MYG

[Laskos]

... I think the use of positive Contempt for rating lists would significantly increase Stockfish rating on these lists.

I consider that a myth repeated to often.

I played a full set of games against all opponents with SF 7 on the 18th of June 2016 and it (Version 7 at that time) gained 4 Elo with a Contempt of 20 (which was the prefered contempt in discussions at that date). 4 Elo is less than half one SD in my list, so basicaly it was noise and I might get the same with just repeating the normal SF games ...
You find the full information on my main page on that date (and on the 20th.06.16 the games of Komodo 10 without contempt).

Regards
Ingo

Maybe something then was wrong in your test or the data was too few.

Your stats from TOP16:

Code: Select all

Komodo 11.2.2    3315 :   3300 (+2053,=1144,-103),  79.5 %

Komodo loses 84.7% of its lost points to Draws.

Code: Select all

Stockfish 8      3299 :   3300 (+1918,=1310,-72),  78.0 %

Stockfish loses 90.0% of its lost points to Draws.

This is a significant difference. Amplified by the fact that Stockfish 8 is second rated, and lower you go, lower that ratio should be.
Also, although in direct match Stockfish 8 beats Komodo 11.2.2:

Code: Select all

220 (   37,  159, 24),  53.0%

On the rating list Stockfish 8 is 16 ELO points lower. And this is repeated in some other rating lists. Also, by regression, I derived that Komodo gains about 15-20 ELO points from its contempt on your rating list, because only one engine is slightly stronger, and some 13 are significantly weaker.

[JJJ]

Plz Ingo, just tell us it is not Stockfish dev.

[Lyudmil Tsvetkov]

Irregular performance might suggest this engines has some weakness or some specific strength. Why does it perform so well against Houdini and Komodo and so less against Stockfish 8 ? Why does it perform slighty less against lesser good opponent ?

because, it is most likely a SF clone(fully consistent with the scores, a new engine should have beaten SF by at least 60/40).

it is a pity Ingo has fallen for that...

[JJJ]

... I think the use of positive Contempt for rating lists would significantly increase Stockfish rating on these lists.

I consider that a myth repeated to often.

I played a full set of games against all opponents with SF 7 on the 18th of June 2016 and it (Version 7 at that time) gained 4 Elo with a Contempt of 20 (which was the prefered contempt in discussions at that date). 4 Elo is less than half one SD in my list, so basicaly it was noise and I might get the same with just repeating the normal SF games ...
You find the full information on my main page on that date (and on the 20th.06.16 the games of Komodo 10 without contempt).

Regards
Ingo

Maybe something then was wrong in your test or the data was too few.

Your stats from TOP16:

Code: Select all

Komodo 11.2.2    3315 :   3300 (+2053,=1144,-103),  79.5 %

Komodo loses 84.7% of its lost points to Draws.

Code: Select all

Stockfish 8      3299 :   3300 (+1918,=1310,-72),  78.0 %

Stockfish loses 90.0% of its lost points to Draws.

This is a significant difference. Amplified by the fact that Stockfish 8 is second rated, and lower you go, lower that fractio should be.
Also, although in direct match Stockfish 8 beats Komodo 11.2.2:

Code: Select all

220 (   37,  159, 24),  53.0%

On the rating list Stockfish 8 is 16 ELO points lower. And this is repeated in some other rating lists. Also, by regression, I derived that Komodo gains about 15-20 ELO points from its contempt on your rating list, because only one engine is slightly stronger, and some 13 are significantly weaker.

I think Houdini 5 is also too strong for contempt, even with the last Komodo. That's make 2 engines in fact, specially knowing Houdini is coming soon.

[IWB]

Maybe something then was wrong in your test or the data was too few.
...

In the test you have seen a contempt of 20 was useless for a rating list (no doubt here). Maybe the draw problem (if any) is not related to the contempt?!

I would rather trust my data and doubt the theorie than the other way around ...

Ingo

[Lyudmil Tsvetkov]

of course, other option is this is Houdini 6, but as Graham says, Houdart does not pre-release.

why I think this is SF clone:

- too heavy win vs Komodo, typical of SF
- almost 50/50 vs SF 8, just 51-52%, what latest dev would score against SF 8
- as Jean Batiste points out, not very convincing scores vs the weaker engines, again typical SF without contempt
- no contempt, well, again, this is SF

pity indeed.

[Laskos]

... I think the use of positive Contempt for rating lists would significantly increase Stockfish rating on these lists.

I consider that a myth repeated to often.

I played a full set of games against all opponents with SF 7 on the 18th of June 2016 and it (Version 7 at that time) gained 4 Elo with a Contempt of 20 (which was the prefered contempt in discussions at that date). 4 Elo is less than half one SD in my list, so basicaly it was noise and I might get the same with just repeating the normal SF games ...
You find the full information on my main page on that date (and on the 20th.06.16 the games of Komodo 10 without contempt).

Regards
Ingo

Maybe something then was wrong in your test or the data was too few.

Your stats from TOP16:

Code: Select all

Komodo 11.2.2    3315 :   3300 (+2053,=1144,-103),  79.5 %

Komodo loses 84.7% of its lost points to Draws.

Code: Select all

Stockfish 8      3299 :   3300 (+1918,=1310,-72),  78.0 %

Stockfish loses 90.0% of its lost points to Draws.

This is a significant difference. Amplified by the fact that Stockfish 8 is second rated, and lower you go, lower that fractio should be.
Also, although in direct match Stockfish 8 beats Komodo 11.2.2:

Code: Select all

220 (   37,  159, 24),  53.0%

On the rating list Stockfish 8 is 16 ELO points lower. And this is repeated in some other rating lists. Also, by regression, I derived that Komodo gains about 15-20 ELO points from its contempt on your rating list, because only one engine is slightly stronger, and some 13 are significantly weaker.

I think Houdini 5 is also too strong for contempt, even with the last Komodo. That's make 2 engines in fact, specially knowing Houdini is coming soon.

Ok, 2 slightly stronger or almost equal, and 13 much weaker on that list. I did do the full regression, and it came at some 15 ELO points gain for Komodo in IPON and some other similar lists.

[Lyudmil Tsvetkov]

The new Andscacs?

why not?

[IWB]

I made a comment at 50% and don't want do more before it is finished, but what if you are wrong in some or many of your conclusions

[Laskos]

Maybe something then was wrong in your test or the data was too few.
...

In the test you have seen a contempt of 20 was useless for a rating list (no doubt here). Maybe the draw problem (if any) is not related to the contempt?!

I would rather trust my data and doubt the theorie than the other way around ...

Ingo

One possible issue is that Contempt in Stockfish is broken or at least too rudimentary. Komodo has much more involved Contempt, maybe that's why it is gaining significantly ELO in rating lists such as yours.