Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Vasik How much more elo point do u think u can add to RYBKA
- - By ChessMate (***) Date 2007-07-19 18:29
I have one of ur earliere interviews in which u have said that 3100 could be the ultimate rating for rybka.
But after just 1 year we have almost 3050   so may be u can go much higher than this.
Any plans of shredder bases like  EGTB which will increase elo almost 40-50 points.
Parent - By h1a8 (***) Date 2007-07-19 23:43
shredder bases are a form of bitbases (gives no distance to conversion).
They must be used in conjuction with tablebases to be effective.
Parent - By turbojuice1122 (Gold) Date 2007-07-20 03:06
I don't think that anyone really thinks that incredibly fast 6-men bitbases will increase elo 40-50 points.  I know that with conventional tablebases, hordes of tests have shown that the elo increase is about 5 elo points, and it would seem rather unlikely that 6-men bitbases would increase it more than an additional 10 points.  Time is better spent improving the engine, since it is easier to have a significant elo gain from doing that.
Parent - - By Vasik Rajlich (Silver) Date 2007-07-20 09:20
There is no way that an Elo of 3100 (or anything that starts with a 3) is the "ultimate". We're just getting warmed up.

Re. EGTB - we wait for a compressed format. Eventually, it will come. It's not an Elo question, it's a user convenience question - raise your hand if you want to have 1.2 TB of tablebases on your hard drive.

Vas
Parent - - By Uri Blass (*****) Date 2007-07-20 12:48
I think that the question is undefined because elo is dependent on who play against who.

I see no way to have confidence that it is possible to get elo that starts with 4 even if players do not play the opponent and do not play for a draw against a superior opponent.

It is possible that the rating of machine that is (2^n) times faster than the 3100 player is given by the formula
3100+               (50+50*0.9+....50*0.9^(n-1))=
3100+50*(1-0.9^n)/(1-0.9)=3100+500*(1-0.9^n)

This formula means that you never get better rating than 3600.

My personal opinion is that we will never see rating that starts with 4 in comp-comp lists like CEGT or CCRL.

Uri
Parent - - By Felix Kling (Gold) Date 2007-07-20 13:01
Imagine a program that doesn't make mistakes AND knows how to maximize its winning chances... I don't know why this shouldn't be possible, but I agree that we won't see that in the next years. What I expect is +50 ELO per year - this means ELO 4000 in 18 years... maybe too optimistic, but who knows?
Parent - By Uri Blass (*****) Date 2007-07-20 13:36
I think that you are too optimistic and I think that there is going to be diminishing returns from a year.

You will probably see more than +50 thanks to software and hardware improvement during 2007 and 2008 but later you are going to see a smaller improvement.

Uri
Parent - - By George Tsavdaris (****) Date 2007-07-20 13:47
As Uri have already said, improving yourself is not the only critical factor for improving ELO. Your opponent's ELO is the other important factor.

We see from CEGT 120/40 that:
1 Rybka 2.3.2a 64 2CPU           : 3045   21  21   750    77.5 %   2831   34.7 %

Imagine Rybka would only draw a game out of all these 750 games!
So it would have  749.5/750. that translates to a 4101 ELO "only" due to opponent's mean 2831 ELO.

(Actually this is not so correct, since the procedure of determining opponent's mean ELO in this big list is more complicated, but the result would not be so different as that i gave.)

Imagine now a more reasonable situation that Rybka would draw the 10% of its games and never lose a game.
With 10% draws then we would have  (750-75)+75/2 = 712.5 out of 750 games.
So we would have 3343 ELO for this Rybka.
Winning all these games and draw only 10% would mean scores of 95%:
Rybka - Fritz  +46 -0 =4
Rybka - Hiarcs +45 -0 =5

Super amazing results! What can be done better? But still "only" a 3343 ELO.
Parent - - By Felix Kling (Gold) Date 2007-07-20 13:51
But you could test it against other Rybka versions- Rybka 20 vs. Rybka 19 ... maybe in 20 years CCRL will stop testing other engines but Rybka ;) or there's a guy who can create a program as strong as Rybka... again, who knows :)
Parent - By George Tsavdaris (****) Date 2007-07-20 13:58

>But you could test it against other Rybka versions- Rybka 20 vs. Rybka 19


I agree. That's why i would like to see CEGT and CCRL to start matching different Rybka versions .
Perhaps Rybka 2.3.2a and Rybka 2.3.2 is not so interesting and logical, but Rybka 2.3.2a 4 CPU versus Rybka 1.1 1 CPU, is perfectly logical and USEFUL for giving valuable (ELO) information.....
Parent - - By George Tsavdaris (****) Date 2007-07-20 13:55 Edited 2007-07-20 14:13
Where did you get the 0.9?
Perhaps 0.9 is actually 0.95 so we would have:
3100 + 50·(1-0.95^n)/(1-0.95)=3100 + 1000·(1-0.95^n) = 4100 when n is too big.

Or even the situation could be that:
A machine that is (2^n) times faster than the 3100 player has ELO:
3100 + 50·n  so we would have no limit.

Although i doubt and your formula with a number between [0.9 , 1) is more logical. But the fact is we don't really know.
Parent - By Uri Blass (*****) Date 2007-07-20 14:08
3100+50n is impossible because when the computer solve chess it cannot get higher rating.

