TalkChess.com

Congratulations to Stockfish, Houdini, Critter and Komodo for qualifying for Stage 4 after Stage 3 ended yesterday.

Stage 4:
16 cores, ponder off.
Up to 16 GB hash.
Time control is 120' + 30".
Hexadeca round robin, 96 games.
Top 2 qualifies for the Superfinal.

http://tcec.chessdom.com/live.php

Full rules can be read on the TCEC website, click "Help" in the top menu then "Rules & Information".

Finished games can be viewed and downloaded in the Archive page:
http://tcec.chessdom.com/archive.php

This Komodo development version is using Syzygy tablebases as well, so all 4 participants are using tablebases of some sort.

Best of luck to all engines!

I think that it may be better to have a rule in the next TCEC that a program that score less than 50% simply does not promote to the next stage.

It is possible to have an exception only for the superfinal so at least 2 programs always promote to the next stage.

It can help to prevent cases when one program that is clearly weaker than the rest of the programs promote.

It happened in this season more than once.

1)Shredder scored only 7 out of 15 and promoted to stage 3
2)Critter scored only 13 out of 28 and promoted to stage 4

This doesn't make any sense to me.
Shredder qualified fair and square, the same did Critter.

If we are supposed to eliminate every "clearly weaker program" as you say it, then why not simply let SF, H and K play each other the whole Season?

Simulations Stage 4 after 4 RR out of 16 RR:


STAGE 4: To Superfinal

SF: 0.83
K: 0.75
H: 0.42
Cr: 0.00
__________

TO WIN STAGE 4

SF: 0.50
K: 0.37
H: 0.13
Cr: 0.00
__________

Victory in Superfinal

SF: 0.55
K: 0.32
H: 0.13
Cr: 0.00
___________

SF adversary in case it's in Superfinal

K: 0.70
H: 0.30
Cr: 0.00

Nice statistique. But why the hell H4 doing less better than Komodo ? I have see Houdini doing really well ins these private tournament except versus stockfish. I don't understand.

JJJ wrote:Nice statistique. But why the hell H4 doing less better than Komodo ? I have see Houdini doing really well ins these private tournament except versus stockfish. I don't understand.

I think that Komodo earns more from 16 cores relative to houdini.

Komodo also may be better than the commercial version that won
latest TCEC so komodo is favourite to do better result than houdini.

The crosstable with a quarter of the games played:

JJJ wrote:Nice statistique. But why the hell H4 doing less better than Komodo ? I have see Houdini doing really well ins these private tournament except versus stockfish. I don't understand.

I started with reasonable assumptions and TCEC ratings. H4 lost to SF and K in the previous season. SF improved quite a bit, K probably too, since that season. So, rating at these extreme LTC and hardware I set to be 20 Elo points SF, 10 Elo points K, 0 Elo points H, and -70 Elo points Cr. They are middle ground from TCEC ratings and common sense. As it shows, even 10 Elo points disadvantage H vs. K is noticeable in predictions, besides that, SF and K already have 1 point advantage after 4 RR out of 16 RR. I have a draw model giving ~55% draws, a bit more for stronger engines, a bit less for weaker like Critter.

The interesting thing is also that the new TCEC format with more games per engine and carefully selected openings is able to discern reasonably 15-20 Elo points differences.

Laskos wrote:

JJJ wrote:Nice statistique. But why the hell H4 doing less better than Komodo ? I have see Houdini doing really well ins these private tournament except versus stockfish. I don't understand.

I started with reasonable assumptions and TCEC ratings. H4 lost to SF and K in the previous season. SF improved quite a bit, K probably too, since that season. So, rating at these extreme LTC and hardware I set to be 20 Elo points SF, 10 Elo points K, 0 Elo points H, and -70 Elo points Cr. They are middle ground from TCEC ratings and common sense. As it shows, even 10 Elo points disadvantage H vs. K is noticeable in predictions, besides that, SF and K already have 1 point advantage after 4 RR out of 16 RR. I have a draw model giving ~55% draws, a bit more for stronger engines, a bit less for weaker like Critter.

The interesting thing is also that the new TCEC format with more games per engine and carefully selected openings is able to discern reasonably 15-20 Elo points differences.

I see. Komodo will be updated one more time if it pass to the final. Do you think it will be able to defeat Stockfish ? Or to be as good than it or better ?

JJJ wrote:

Laskos wrote:

JJJ wrote:Nice statistique. But why the hell H4 doing less better than Komodo ? I have see Houdini doing really well ins these private tournament except versus stockfish. I don't understand.

I started with reasonable assumptions and TCEC ratings. H4 lost to SF and K in the previous season. SF improved quite a bit, K probably too, since that season. So, rating at these extreme LTC and hardware I set to be 20 Elo points SF, 10 Elo points K, 0 Elo points H, and -70 Elo points Cr. They are middle ground from TCEC ratings and common sense. As it shows, even 10 Elo points disadvantage H vs. K is noticeable in predictions, besides that, SF and K already have 1 point advantage after 4 RR out of 16 RR. I have a draw model giving ~55% draws, a bit more for stronger engines, a bit less for weaker like Critter.

The interesting thing is also that the new TCEC format with more games per engine and carefully selected openings is able to discern reasonably 15-20 Elo points differences.

I see. Komodo will be updated one more time if it pass to the final. Do you think it will be able to defeat Stockfish ? Or to be as good than it or better ?

SF has a patch too, regarding multithreading. I keep the same supposed ratings (and assume H4 is not updated) , and the new predictions after 5 RR out of 16 RR are:

STAGE 4: To Superfinal

SF: 0.87
K: 0.74
H4: 0.39
Cr: 0.00

__________

TO WIN STAGE 4

SF: 0.57
K: 0.31
H4: 0.12
Cr: 0.00

__________

Victory in Superfinal

SF: 0.58
K: 0.30
H4: 0.12
Cr: 0.00

__________

SF adversary if it's in Superfinal

K: 0.70
H4: 0.30
Cr: 0.00