Author: Chris Carson
Date: 12:01:09 01/08/00
Go up one level in this thread
-g You are right. SSDF is based on lots of different programs and hardware. Tony Hedlund is one of the testers and posts here regularly with results and his home page has all his games for download and review. No rating is based on one game or one opponent. No extrapolation as described above is done, based on just one opponent or game. :) The SSDF has a link to Tony's page, has the ratings list, and if you download the complete list, you will see the results for all the programs (head to head results and opponents average rating). http://home3.swipnet.se/~w-36794/ssdf/ The main dispute is two fold, not enough human games (some are included), and the games (with humans) are two old. The argument is that the mean of the data should be moved down 200 points and the spread should be reduced by 100 points (from 150 to 50). I am not in favor of artificial adjustments to the data. I would love to see more human games included. SSDF is willing, the humans are not. :) IMHO. No one has produced any data that disputes the SSDF, only argument with the methodology. IMHO: Facts (data) hold water. I am excited about the Isreal chess league, all teams will be allowed to have programs next year! IMHO: I believe the SSDF is valid for both computer and FIDE comparison, (SSDF documented intention, High correlation with AEGON results, ...). The SSDF also has a FAQ link on thier page for more information. I respect the opinion of everyone. I am influenced by experts such as Dr Hyatt, Ed (author of Rebel), Bertil (SSDF), IM Kaufman, Dann C., GM Soltis (who has written about the strength of computers and how GM's rely on them in Chess Life, and others). IMHO: SSDF is reliable within 3 error of measures for Tournament Performance Ratings (not match, not GM norms, ...). Just my opinion. :) Some GM's are great against computers, others are not. Paid matches with GM's that have time to prepare against the program is not provide a fair TPR, although it is very useful information and shows where that program is against match play GM's on specific hardware (some of the games are on K2-300's, some on Intel boxes, and some on Kryotech AMD 600's, a mix of hardware, so the ratings need to be calculated base on hardwar as well, not mixed as we see published as 2485 so often). :) I do hope everyone can stay on topic. Disagree, but not disrespect. :) I think Dr. Hyatt will agree with me that the SSDF does good work. I think Dr. Hyatt will disagree with the validity of comparing SSDF with FIDE. Dr. Hyatt makes good points based on facts, even if I disagree with the interpretation. I would like to see about 2092 (given 3SEM, 99.7 confidence level on SSDF with 69727 games) GM games against programs that clearly show that programs are 200 point lower (median) and the spread should be reduced by 100 points. There is no data (0 games) to support the SSDF data is bad argument (at least I have not seen it) IMHO, however, it is just as valid to ask for more proof of human vs program games that support the SSDF. I can understand both sides, but make up my own mind. I think the view of IM Kaufman, Bertil (SSDF), myself, and others is also a valid view, but more data would improve the validity. :) Best Regards, Chris Carson
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.