Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 3 to be much more efficient
Poll Should Rybka 3 display higher node counts? (Closed)
Yes, other engines also display high node counts 2 13%
No, this is useless and leads to confusion 4 27%
I don't care 9 60%
- - By premraj_n (**) Date 2007-12-20 14:56
Hi Vas,
Is there any way that you can make next Rybka generate nodes faster like about 1.5times to 2times more than present 2.3.2a's capability and still add 50 or more ELO to the strength.I think it may require you to clean up some unwanted things and find better way to achieve more efficient search.

I know its easier said than done, but still I think its possible.

Thanks,
Prem
Parent - - By Fulcrum2000 (****) Date 2007-12-20 17:54 Edited 2007-12-20 18:01
It's very simple, just change the NPS multiplication/division constant (e.g. from 14 to 7) :)
Parent - By Vempele (Silver) Date 2007-12-20 18:10
That's assuming it's a constant.
Parent - By turbojuice1122 (Gold) Date 2007-12-20 18:46
I say change the parameter to about 0.2--that should increase sales pretty nicely! :-)
Parent - By Felix Kling (Gold) Date 2007-12-20 19:07
Yes, a high node count is just cool, I add a poll about it :)
Parent - - By Uri Blass (*****) Date 2007-12-20 21:15
There is no connection between more efficient search and displayed nodes per second.
I also do not see why more nodes per second mean more efficient search.

Uri
Parent - By Uly (Gold) Date 2007-12-21 01:34
Ah, I think that he's not talking about displayed node counts, but about searched node counts, as if Rybka magically analyzed twice the nodes at a given time.

In practice, it could be harmful, and if it somehow was optimal, as in doubling the speed, it would only give about 50 rating points (wouln't it?)

Anyway, I think that Rybka already has the best search efficiency out there, and that if this search was implemented into other engines, they would become dramatically stronger.
Parent - - By premraj_n (**) Date 2007-12-21 03:47 Edited 2007-12-21 04:07
Hi Uri its not like i would be happy if i see nodes speed faster and there is no difference yet. What i meant was currently say if i start a analysis on a quad cpu like q6600 stock speed and for a given position lets say Rybka takes 1:14 minutes and reaches depth 20 and generates about 25000kN in total and completes its analysis for that depth.
Instead if Rybka 3 could do that in 35seconds to do same thing on same hardware the node speed is twice fast isn't it. That is what i mean.
Its like Rybka performing on a dual system like as though its is playing on a quad.
Also on a quad system like its playing on a Octal core and etc for all time controls.

Also the 50ELO point improvement i said is not related to this.
If this would improve Rybka 3 naturally to say 150 ELO points then also Vas would be working on implementing more knowledge to increase 50 ELO more also that is different.
Parent - - By turbojuice1122 (Gold) Date 2007-12-21 05:14
You can be sure that this has been one of the Rybka team's top priorities for over two years, and it does this better and more accurately than any other engine.
Parent - - By Uly (Gold) Date 2007-12-21 14:24
If he's talking about the effective speed improvement of increasing cores (if I understand him right, he wants 2 cores to be twice as effective as 1 core, and 8 cores to be twice as effective as 4 cores), then I think that Zappa does this better than Rybka.
Parent - - By turbojuice1122 (Gold) Date 2007-12-21 22:17
Just interpreting his wording literally, it is that he is wondering how a dual core can become twice as fast (i.e. he mentions wanting to increase speed 2x on the same hardware).
Parent - - By Uly (Gold) Date 2007-12-21 22:53
So he wants twice what I was saying (2 cores to be x4 as effective as 1 core, and 8 cores to be x4 as effective as 4 cores)... From where the additional speed will come from, I wonder?
Parent - - By Felix Kling (Gold) Date 2007-12-21 23:31
It's somehow a philosophical question if it's in theory possible to more than double the speed by doubling the processors. in practice I'm quite sure that this isn't possible :)
Parent - - By premraj_n (**) Date 2007-12-22 02:41
Hey guys already miscommunication in requirements phase of SDLC.

Always keep in mind I am not asking this so that displayed kN/s will be high and I will be happy.

