Rybka vs Zappa - An Author's Perspective

[CRoberson]

Its a great match! Also, a waycool coin toss.

Going into the match it was almost a coin toss for
who is going to win. Its only 10 games! So, I tossed a coin
10 times and had heads 7 times. Is that statistically unlikely?
No! It is not with only 10 tosses.

So, lets look at ratinglists using 4 procs or more and long TC and
have a large number of games. In a quick search, I found CCRL
has a 4 proc 40/40 list. Not very long TC but best I could find.

The rating difference between the two is 66 points (3117 to 3051).
That puts the match odds at 59.14% in favor of Rybka for a long running
match. The thing to remember is that the probability of draws ( P(D))
has a considerable influence on a small match such as 10 games.

In a 100 game match, Rybka should gain 59% of the points. This
could happen by any combination between two extremes.
First, Rybka could win 59 games and Zappa could get 41 wins.
Second, Rybka could win 18 games and draw 82 games.

There are two key statistical points here.

• 1) Each outcome is equally likely.
  2) Statistics can not and do not specify the order of individual game results.

Given that the Mexico match is 10 games, point #2 is the most
important. So, a 10 game match could result in all draws, All wins
for either side, an even number of wins for either side or a score
of 5.5 to 4.5 with several permutations.

Even better yet, the individual game odds are not equal to the
match odds of 59%! This is due to the effect of draws.

The max wins Rybka should get, at 59% match odds, is 59 of 100.
The min wins Rybka should get is 18 of 100 - Zappa could gain its 41 pts
by 82 draws. If we average the two we get the probability of Rybka
winning a specific game (not match).
Thus, P(Rybka wins a specific game) is 38.5% - P(Draw) is large.

More practically speaking, the oddities of two of the games and
some tests I've run indicate there might be a bug in Rybka's
parallel search. (I said "might" be a bug.) If that is so, it could be
exaggerated by using an 8 proc machine.

[CRoberson]

I left out two important perspectives:

• 1) Both sides have their Book Authors available or on site.
  2) Rybka has the program author on site.

Point #1: This is a very dynamic issue. The book can make a big
difference and the book authors on both sides are very
experienced. Also, the Rybka team is full of Chess Masters.
This could be an edge for Rybka. Eventhough, Rybka is currently
behind and Erdo is very experienced at this.

Point #2: The program author is available. Add this to the possibility of
a bug in the parallel search and there could be an edge for Rybka.
However, any chess program author will tell you that changes made
with less than a weeks worth of testing can be very risky. Thus,
I don't expect Vasik to make search code mods during the tournament,
but he could.

It is far easier to redirect the game via book but that isn't easy.

There is a chance for Rybka to win the match.

Who or what am I rooting for?
I am looking for an exciting match and for the last games
to be excellent.

[Nid Hogge]

I Agree mostly.

1) Both sides have their Book Authors available or on site.

Yes, Very crucial. we have seen the games very dependent on the openings. To be expected in tournaments with small # of games.

2. Indeed, But I think Anthony is there as well.



Check out Susan Polgars blog:
http://susanpolgar.blogspot.com/2007/09 ... chine.html

[playjunior]

I think those testings are rather irrelevant: because engines are so book-dependent. What is the statistical significance of Rybka scoring 60% against Zappa in 1.c4 c5 when both teams played almost exclusively Ruy Lopez in the match?

[Uri Blass]

I think those testings are rather irrelevant: because engines are so book-dependent. What is the statistical significance of Rybka scoring 60% against Zappa in 1.c4 c5 when both teams played almost exclusively Ruy Lopez in the match?

Correct.

Another important point is that not the commercial version of rybka is used but something different.

Uri

[Mike S.]

The rating difference between the two is 66 points (3117 to 3051).

Did you notice that on the CCRL blitz 40/4m list, Zappa Mexico has already 520 games and it is a difference of 71 Elo between 2 CPU (R.) and 4 CPU (Z.)? Blitz or longer, whatever-book, if Zappa Mexico 4 CPU is 71 Elo behind "old" Rybka 2.3.2a 2 CPU, I do not expect that Zappa Mexico can make it to rank #1 anywhere if equal hardware is compared, in the long run.

The better scaling as programmer's call it, has allowed Zappa Zanzibar to reduce the difference to Rybka by 64 Elo, at CCRL 40/40m rankings:

Code: Select all

Rybka 2.3.2a 32-bit            3015
Zap!Chess Zanzibar 32-bit      2885 -130
-----
Rybka 2.3.2a, 64-bit 4CPU:     3117
Zap!Chess Zanzibar 64-bit 4CPU 3051 - 66

Nevertheless, on 2 CPU or more it has the clear #2 spot behind Rybka at the moment (and for quite a time now) and that is very strong and respectable! I am not somebody for whom only the number one counts...

[Uri Blass]

The rating difference between the two is 66 points (3117 to 3051).

Did you notice that on the CCRL blitz 40/4m list, Zappa Mexico has already 520 games and it is a difference of 71 Elo between 2 CPU (R.) and 4 CPU (Z.)? Blitz or longer, whatever-book, if Zappa Mexico 4 CPU is 71 Elo behind "old" Rybka 2.3.2a 2 CPU, I do not expect that Zappa Mexico can make it to rank #1 anywhere if equal hardware is compared, in the long run.

The better scaling as programmer's call it, has allowed Zappa Zanzibar to reduce the difference to Rybka by 64 Elo, at CCRL 40/40m rankings:

Code: Select all

Rybka 2.3.2a 32-bit            3015
Zap!Chess Zanzibar 32-bit      2885 -130
-----
Rybka 2.3.2a, 64-bit 4CPU:     3117
Zap!Chess Zanzibar 64-bit 4CPU 3051 - 66

Nevertheless, on 2 CPU or more it has the clear #2 spot behind Rybka at the moment (and for quite a time now) and that is very strong and respectable! I am not somebody for whom only the number one counts...

Zappa mexico has more games in the ccrl blitz list

Rybka 2.3.2a is significantly stronger in blitz even with single processor.

Here is a link for comparison(to get it I made custom comparison by comparing between zappa and zeus and replaced zeus with rybka

http://computerchess.org.uk/ccrl/404.li ... librate=no