I think unknown result games should be discarded

[Dann Corbit]

I tried a 2 GB pgn file, and got a few hundred of these warnings:
Warning: unknown game result set to draw

I think it is a mistake to set it to a draw. Discard the game instead.

In your todo list, I saw an item to add other computational methods.
There are three of them here:
ftp://cap.connx.com/pub/tournament_software/prog10.cpp
Here is the permission to use this file:
ftp://cap.connx.com/pub/tournament_software/Re Your chess rating systems.txt

[Dann Corbit]

Directory of C:\eco
12/29/2004 01:26 PM 20,903 junkbase0.dat
[snip of 24102 files]
12/29/2004 01:24 PM 20,903 junkbase9998.dat
12/29/2004 01:24 PM 20,903 junkbase9999.dat
24105 File(s) 503,866,815 bytes

Twenty four thousand and 500 MB worth, in this case.

I was expecting a single file containing Elo ratios.
What are all these files?

[Uri Blass]

I tried a 2 GB pgn file, and got a few hundred of these warnings:
Warning: unknown game result set to draw

I think it is a mistake to set it to a draw. Discard the game instead.

In your todo list, I saw an item to add other computational methods.
There are three of them here:
ftp://cap.connx.com/pub/tournament_software/prog10.cpp
Here is the permission to use this file:
ftp://cap.connx.com/pub/tournament_software/Re Your chess rating systems.txt

I do not understand what is the software that you used for the 2 GB pgn file.

I do not understand who is the person that you saw things in his todo list.
The links did not work for me and I understood nothing from your post.

Uri

[Rémi Coulom]

Directory of C:\eco
12/29/2004 01:26 PM 20,903 junkbase0.dat
[snip of 24102 files]
12/29/2004 01:24 PM 20,903 junkbase9998.dat
12/29/2004 01:24 PM 20,903 junkbase9999.dat
24105 File(s) 503,866,815 bytes

Twenty four thousand and 500 MB worth, in this case.

I was expecting a single file containing Elo ratios.
What are all these files?

Each of these files contains the likelihood distribution of a player.

Here is a little more details about how the program works: in order to compute confidence intervals, a likelihood distribution is stored for every player. The number of bars of this histogram is controlled with the "resolution" command. Two method are available to compute those distribution: "jointdist", that is most accurate, but has a cost exponential with the number of players, and "exactdist", that is an approximation (not a well-chosen command name, I admit. It is called exact because it assumes that the estimated Elo of opponents is their true Elo).

The "writedist" function is available so that you can plot likelihood distributions with an external tool, such as gnuplot or Excel. Each file that it produces contains two columns. The first colum is the Elo rating, and the second column is its likelihood.

Note that those "dist" functions do not work well in case the number of games of one player is really big, as Tim noticed in the other thread. That will be easy to fix in the next version.

Also, thanks for indicating those Elo-calculation methods. Implementing other methods is not my priority, but I will probably take a look later.

R?mi

[Rémi Coulom]

I do not understand what is the software that you used for the 2 GB pgn file.

I do not understand who is the person that you saw things in his todo list.
The links did not work for me and I understood nothing from your post.

Uri

http://remi.coulom.free.fr/Bayesian-Elo/
http://wbforum.volker-pittlik.name/viewtopic.php?t=949

You need to change the "passive mode" of your ftp client to access Dann's site.

R?mi

[Rémi Coulom]

I forgot to reply to this:

I think it is a mistake to set it to a draw. Discard the game instead.

You are right, of course. This will be fixed in the next version.

R?mi

[Dann Corbit]

[snip]
Also, thanks for indicating those Elo-calculation methods. Implementing other methods is not my priority, but I will probably take a look later.

Here is the web site with an extensive discussion about his rating ideas. It is a very interesting read for everyone to examine:
http://www.ratingtheory.com/