What i am trying to tell to Vasik is For example lets us take Dual system E6600 running on stock speed and if 64bit version of Rybka 2.3.2a MP is used and if it is generating 240kN/s for a starting position of the game then i want Rybka 3 to generate 480kN/s and be more specific its not for just display purpose I am looking for analysis also getting completed in half the time taken by Rybka 3 if Vasik is able to achieve this.

Basically I am cannot say its easy to do this so I am telling Vas to try for 1.5 times the speed improvement over 2.3.2a i.e. 1.5x240=360kN/s so he may end-up at increase to 1.2times or 1.3times which maynot be big notable difference but which is still a difference.
Parent - - By Permanent Brain (*****) Date 2007-12-22 02:57
It is not required to tell a chess programmer, that faster is better, and desireable. They know that! :-)

But keep in mind, a "fast(er)" engine (in terms of tactical speed, or calculating combinations) is not necessarily the stronger engine. Neither Rybka nor Shredder(9) which was leading before Rybka, are/were fastest, tactically. Faster are for example Hiarcs 9 or King 3.x - although their node rates are relatively small. IMO, a node rate is useful for hardware benchmarks only, anyway and not useful if we discuss improvements in terms of chess strength.
Parent - By premraj_n (**) Date 2007-12-22 04:21 Edited 2007-12-22 04:27
I know he knows but I just want him to try to improve and if u say I don't have to tell him then I won't have anything to tell and I dont have to come to Rybka forums to say something or other so you better keep out of this. Or if u insist i will keep away.
Or do u want me to tell Vas to implement Bishop underpromotion or some position in endgame where rybka lacks knowledge I did all those before 2.3.2 was released and  he didnot fix them so there is no meaning in telling that again and again.
Parent - - By Uly (Gold) Date 2007-12-22 15:27
I actually, I'm pretty sure that Rybka is the fastest engine in town. (For example, I think that Zappa calculates better, but it's just so slow so in the end Rybka's speed overcomes Zappa's better chess.)
Parent - - By Lukas Cimiotti (Bronze) Date 2007-12-25 09:34
You can test your theory by making an engines match at a fixed depth of 1 ply.

I did this just for fun, the result of 100 games was: +95 =5 -0 for Rybka.

I can only imagine 2 explanations:
Either Rybka is much better positionally because of much better knowledge.
Or Rybka and Zappa have different opinions what 1 ply is.

Regards,
Lukas
Parent - By Vempele (Silver) Date 2007-12-25 10:31

>Either Rybka is much better positionally because of much better knowledge.


Search dominates when the depth is that small.

> Or Rybka and Zappa have different opinions what 1 ply is.


Or Rybka's quiescence search is much more accurate (well, most likely more extensive too). It's about not losing material to trivial tactics.
Parent - - By Uri Blass (*****) Date 2007-12-25 14:17
It is clear that rybka and zappa have different opinions what 1 ply is.
I already found with rybka beta that rybka simply thinks that 1 ply is more than 1 ply and strelka is the proof.

Strelka is almost the same as rybka beta but 1 ply of rybka beta is equivalent to 3 or 4 plies of strelka(I am not sure about the exact difference and it is not important)

Note that 1 ply games are based mainly on tactics and not on positional understanding.

Uri
Parent - - By lopezexchange Date 2007-12-29 02:17
Mr Uri,

this is interesting.  this diference in what a ply is considered to be.  does this apply to greater depths as well, or only at depth 1 ply?
Parent - - By Uly (Gold) Date 2007-12-29 02:26
To all plies, engines just count plies differently.
Parent - By Vempele (Silver) Date 2007-12-29 09:41

>To all plies, engines just count plies differently.


Which engines? Open source if possible, please.

Sure, the search is wildly different from engine to engine (as said by Bob Hyatt about LMR: "There are plies, and then there are plies"), but the actual counting is pretty straightforward: quiescence,1,2,3,4,5,6,7... (there are also fractional plies, e.g. quiescence,4,8,12 etc).

Obvious obfuscation doesn't count.
Parent - By Vasik Rajlich (Silver) Date 2007-12-25 15:39
FWIW - generally, I think that it's good to express things simply and elegantly, but any optimization beyond that just doesn't give enough Elo per developer hour.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka 3 to be much more efficient

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill