Engine Tournament

[q30]

The rating of Phoenix LeelaZero versions (details):
1) 0.16
2) lz_v0.33(0.15)
3) lz_v0.32(0.15)
4) lz_orig(0.15)
5) lizzie(0.15)
6) 0.14

But I don't understand, what is the reason of all these versions creating without the neuronet updating...

[spook]

Yes, the description is the same and the lists are different (the list was updated: was added Zenith and was removed Ray(RLO) because of about sense). What is the problem

I never have problems, only challenges.
It wasn't obvious, but now it is. Thanks.

[as0770]

This is really ridiculous, how You read posts...
Where You had found in the post by this link word "Phoenix" (or "Феникс")?!

We did not talk about Phoenix but about the significance of the results you posted. You pervert the facts once again. I just don't know if you do it on purpose or ignorantly.

[Uberdude]

as0770, why not ignore this thread instead of engaging in these pointless arguments?

[as0770]

as0770, why not ignore this thread instead of engaging in these pointless arguments?

Oh well, I tried to delete my account here but I failed. So I will go on answering if someone tells nonsense. But feel free to delete my account.

[AloneAgainstAll]

3 nov 2018

I promise this is my last post in this thread...

25 jul 2019

We did not talk about Phoenix but about the significance of the results you posted. You pervert the facts once again. I just don't know if you do it on purpose or ignorantly.

I see strong contradiction here.

[as0770]

3 nov 2018

I promise this is my last post in this thread...

25 jul 2019

We did not talk about Phoenix but about the significance of the results you posted. You pervert the facts once again. I just don't know if you do it on purpose or ignorantly.

I see strong contradiction here.

Congrats, you found contradiction in a public forum.

I know I shouldn't read posts by blacklisted people, but now and then I forget my deliberate intention and read them because the forum software don't allow to blacklist someone completely. And then I feel this inner constraint and have to answer to the nonsense.

BTW I ignored it for a long time, but since I started this thread this guy wrote deprecative comments to me. I could live with it if there would be at least some basic expertise. But neither there is some know how, nor he understands english.

So after all I think it is OK to answer him once in a half year. Judge yourself if it is OK to blame me for that.

[q30]

Yes, the description is the same and the lists are different (the list was updated: was added Zenith and was removed Ray(RLO) because of about sense). What is the problem

I never have problems, only challenges. :)
It wasn't obvious, but now it is. Thanks.

I had thought, that my post has problem...

[q30]

This is really ridiculous, how You read posts...
Where You had found in the post by this link word "Phoenix" (or "Феникс")?!

We did not talk about Phoenix but about the significance of the results you posted. You pervert the facts once again. I just don't know if you do it on purpose or ignorantly.

I don't know, what abstract significance You had told about, but I had told specifically about the significance in first case of original LeelaZero, when there was small overweight in the account in a match with small number of games, but big number of playouts,that was confirmed in a match with big number of games, but small number of playouts, and about the significance in second case of LeelaZero Phoenix (that post was about what), when there was big overweight in the account in a match with small number of games, but big number of playouts (and there wasn't any additional matches).
So there wasn't any fact perversion, but was Your inattentive reading of posts...

[q30]

The rating of original (weights file format) LeelaZero versions (details):
1) 0.17
2) lz_next_190222(0.16)
3&4) 0.16&0.14
5) 0.15(all)

The only 0.17 version is significantly stronger than previous.

One "trick": for getting not only the LeelaZero version, but and used weights file, replace in GTP.cpp the stroke

Code: Select all

gtp_printf(id, PROGRAM_VERSION);

by the one

Code: Select all

gtp_printf(id, ' %s + %s', PROGRAM_VERSION, cfg_weightsfile.c_str());

before engine compiling.

[as0770]

in a match with small number of games, but big number of playouts,that was confirmed in a match with big number of games, but small number of playouts,

Your main error in reasoning is that the statistical significance of a result will not increase with a higher number of playouts but only with a higher number of games.

[jann]

the statistical significance of a result will not increase with a higher number of playouts but only with a higher number of games.

This is not entirely correct. Higher playouts reduce the random factor in individual matches somewhat, making the result more representative. OC this is a weaker effect than the statistical validity coming from the number of samples (only increasing the weight of samples towards 1, whereas a match on low playouts may only worth 0.7, for example).

[as0770]

the statistical significance of a result will not increase with a higher number of playouts but only with a higher number of games.

This is not entirely correct. Higher playouts reduce the random factor in individual matches somewhat, making the result more representative. OC this is a weaker effect than the statistical validity coming from the number of samples (only increasing the weight of samples towards 1, whereas a match on low playouts may only worth 0.7, for example).

In a match with x playouts the winning chance for an engine is y %. There is nothing like a random factor. What you mean is: Results with a higher number of playouts are more representative for the engines strength. Of course that's true. But that don't mean you get a statistical significant result with less games.

[Bill Spight]

What you mean is: Results with a higher number of playouts are more representative for the engines strength.

The number of playouts is one parameter of an engine's strength.

[jann]

There is nothing like a random factor.

Without random factor the stronger net would always win (and the games may even be identical).

A winrate of eg. 54% may go up to 58% with quadruple playouts. This 58% makes slightly more statistical mass from the same number of games (because each sample weights nearly 1, while at very low playouts game results are more random, thus weight less than 1 - carry less information).

Results with a higher number of playouts are more representative for the engines strength. Of course that's true. But that don't mean you get a statistical significant result with less games.

The same number of more representative samples weights more than the same number of less representative samples. Maybe you understand better from a specific example: 102 games with 200 playouts are statistically less significant than 101 games with 2000 playouts (a weaker effect as mentioned).