I agree that we do not know and my opinion is that rating that is more than 4000 is impossible.
Uri
Parent - By SillyFunction (**) Date 2007-07-20 15:28
Ah, then you've just revealed your ultimate goal, Vas. You want to create God Father of the chess world!
Parent - By herakleitos (***) Date 2007-07-20 15:51
> We're just getting warmed up.

Scary. Like facing Aljechin in 1930 or Fischer in 1971...
Parent - - By turbojuice1122 (Gold) Date 2007-07-21 01:19
Of course, I've thought before that 4000 elo is impossible, but reading the article on checkers being solved in the Chessbase site makes my belief even more firm.  Really good human players were getting draws against this machine when it was basically on the verge of solving the game.  The humans were undoubtedly making a few mistakes, but these mistakes weren't serious enough to cause a loss.  Obviously chess is a lot more complex, but I think that when computers, due to various abilities in pruning and programming are effectively able to search past depth 30 or 35 (i.e. twice the depth required for checkers--I do this because depth is related to the logarithm of the number of possible positions, and the number of possible positions in chess is roughly the square of the number possible in checkers) in typical middlegame positions, then we'll be in a similar situation in chess.  It seems very difficult to believe that Rybka able to search to a depth in the thirties (and with better knowledge) would give it 4000 elo. 
Parent - - By M ANSARI (*****) Date 2007-07-21 07:29
It is hard to really estimate an engine's ELO rating because it is relative to hardware used.  But if we were to compare ELO based on engines playing today on today's hardware I would say that if Rybka today is 3100 ELO ... the highest possible ELO it can get would be around 3200 to 3250 ELO.  I say that will the knowledge that if you had asked me that same question 2 years ago ... I would have estimated the highest ELO possible to be around 3100 ELO.  That would be mainly due to improved hardware in the future and ofcourse improved software of the engine itself.  This is ofcourse a guess and that could be wrong .... but that is what I believe is the "wall".  Rybka still plays poorly in some odd positions and some endgames ... but I am not sure that can be worth more than 50 ELO points.  Another thing to realize is that as engines get bigger hardware they might be able to see very deeply into a position and thus not play many interesting moves because they can see the refutation.  Thus I believe you will see a very high number of draws and the engine will seem to play too conservatively.  I guess changing the contempt to force the engine to play weaker moves with better chances for a win (based on weaker play of opponent) is possible ... but I am assuming a Rybka that would play the best move it sees in the position with no contempt factor.  Another possibility to gain ELO points against the weaker engines is to allow the engine to be smart enough to go into openings where while the position can be even or even slightly worse (especially unbalanced material positions), the stronger engine can slowly outplay the weaker one with more precise play.  But it will always be a guess ... and my guess would be as good as any other.  But 4000 ELO I think is unrealistic.
Parent - - By Vasik Rajlich (Silver) Date 2007-07-22 07:45
FWIW, when Zappa played in Turino on 512 processors and made a ton of draws, this was Anthony's explanation (or at least candidate explanation) - that the deep searches were discouraging Zappa from trying many interesting things.

Personally, I think these walls and too-deep searches are all really far away.

Vas
Parent - - By VP (*) Date 2007-07-24 01:09
FWIW, when Zappa played in Turino on 512 processors and made a ton of draws, this was Anthony's explanation (or at least candidate explanation) - that the deep searches were discouraging Zappa from trying many interesting things.

Personally, I think these walls and too-deep searches are all really far away.

Vas


Vas,
In effect, are you insinuating that since the engine was too strong, it was discarding certain lines that it may have played against weaker opponents?
Can we not fine tune engines based on the difference of ELO to play more enterprising chess?
Say, had Zappa played a little more enterprisingly (contempt factor increases ;-))- and the formula for contempt factor was a function of the difference in ELO, then we would have seen imperfect-but enterprising chess.
What are your views?
regards,
Varun
Parent - By Vasik Rajlich (Silver) Date 2007-07-24 06:56
When you get a huge machine on which you can't test before the event, there are just too many variables and it's really impossible to draw conclusions. Personally, I think it was either bad luck or some subtle bugs.

Vas
Parent - By Jitugo (**) Date 2007-07-22 09:43
Hi,

why don't you just take the scorpio bases. They are free, open source and compressed. You also only would have to put the source for the scorpiobase access in the distribution of rybka. They are already used in a couple of programs. Did you look at these, and if so did you find anything preventing you from using them?

regards,

jitugo
Parent - - By Banned for Life (Gold) Date 2007-07-23 01:55
My hand is raised. Unfortunately, the easiest way for me to get them would be to fly to the DC area, break into Nelson's house, and copy them. To my credit, I only seriously considered this for a short while. :-)

Regards,
Alan
Parent - - By Vasik Rajlich (Silver) Date 2007-07-24 06:56
Well, as long as you didn't book the plane tickets ..

Vas
Parent - - By VP (*) Date 2007-07-24 06:59
Maybe the tickets were not discounted ;-)
Parent - By Vasik Rajlich (Silver) Date 2007-07-24 07:44
If Alan starts playing on Playchess with all 6-man TBs shortly after air prices drop then we can start getting suspicious :)

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Vasik How much more elo point do u think u can add to RYBKA

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill