3Champs reloaded

M ANSARI · Post by **M ANSARI** » Tue Dec 17, 2013 10:46 pm

It is interesting how Houdini 4 is clobbered so often at longer time controls. Both Stockfish and Komodo seem to be able to outplay it more often that they get outplayed. Stockfish seems incredibly strong at longer time controls and it can get deep in critical lines very quickly, especially in endgames. It really is a tactical monster in the endgame! I also can't help but enjoy Komodo's games, it is one engine that I can understand many moves and it plays by far the most human chess. My guess is that Houdini must have some excellent time management tuning which makes Houdini play really well at short time controls. This time management optimization seems something that Houdini is still ahead of the competition in, but TM optimizations seem to not make much more of a difference at LTC.

IGarcia · Post by **IGarcia** » Wed Dec 18, 2013 12:49 pm

Very nice tournament! Was like a triple TCEC superfinal, with some differences.

Thanks for running and sharing.

ouachita · Post by **ouachita** » Thu Dec 19, 2013 1:10 am

Nice event Clemens. The statisticians will argue that 150 games are not enough, so please run another match using SF's current best.

M ANSARI · Post by **M ANSARI** » Thu Dec 19, 2013 5:38 am

Nice tournament ... it pretty much confirms TCEC in that Houdini 4 does not scale as well as Komodo and Stockfish. Stockfish is quite a monster on big hardware, and Komodo seems to gain as well as more time and bigger hardware covers for weakness in tactics when compared to the other big guns. Quite a big gap between SF and H4 ... SF really seems to have Houdini's number.

Milos · Post by **Milos** » Thu Dec 19, 2013 6:49 am

M ANSARI wrote:Nice tournament ... it pretty much confirms TCEC in that Houdini 4 does not scale as well as Komodo and Stockfish. Stockfish is quite a monster on big hardware, and Komodo seems to gain as well as more time and bigger hardware covers for weakness in tactics when compared to the other big guns. Quite a big gap between SF and H4 ... SF really seems to have Houdini's number.

That's completely false.
SF and K advantage over H is only due to RH's decision to pursue large contempt (around 15cp) in order to benefit more from rating lists (and justify more Elo gain between H3 and H4) assuming in his business model that potential customers rely more on rating lists than on some "geeky" tournaments.
Problem here is that there are many SF/K fans who simply don't want to run tournaments with H with contempt 0 even though this is advised model for the analysis by RH (i.e. for LTC and correspondence).
Replay the same tournament as this one with H4 contempt 0 and H4 would comfortably win it.

Hugo · Post by **Hugo** » Thu Dec 19, 2013 7:26 am

Hi All

sorry for late post but I am so much bussy with my job theres days.

Congratulations To Stockfish for clear win of the 3Champs tournament.

Code: Select all

Games        &#58;    150 &#40;finished&#41;

White Wins   &#58;     37 &#40;24.7 %)
Black Wins   &#58;     16 &#40;10.7 %)
Draws        &#58;     97 &#40;64.7 %)
Unfinished   &#58;      0

White Perf.  &#58; 57.0 %
Black Perf.  &#58; 43.0 %

ECO A =     24 Games &#40;16.0 %)
ECO B =     48 Games &#40;32.0 %)
ECO C =     20 Games &#40;13.3 %)
ECO D =     36 Games &#40;24.0 %)
ECO E =     22 Games &#40;14.7 %)

thank you for all your interests.

regards, Clemens Keck

Martin Thoresen · Post by **Martin Thoresen** » Thu Dec 19, 2013 10:53 am

Milos wrote: Replay the same tournament as this one with H4 contempt 0 and H4 would comfortably win it.

Well, Houdini played with contempt 0 in Stage 4 of TCEC and failed to qualify for the Superfinal. It's not like it's a magical setting or anything.

IGarcia · Post by **IGarcia** » Thu Dec 19, 2013 11:45 am

Martin Thoresen wrote:
Milos wrote: Replay the same tournament as this one with H4 contempt 0 and H4 would comfortably win it.
Well, Houdini played with contempt 0 in Stage 4 of TCEC and failed to qualify for the Superfinal. It's not like it's a magical setting or anything.

Hehehe some people do think its magical, remember this is "houdini"

SD and K are very close and the advantage obtained by SF is mainly because beating houdini. Each engine played 100 games (not 150) so is less representative ... but is very clear the tendency when you sum up this results with nTCEC and several other LTC tournaments.

Regards.

Milos · Post by **Milos** » Thu Dec 19, 2013 3:22 pm

Martin Thoresen wrote:
Milos wrote: Replay the same tournament as this one with H4 contempt 0 and H4 would comfortably win it.
Well, Houdini played with contempt 0 in Stage 4 of TCEC and failed to qualify for the Superfinal. It's not like it's a magical setting or anything.

First that is not true no matter what RH says, run your TCEC version on couple of positions (in tournament mode), run H4 with contempt 0 on those and observe the difference.
Second, TCEC??? 100 games Stage 4???
Give me a break, your tournament is statistically relevant as rolling a dice.
It's like saying I throw a coin 30 times (assuming your Stage 4 70% draw rate), got 16 or 17 heads and claim coin is biased for heads.
Do you even know what you are talking about?
I would be quiet in your place regarding bragging about importance of TCEC results in any possible sense.
It was just a fun tournament for CC geeks, nothing more, nothing less.

Martin Thoresen · Post by **Martin Thoresen** » Thu Dec 19, 2013 3:50 pm

Milos wrote: First that is not true no matter what RH says, run your TCEC version on couple of positions (in tournament mode), run H4 with contempt 0 on those and observe the difference.

So RH lied to me and everyone else about this then. Just a FYI, contempt is disabled in the 9601 version that played in Stage 4. Contempt is not even an UCI option.

Milos wrote: Give me a break, your tournament is statistically relevant as rolling a dice.
It's like saying I throw a coin 30 times (assuming your Stage 4 70% draw rate), got 16 or 17 heads and claim coin is biased for heads.

Milos wrote: Replay the same tournament as this one with H4 contempt 0 and H4 would comfortably win it.

I have never claimed it to be statistically relevant because I know it isn't.
It is you who are saying that 100 games for H4 with contempt 0 in this 3Champs would be statistically relevant because then H4 would have won comfortably.

So basically in your world, 30 games are statistically totally irrelevant while 100 games are statistically totally relevant. Interesting, since I thought the ELO error margin for 100 games are about +/- 55.

I assume you have suggested for the SF team to stop running 20.000-50.000+ games for each test then and just do 100.

Milos wrote: I would be quiet in your place regarding bragging about importance of TCEC results in any possible sense.

Sure I'll be quiet if you tell me to.

3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded

Re: 3Champs reloaded