer.

It should be possible to release a complete view of your variations in the demo page of the Brainfish website, I will do this when I have time.

Thomas

I'm just too happy to hear that I gave a valuable contribution!
Any time I find something interesting I'll mail you an analyzed pgn.

About the page, I wouldn't look at it until all three games are at their end. Current situation is:

Game one: (novelty was by my opponent) 16.a3 Ouch!! Really an aggressive move. After hard fight it's draw.

Game 2: my novelty 11.b4. I played my 16th move of that game and I see about 350 cps advantage because of an opponent imprecision. He probably still thinks to be about -120 cpd disavantage with a normal Stockfish. All the difference is made by the score propagation of Stockfish_savehash by Daniel Jose. So we can consider this game as if already ended.

Game 3: my novelty 12.0-0. This will take more time to end. My current advantage is only about 110 cps. So do not rush to build the page...

Regards,
Rodolfo

Uri Blass wrote:After you find the losing mistake take the top chess program of the time that the loser play the move and give it 24 hours to analyze

Well, I haven't gone as far as 24 hours, but what I have seen from leaving analyzing the engine unattended for 8 hours (I only do this when I'm not there) is that when I come back it has produced a line that I'm able to refute in an hour, and that either I could have found that best move much faster by using a different engine (you may be surprised how often those moves that take hours for an engine to find, and like, and stabilize to, are found in the first minute of some other engine), or I'm able to find a better line than what was found in 8 hours, with my common, much faster methods.

I actually could produce those "8 hour" moves daily, you know? When I go to sleep my computer is sitting idle, you would think I can let my computer analyzing some move from my game so I get up to speed in the morning.

The reason I don't do it is because I have found it's a big waste of time. The main problem is that you have to refute the 8 hour line at the same relative depth, which takes a lot of time, and the higher the depth reached by the engine overnight, the longer it takes!

And, anyway, isn't this 8 hour analysis that I get equivalent in strength to some 24 hour analysis of the past? Because if back then I used 24 hour move things for my move choices, I'd have made those low quality moves, and 24 hour/move of today is just 8 hour/move of the future (say, of 3500 elo engines).

The secret is that you can get ahead of any high depth analysis easily. At what depth do you analyze? I do it at Depth 28, so if overnight my engine reached depth 36, I can get ahead and refute it if I go to its 6th move and reach depth 28 there, and see beyond its horizon.

Sometimes I get a better PV if I just let the engine reach 28 on every move and let it play 12 plies! And this takes about 12 minutes as opposed to 8 hours! And the 12th ply will be higher quality than the 12th ply of the 8 hour engine!

Perhaps this needs to be exemplified, give me a position where the engine needs 20 or some hours to find the best move, I can show how I'd find that move faster.

At least, I've been able to find the moves of Ed's "Start Your Engines" competition with relative ease and speed (never more than 10 minutes, I think?), and the only time I failed, I sent a winning move, so on real conditions I'd have won the game anyway, who cares if it's not the fastest mate?