back


29164WCN
2019-02-22 12:16:59
The two nobodies! on abrok.eu/stockfish Argue amongst themselves because they both just realised no one else is reading what they are saying!
So when the Sh#t hits the fan and they both finally find out they are both nobodies! because there internal chat ? is nonsensical! :(


29157 queen13
2019-02-21 23:05:16 I might as well treat you like arb.

I looked first 10 pages and there's nothing of the sort you are describing.


29160 noname
2019-02-22 00:01:04 You offend me queen13.
Have I ever insulted you.
I could treat you as a troll.
It's time that I end all communication with you.
You are a bad person.



WCN
29163noname
2019-02-22 05:29:52
29161 But at least I'm not so stupid like you to think tcec can't afford to play chess games.
That is so funny.
Haa ha haa
29162noname
2019-02-22 05:18:03
That's me.
You know me so well.
29161SilverFang
2019-02-22 05:00:32
@29160 Hahaha I can imagine what a poor moe sobbing loli you are!
29160noname
2019-02-22 00:01:04
You offend me queen13.
Have I ever insulted you.
I could treat you as a troll.
It's time that I end all communication with you.
You are a bad person.
29159noname
2019-02-21 23:28:17
You can be quite insulting queen13.
I see you as a bad person.
29158queen13
2019-02-21 23:07:59
Stockfish has no regression. Either it was unlucky, or lc0 is really that strong or a

combination of those 2. That's what I think is most likely the case.
29157queen13
2019-02-21 23:05:16
I might as well treat you like arb.

I looked first 10 pages and there's nothing of the sort you are describing.
29156noname
2019-02-21 22:46:58
The deeper you look the more you will see.
29155queen13
2019-02-21 22:41:32
I did, they are nowhere.
29154noname
2019-02-21 22:10:10
You're welcome anst.
29153noname
2019-02-21 22:08:10
They are here.
Try to look harder.
29152anst
2019-02-21 20:16:37
@noname 29110

Thanks for the interesting analysis. I will study it more closely later.
Best regards
anst
29151queen13
2019-02-21 17:55:40
Which page are you talking about?
29150noname
2019-02-21 17:30:32
Have you found them yet ❓
29149queen13
2019-02-21 17:18:54
I couldn't in first 2 pages
Which page should I look ?

I hope you don't mean not functional or optimization changes
29148noname
2019-02-21 12:45:24
Yes I can find them.
Can't you ❓
29147queen13
2019-02-21 12:27:36
Well, at least it can't be that they run ltc tests and then allow everything to pass

regardless of the result don't you think? Also, if it's true that

The list of patches that passed stc and failed ltc is extensive but all of them were passed

then those patches that passed stc and failed ltc but passed can be found here.

Can you find them?
29146noname
2019-02-21 12:12:52
I know what he wrote.
It's the interpretation of the post I'm having difficulty with.
You interpret it differently.
29145queen13
2019-02-21 11:49:37
He took some patches that passed stc and noticed what happens at ltc:

<Total number is 17 red 7 green and 11 yellow.
about half of the patches that pass STC have negative score at LTC and it suggest that many patches that are good at STC are bad at LTC>

so 17 of those were red, only 7 green and 11 yellow.
So many patches that are good at STC are bad at ltc.

Then he said that this means that maybe even passing ltc does not mean that
it is a possitive change for even longer time controls like tcec for example.

He said this:

<The fact that passing [0.5,4.5] at (10+0.1) does not mean not having big probability to be red with [0,3.5] at (60+0.6) suggest also that even passing [0,4.5] at (60+0.6) often does not mean improvement at (360+3.6) time control and practically patches only need to pass [0,3.5] at LTC that is even worse.>

All these in the original post.

So then someone replied this:
<No.
I've tried stuff that passed [0;5] back lately and in multiple variations of the same idea and guess what. Nothing of this passed [0.5; 4.5]
So I can say that passing STC a lot of the times is just luck. Patch is worth +0.3 elo but gets lucky and performs as +3 elo at STC but at LTC it shows it real +0.3 elo and doesn't pass. This just happens.>

to which Uri replied:

<1)I do not believe that patches that pass STC are usually worth 0.3 elo at STC espacielly when there are cases when more than one variation of the same idea pass STC even with [0,5,4.5] but fail LTC with [0,3.5].

Based on the name
TropismKDeval2 and TropismKDeval3 are similiar ideas that passed STC and failed LTC
sac1 and sac1^ are similiar ideas that passed STC and failed LTC
SearchTreeG and SearchTreeG^ are similiar ideas that passed STC and failed LTC

Facts are that there are patches that give an improvement at STC but do not give an improvement at LTC.

2)It is logical to believe that the opposite is also correct and there are patches that give an improvement at LTC but do not give an improvement at STC but deciding to test first with [0,5,4.5] at STC means that you have almost 0 probably to accept them.>

So either than the original where he lists a bunch of patches that passed stc but many failed ltc, that's another way for you to know that they don't only test stc.

In 2 he said that testing at stc only may prevent the acceptance of patches that would be good
for ltc as failing stc means not getting tested at ltc.

I don't understand why you think that tests that pass stc and fail ltc are accepted ever...
29144noname
2019-02-21 11:14:59
If anyone knows exactly what Uri Blass meant in his post, then do let us know.
29143noname
2019-02-21 11:13:05
In anycase presumably bounds were changed in order to pass patches that would otherwise have been rejected.
29142noname
2019-02-21 11:09:20
>>
Why do you say that all are accepted?
<<

Uri posted...

"There are many changes that seem good for STC but bad for LTC"

"changes"

That is what is implied by Uri's post isn't it ❓

Else if they're not accepted what's the point of the post ❓

How do you interpret it ❓
29141queen13
2019-02-21 10:21:07
>The list of patches that passed stc and failed ltc is extensive but all of them were passed.

Why do you say that all are accepted?

I went to the development versions here and didn't see that in the first 2 pages.

Maybe Alpha-reduction2 is a speed-up and does not depend on time control, maybe a non-functional

change, or maybe it isn't in stockfish even(How do you know it is?)
29140noname
2019-02-21 10:18:44
In addition to the two that were not tested or fully tested
there was 17 red patches at ltc that were still passed and will probably be bad for SF at tcec.
and 11 yellow at ltc that were still passed and could be bad at tcec.
Only 7 green at ltc passed.

A total of 30 patches that could be bad for SF at tcec.

When a certain position or opening arises in a game which a rogue patch is dealing with, then SF could be forced by that patch to play a passive move.

When that happens multiple times or with more than one rogue patch in the same game SF could be forced by those patches to play a number of passive moves, the effects of which accumulate, and could result in either only a draw from an advantageous position, or a loss.

A loss not only to lc0 but also to other engines like Houdini.

Losing just one game to H would have been unthinkable during div p.
A reasonable person could think the same for later SF dev versions, yet after div p Houdini beat SF in the tcec cup.

The combined effects of patches could be unpredictable when there are so many that fail ltc but have still been passed.

Conversely, in some games with different positions, the problematic code in those rogue patches may not be used at all and SF will not make passive moves, leading to either an easy draw for SF, or a win.

So it depends on what's happening in any particular game.

Uri's point is that they think passing at stc alone means that's fine.

But including patches after bounds were changed that failed at ltc seems crazy suicidal.

next page >
single page