Not logged inRybka Chess Community Forum
1 2 Previous Next  
- - By Uly (Gold) [mx] Date 2010-05-20 05:52
There probably won't be a Human or Dynamic personality for Rybka 4, and so, KingHunter or Wild personalities are out of the question (though Rybka is wild as it comes, anyway).

Instead, we get this!:



:)

(don't mind the colors...)

I thought many members would be dreaming of this.
Parent - - By BigBen (****) Date 2010-05-20 07:02
Hi,
    Not sure I believe the values shown for the pieces are the ones being used ... These classic values 1, 3, 5 etc I thought had long changed?

Regards
Parent - By Kapaun (****) [de] Date 2010-05-20 07:10
Just consider it to be some meaningless numbers. The question is only: Will you raise or lower them?
Parent - - By grolich (***) Date 2010-05-20 18:38
Probably just to simplify.
So the 300 for knight most probably represents the default value given to a knigh, be it 300, 301, or 295.
Your changes are computed with respect to the default, so if you raise by 10% to 330, and the internal value is 310, the value really changes to 341.

That's what makes sense to me, anyway.
Parent - - By Uly (Gold) [mx] Date 2010-05-20 19:15
Apparently, it's not a % value, so if you set knight to 250, and the "true" value is 330, the value will change to 280 (i.e. you subtracted 50).
Parent - - By turbojuice1122 (Gold) [us] Date 2010-05-20 20:22
Oh, so someone could unwittingly set a piece to a negative value without realizing it!  It would be funny if the first thing the computer tries to do is to get rid of that piece at all costs!
Parent - - By Vempele (Silver) [fi] Date 2010-05-20 20:33
The minimum is 0 for all pieces.

> It would be funny if the first thing the computer tries to do is to get rid of that piece at all costs!


I played a game with the Queen values at zero. In practice she'll want to trade the queen for a minor piece (a queen-for-pawn trade is always in the horizon, so there's no need to do it immediately).
Parent - By Sesse (****) [ch] Date 2010-05-20 20:50
With beta 10 this seems to work, since the sign is wrong. I had a game where seemingly white's primary goal was not taking any of black's pawns, while black was happily eating white's pawns.

Nevertheless, white won in the end, after seeing a mate black was too busy eating pawns to find...

/* Steinar */
Parent - - By InspectorGadget (*****) [za] Date 2010-05-20 07:32
Uly,

people have been crying about that, so I believe that is good move by Vas. Now everybody can start tempering with those values and make their own personalities. Nice! :)
Parent - - By yanquis1972 (****) [us] Date 2010-05-20 07:40
how exactly does that constitute personalities...? i don't know, surely someone much smarter than i will be able to make a tal or karpovian style of chess emerge from a rearrangement of piece values...
Parent - By InspectorGadget (*****) [za] Date 2010-05-20 08:41
You can adjust the piece values to change the style of play. Example, if you've seen Wael Deeb's Zappa settings, you'll notice that Zappa plays much like Rybka. Someone said fiddling with Stockfish values makes it play it bad.
Parent - - By Nelson Hernandez (Gold) [us] Date 2010-05-20 21:05
Check out how it was done in Chessmaster.  Same concept.
Parent - By yanquis1972 (****) [us] Date 2010-05-21 03:26
i never really got into actuallly creating the chessmaster personalities (in terms of making it a hobby) but i do know there are probably 2x the variables, including the kind of stuff you see in zappa and especially shredder such as king safety, aggression, positional/material balance, etc -- and of course strength, although i'm sure with experimenting one could figure this out with cpu usage. it's much much more than piece values.
Parent - - By Fulcrum2000 (****) [nl] Date 2010-05-20 07:32
I think we need some explanation of the different TC items ;-)
Parent - - By Uly (Gold) [mx] Date 2010-05-20 08:04

> I think we need some explanation of the different TC items ;-)


TBD.

