The evaluation value and value returned by minimax search

[diep]

Is this again some KGB crap?

Even Uri Blass understands what you wrote is not correct and he gives a disproof for it, just using a few moves Marcel.

Are you understanding what he wrote down?

[zamar]

[snip]
Sorry, but your writing is complete crap and disinformation. You don't seem to understand much about the subject.

So the guy who parameter tunes himself using positions his own program based upon the evaluation difference of positions,
he writes about me that i spread 'desinformation and crap'.

Whereas the only out of us 2 who is clearly lying, is you, as you encouraged the Chinese guy to NOT do the same.

Now *that* is hypocrisy.

I wrote about one tuning experiment which failed. Of course it's possible that it works, but I got the details wrong... I'm not intentionally encougaring/discouraging anybody.

Your main vehicle of tuning is exactly this - you obviously spit out desinformation here deliberately that it failed for you, as it didn't.

It's your main form of tuning.

In case you didn't know, our main tuning method is documented in here:
http://chessprogramming.wikispaces.com/ ... ing+method

I've of course tried various alternatives, but the experiment I told about this thread failed. I've also heard that Komodo team tried something similar and failed too (not 100% sure about this)

[zamar]

Is this again some KGB crap?

Even Uri Blass understands what you wrote is not correct and he gives a disproof for it, just using a few moves Marcel.

Are you understanding what he wrote down?

Yes, but you don't seem to have a clue of what I'm talking about.

[diep]

[snip]
Sorry, but your writing is complete crap and disinformation. You don't seem to understand much about the subject.

So the guy who parameter tunes himself using positions his own program based upon the evaluation difference of positions,
he writes about me that i spread 'desinformation and crap'.

Whereas the only out of us 2 who is clearly lying, is you, as you encouraged the Chinese guy to NOT do the same.

Now *that* is hypocrisy.

I wrote about one tuning experiment which failed. Of course it's possible that it works, but I got the details wrong... I'm not intentionally encougaring/discouraging anybody.

Your main vehicle of tuning is exactly this - you obviously spit out desinformation here deliberately that it failed for you, as it didn't.

It's your main form of tuning.

In case you didn't know, our main tuning method is documented in here:
http://chessprogramming.wikispaces.com/ ... ing+method

I've of course tried various alternatives, but the experiment I told about this thread failed. I've also heard that Komodo team tried something similar and failed too (not 100% sure about this)

Initially there was a posting from Marco Costalba claiming Stockfish was getting tuned by playing sets of 1000 games 1 minute all.

We did not get any answers then on how you could actually measure elo differences between engines then, as independantly several authors concluded they needed 10k to 40k games for that as a minimum.

Then some years later you posted yourself some tuningsstory on that you're now playing very superbullet games and 5000 games each set.

You really think we're eating each time a new tuningsstory as being the truth?

After years of playing at 128+ cores, crafty still is badly tuned, unlike Stockfish.

Oh by the way it needs 40k games for each run. 5000 really is getting closer than 1000, yet to progress 0.5 elopoint that won't be enough.

This whereas the deep blue paper is pretty clear huh. Add statistical significance to what they were doing back then, and you're getting exactly to the values that you ended up in SF.

Why each time a new story here and discouraging the Chinese guy to do something that in the first place had your full attention.

In fact i very well remember some online conversations i had of a few years ago, where very precise was formulated what's getting done and what you guys were exploring/wondering about.

Vincent

p.s. if you automatically tune playing games you tune far more passive than with other methods. It's easy to figure out therefore which engines were tuned by playing games and which are not.

[mcostalba]

Initially there was a posting from Marco Costalba claiming Stockfish was getting tuned by playing sets of 1000 games 1 minute all.

Sorry to chime in, just to calrify that it was not tuning. Testing on 1000 games at 1' per game was used to validate a new feature (not auto-tuning a parameter).

And it worked ! If you look at SF commit history in github you will see that until about 1-2 years ago many changes were committed after 1K games of testing. Today SF is much more mature and finding an increase with just 1K games is not realistic, but at the time the changes were in the range of 10-15 ELO each and the signal/noise level was high enough that even 1K games validation worked most of the times and, on average, let SF to steadily increase ELO from Glaurung.

[diep]

Initially there was a posting from Marco Costalba claiming Stockfish was getting tuned by playing sets of 1000 games 1 minute all.

Sorry to chime in, just to calrify that it was not tuning. Testing on 1000 games at 1' per game was used to validate a new feature (not auto-tuning a parameter).

And it worked ! If you look at SF commit history in github you will see that until about 1-2 years ago many changes were committed after 1K games of testing. Today SF is much more mature and finding an increase with just 1K games is not realistic, but at the time the changes were in the range of 10-15 ELO each and the signal/noise level was high enough that even 1K games validation worked most of the times and, on average, let SF to steadily increase ELO from Glaurung.

The discussion back then WAS about automatic parameter tuning.

Not about validation at all...

Basically you're trying to talk your way out now giving 2 'new arguments'.
Arguments never used back then. Just a deadly silence there was back then, when questions were asked.

And about 'githubs', you posted that that tuningsframework of SF was integrated into SF and your own proprietary code and topsecret therefore, which is why nothing of the code, especially not how you modify parameters, was shown.

Of course we see nothing about all that - you tune in a different manner all those years, as is so obvious to all readers who followed over the years those arguments and discussions, and now a Chinese guy who is on the right path gets 100% DESINFORMATION from your buddy.

Not only a deadly silience, which is very legal, but a posting claiming it didn't work for you to tune in the manner how some of the top engines get tuned, including your 'own'. It's ok to say nothing if something works for you, but spreading 100% desinformation is just very very sick.

[zamar]

Not only a deadly silience, which is very legal, but a posting claiming it didn't work for you to tune in the manner how some of the top engines get tuned, including your 'own'. It's ok to say nothing if something works for you, but spreading 100% desinformation is just very very sick.

You really sound like, you should get some rest...

Nothing I say could convince you, so I am not going to even try...