What happend to TCEC?

[Dann Corbit]

I misremembered you as supporting the ICGA's findings.
Therefore, I do apologize for that.

Thank you, apology accepted! I knew you are a gentleman.

Until such time as they should apologize for the Rybka affair, I would not join the ICGA.

I respect that, but you cannot expect or coerce others to join you in such a crusade. I'll be damned if I let Vas' misbehavior affect my life. Why should I suffer from other people's faults?

I think we see data in a fundamentally different way. It is OK to see things differently.

Apparently so. But I did nothing else than pointing out what the standard analysis of the data tells us. 200-Elo error bars deny a 300-Elo Elo difference with a confidence larger than that of the error bars, and you need 10,000 times fewer games to get 200-Elo error bars than to get 2-Elo error bars of the same confidence.

I was wrong on the 300 Elo figure also. It's only 245.

[Lyudmil Tsvetkov]

I still think that people publishing open source code are communists. So I don't want to be associated with that group.

I also think that ICGA should stay strict in only let play engines that are original.

Commonfish is a nice case in point.

[hgm]

While there is some chance that a revolutionary pondering algorithm gives Jonny a big boost, we would expect to see it at 44 cores to some degree.

Not really. I don't think it uses this algorithm at all, on TCEC hardware. It is specific to the cluster version, to solve the problem of weakly-coupled nodes one typically finds in big clusters (i.e. no RAM shared between them at all, only communication through network connections). Programs that are not specifically written for it (i.e. do not contain code to explicitly use the network to exchange results between the weakly coupled nodes) would not be able to benefit at all from using more than one such a node. And I don't expect the TCEC to use hardware that was completely useless to virtually all contestants, so those 44 (or 20 in the case of Johnny, as someone pointed out) cores must have been tightly coupled.

I would expect the CEGT results to be far more accurate, since there are 2700+ games for Jonny in that rating list.

No reason to 'expect' anything: the error bars are reported. They are 14 Elo in the CEGT list, against 65 Elo for the WCCC results. So about 4 times more accurate.

Error bars don't tell everything, however, as errors can be correlated. The accuracy of the Komodo-Johnny difference is better in the WCCC results than you might expect based on the individual error bars, because there were so many games directly between the two. Measuring an Elo difference from playing engines against each other has a much lower statistical error than when playing them against a 'third party'. (Of course it can have systematic errors.)

And accurracy is meaningless if the same thing isn't measured. The whole argument that Komodo on TCEC hardware will completely crush a 2400-core Johnny hinges on many completely uninformed guesses. You cannot be certain that Johnny cannot make more progress per year than Stockfish. Johnny can learn from what Stockfish does, but not the other way around, and going is much less tough if you are 300 Elo behind. You don't know what the effect of the opening books is, as Johnny's book is not public, and GECT does not play with own books. You don't know how much Johnny gains by using 100 times as many cores, as he is not using the cores just to parallellize an alpha-beta search, so Amdahl's law might not apply. All in all you know preciously little for how to adjust the somewhat-more-accurate CEGT rating for WCCC conditions.

And the fact is that the landslide victory predicted by those boasting that TCEC or CEGT tell it all simply did not happen. Proper statistical analysis shows that 9 games are good enough to exclude 300-Elo flukes with pretty high confidence.

[Henk]

I still think that people publishing open source code are communists. So I don't want to be associated with that group.

I also think that ICGA should stay strict in only let play engines that are original.

Commonfish is a nice case in point.

Doesn't say anything about the quality. Wasn't Picasso a communist too.

[Lyudmil Tsvetkov]

I still think that people publishing open source code are communists. So I don't want to be associated with that group.

I also think that ICGA should stay strict in only let play engines that are original.

Commonfish is a nice case in point.

Doesn't say anything about the quality. Wasn't Picasso a communist too.

and he is low quality, of course.

there is not a single communist in the whole wide world who would be worthy of anything(I do not know why, but I suddenly became mad, the word communist is simply something I can not stand)

I hope this was the last maddening exercise.

[Dann Corbit]

I still think that people publishing open source code are communists. So I don't want to be associated with that group.

I also think that ICGA should stay strict in only let play engines that are original.

Commonfish is a nice case in point.

Doesn't say anything about the quality. Wasn't Picasso a communist too.

and he is low quality, of course.

there is not a single communist in the whole wide world who would be worthy of anything(I do not know why, but I suddenly became mad, the word communist is simply something I can not stand)

I hope this was the last maddening exercise.

What fun is life, with no maddening exercises left in the tank?

I hope we never run out.

Now, I don't like things that cause people to die. But that is another story.

[hgm]

OK, so can we go back to how to continue/replace TCEC now?

One observation was that running SMP is a really inefficient way to get high-quality chess games (Amdahl's law, and all that). It would be much cheaper to run multiple games independently on a single core or hyper-thread, with a much longer TC. (And then broadcast them at a speed that can be independently tuned to optimally suite the spectators.)

If it costs 10k$ to run TCEC, it would seem more effective to offer the 10k up for prize money, and have the participants pay for their own electricity

To satisfy people that demand lower-level games to reduce the draw rate, it could be an idea to move the tourney to a Raspberry Pi platform.

[Henk]

I don't understand why you need TCEC. We can watch computer chess all the time for many years. You only need to install Tom's Live Chess Viewer and connect. Not many people watching these games. So there probably is not much demand.

Also better to run your own tournament with your own favorite engines.

Probably there might be a problem if you want to watch computer chess on your mobile phone. I don't know.

Normal people only use engines for preparation or analyzing their stupid games if they have time.

[Kotlov]

I don't understand why you need TCEC. We can watch computer chess all the time for many years. You only need to install Tom's Live Chess Viewer and connect. Not many people watching these games. So there probably is not much demand.

Also better to run your own tournament with your own favorite engines.

Probably there might be a problem if you want to watch computer chess on your mobile phone. I don't know.

Normal people only use engines for preparation or analyzing their stupid games if they have time.

I think that if it was in the web interface (+ chat), then the audience would have been much more.

Why do not we all help Graham in this together?

[Nordlandia]

I think that if it was in the web interface (+ chat), then the audience would have been much more.

Why do not we all help Graham in this together?

Good idea.

This was also said in the WCCC chat, GUI/live interface need to be improved.