Anyway, these are there because R4 time management is very sensitive, so that the optimal value is unknown, R4 will come with some default value that is "okay", the user will have the chance to find better values for their time control/hardware/opponent (yes, the opponent is also a big factor, I've seen Rybka 4b beating the derivatives by a bigger margin than it beats R3 with some TC settings) and find an optimal setting.
Parent - - By pokerpawn (***) [be] Date 2010-05-20 11:58
maybe vas could put a reset button in to get back to these standard ok values if the user messes with them a bit too much ?
Parent - By Uly (Gold) [mx] Date 2010-05-20 17:23
You'd like to create personalities for the default values so they're available at all times.
Parent - - By Lukas Cimiotti (Bronze) [de] Date 2010-05-20 11:47
Most of these TC settings will be kicked out in the release version.
Parent - - By Androcles (**) [fr] Date 2010-05-20 15:54
Please, add a couple of buttons "Load settings" and "Save settings".
Then users will have a great way to define and publish "personalities" :)
Parent - By Uly (Gold) [mx] Date 2010-05-20 17:25
This is a job for the GUI.
Parent - - By staylor (***) [il] Date 2010-05-20 07:46
Was it so useful in Rybka 3?
Or was it like just speculative ideas, without soundness?
Parent - - By Uly (Gold) [mx] Date 2010-05-20 08:00

> Was it so useful in Rybka 3?


Yes, there are positions that only Dynamic, or only Human were getting right, and they were the fastest in finding the right moves (though I never tried the derivatives much). Now you can create your own Dynamic or Human, or Whathaveyou personalities.

It's also potentially useful for corr games, where e.g. Rybka insists on trying a plan on exchanging queens that you know is bad, you just increase your queen value, and Rybka will avoid queen exchanges, etc.
Parent - By rivaldo (***) [de] Date 2010-05-20 09:30
this is indeed very useful. stick with it!
Parent - By yanquis1972 (****) [us] Date 2010-05-20 14:05
how would you go about creating a dynamic personality, for example? or human? i do recall human was essentially just changed material values, but in this case we don't even know what the real values are as baseline.
Parent - - By Androcles (**) [fr] Date 2010-05-20 15:57
I often wondered if piece values should not be allowed to be dynamically recast (e.g. when the game is entering a new phase or with certain type of positions). Any experiments with this?
Parent - By Uly (Gold) [mx] Date 2010-05-20 17:26
Pro Deo experimented with this, and found the improvements were so small it wasn't worth the hassle.

This probably will be in Remote Rybka.
Parent - By Geomusic (*****) Date 2010-05-20 20:38 Edited 2010-05-20 20:54
It is my uninformed opinion that yes, eventually-- this will be a good idea! When monte-carlo is more of a factor tied within rybka's evaluation function. Until then, and without such a dynamic piece value, any dynamic adjustment would be tedious for the program, i.e. substantive knowledge added to benefit this variable would slow down it's search diametrically without increased search efficiency and/or more processing cores.
Parent - - By Dragon Mist (****) [hr] Date 2010-05-20 08:12
Log and xLog things are ...?
Parent - By Uly (Gold) [mx] Date 2010-05-20 08:24
I don't know the details, they are so Rybka writes to a file the communication with the GUI and what is going on in general, I think. Probably will not be there in release.
Parent - - By Lukas Cimiotti (Bronze) [de] Date 2010-05-20 11:49
Log will be kicked out, it's only for testing. I hope XLog will stay - it's a very cool way of writing engine output to a file you can read with a browser.
Parent - By Dragon Mist (****) [hr] Date 2010-05-20 14:34
That's the thing I'm looking for! Let's hope it stays for the release.
Parent - - By SpuddyBuddy (*) [ae] Date 2010-05-20 12:34
Hi there

i observe processor...2048, does this mean that i will be able to run on my 20 x 6 core = 120 core ...sgi octane iii , hope so, so i can can cause lucas 1st loss on his cluster., woohoo
Parent - - By Lukas Cimiotti (Bronze) [de] Date 2010-05-20 12:45
You can, but you can only use 12 cores. Btw. my cluster has 112 cores atm. Maybe you'll win one game out of 100.
Parent - - By SpuddyBuddy (*) [ae] Date 2010-05-20 13:14
envy u, btw is it for sale
Parent - - By Lukas Cimiotti (Bronze) [de] Date 2010-05-20 13:36
Rybka cluster is not for sale. My cluster without Rybka cluster wouldn't help you.
Parent - By Birsan Silviu (*) [ro] Date 2010-05-20 13:48

> Rybka cluster is not for sale. My cluster without Rybka cluster wouldn't help you.


Not even Topalov =))
Parent - By Mark (****) [us] Date 2010-05-20 14:25
Well, you could make a nice IDeA tree with the commercial Rybka using 112 cores!
Parent - By SpuddyBuddy (*) [ae] Date 2010-05-20 15:37
makes sense
Parent - - By tomgdrums (****) Date 2010-05-20 15:01
Will the default settings be the best for analysis?

Thanks!
Parent - - By Vempele (Silver) [fi] Date 2010-05-20 15:18 Edited 2010-05-20 15:38
No. Always Score Main Move should be on for analysis but is slightly worse in games, and Display Upperbounds is a useful (cosmetic) option for those who understand it (and use a GUI that shows them).
Parent - - By Kapaun (****) [de] Date 2010-05-20 21:25
Is this valid if using IDeA, too? And is there a difference in this respect between old IdEA (Aq 3) and new IdEA?
Parent - - By Vempele (Silver) [fi] Date 2010-05-20 21:48
Depends on your IDeA settings. If you're using constant depth (with a very high maximum time), ASMM=false might be better. Otherwise, if a task finishes without resolving a fail low, you not only get a bogus eval, you also waste time getting it!

You should set it to false in the old IDeA, because it can end the task right there without letting Rybka look for a better move.

I don't recall the exact details right now, but it's safest to turn on "wait for next depth".
Parent - - By Kapaun (****) [de] Date 2010-05-20 21:55
Depends on your IDeA settings. If you're using constant depth (with a very high maximum time), ASMM=false might be better. Otherwise, if a task finishes without resolving a fail low, you not only get a bogus eval, you also waste time getting it!


Thanks. Yes, actually I'm using IDeA this way.

I don't recall the exact details right now, but it's safest to turn on "wait for next depth".

Why is that?
Parent - By Vempele (Silver) [fi] Date 2010-05-21 04:49

> I don't recall the exact details right now, but it's safest to turn on "wait for next depth".
>
> Why is that?


It avoids the case where IDeA takes the first eval of the last depth and that eval happens to be the old best move failing low.
Parent - - By Androcles (**) [fr] Date 2010-05-20 15:41
I'm afraid this is an interim console for Beta Testers, (Kullberg hinted at this) and not the
dialog the average Joe will be presented with the final Rybka 4.

Hopefully, I'm wrong. Then, Vas should add a load and save pieces value couple of buttons opening a file chooser and everybody can
use for that configuration file a personality name and then publish it ;-)
Parent - By InspectorGadget (*****) [za] Date 2010-05-20 16:04

> Vas should add a load and save pieces value couple of buttons opening a file chooser


I hope Vytron passes this point to Vas. That's a very good point.
Parent - - By M ANSARI (*****) [kw] Date 2010-05-20 17:58
I think this will allow people to come up with optimal settings that we during beta testing will not be able to do.  Basically with these settings you can try thousands of combinations and tune your engine to play the way you like it.  I am sure someone will come up with a substantially better setting than we will be able to come up with during beta testing.  These settings are very sensitive and especially the time controls are extremely sensitive ... and to be honest even after several thousands of games I am not quite sure which is the best setting.  I think we will get close, but the formula for large hardware seems different to the formula for weaker hardware.  Anyway Playchess will change dramatically and most likely the strongest R4 setting will come from there.
Parent - - By Loboestepario (****) [us] Date 2010-05-20 18:33
Well, you will have to establish default settings...a baseline parameter
Parent - By RFK (Gold) Date 2010-05-20 21:40
Ironically, it is characteristic of the learning curve present in Aquarium. It doesn't surprise me.
Parent - By Nelson Hernandez (Gold) [us] Date 2010-05-20 23:20
One thing that seems a little off here is that obviously there are "default" settings that correspond to the piece values shown.  But whereas the increment for the pawns is 1 or 1%, for the queen it is 1 or 0.1%.  I'd be happier (though I might be alone in this) if all pieces were set to a default value of 1.000 and you could move them all from 0.000 to 2.000, where 2.000 would be 2x the default value.  Then you'd have a really full range of possibilities.
1 2 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill