Home 
Search 
Today's Posts 
#1




Statistical significance of score differences  new release of ChessDB
I've made a new release of ChessDB, a chess database based on Scid from
Shane Hudson. There is a fork too of ChessDB from the lying plagiarist Pascal Georges who passes of work of mine as his own, as I have documented at: http://groups.google.co.uk/group/rec...b9e3c5e4e7266a Anyway, the main reason for my post it to introduce a new feature in ChessDB and I would be interested in comments from others about this. Basically ChessDB has a tree window, like many databases (Scid, ChessBase, Chess Assistant etc). But I've added code that will determine if the difference in score between two moves is really real ('statistically significant') , or if it could be due to chance. (If you toss a coin 20 times and it lands on heads 12 times and tails 8 times, you can't deduce the coin is biased)  such a small difference can be due to chance with only 20 tosses. In contrast, if it landed on the head 19 times and the tail only once, you be pretty sure it is biased. It is assumed that the difference in score between two moves is not due to chance if the probability of the observed (or any larger) difference being due to chance, with no underlying reason, is less than 0.05. See: http://chessdb.sourceforge.net/tutor...earch_tree.php Some interesting observations can be seen looking at my database of 3.5 million games 1) 1.d4 scores better than 1.e4 with a pvalue of less than 0.01. In other words, the chance of the observed or any large score difference being due to chance is less than 1%. 2) In my database, the opening move with the highest score is 1.Na3. Despite the score being a lot higher than 1.e4 or 1.d4, this is *not* statistically significant. In other words, whilst we can't say it 1.Na3 is any better or worst than 1.e4 or 1.d4, we can say that that there is a high probability that the observed difference is due to chance. As such, we should pay very little attention to the relative scores. 3) In my database, 3.Nd2 (Tarrash variation) in the French (1.e4 e6 2.d4 d5) scores higher than 3.Nc3 (main line) and is statistically significant at the 5% level, but not at the 1% level. In other words, we can be 95% sure there is a real difference in score between 3.Nf3 and 3.Nd2 in my database, but we can't be 99% sure. In contrast, the difference in scores of 3.Nf3 (or 3.Nd2) to the exchange (3.exd5) or advance (3.e5) variations is statistically significant at the p=0.01 level, so there is less than a 1% chance the observed difference in score is due to chance and can be more than 99% sure there is an underlying reason. (The reason can't be determined in ChessDB, but one might strongly suspect the advance or exchange are inferior for white than the main line (3.Nc3) or Tarrasch (3.Nd2). (I personally have a much better success rate with the Tarrasch than the advance too. I will not contemplate the exchange as it is too boring and while it is drawish, it scores pretty low for white.) Anyone with a reasonable knowledge of statistics might guess I am using a chisquared test, which is what I am doing. Chisquared is calculated then the pvalue determined from that, using an algorithm good to 4 decimal places. I intend changing that to a more accurate approximation soon. Other changes in ChessDB include * Native support for UCI engines (using some code from P. Georges, which I fully acknowledge, unlike him when he uses my code). * The facility to download a database of either 100,000 or 3.5 million games. The database is split into multiple parts for easy downloading, then reconstructed by ChessDB and an MD5 checksum used to verify the database has not been corrupted in transmission. * Quickly download games from the history of anyone on ICC or FICS. * Numerous other changes documented at: http://chessdb.sourceforge.net/Scid/ I'm interested in what others think of the idea of testing the statistical significance in the difference of two moves. To the best of my knowledge, no other chess database does this, yet it seems to me quite logical. 
#2




Statistical significance of score differences  new release ofChessDB
Dave wrote:
I've made a new release of ChessDB, a chess database based on Scid from Shane Hudson. I forgot to say, if you want to try ChessDB or use it to download a large database, see: http://chessdb.sourceforge.net/downloads/ 
#3




Statistical significance of score differences  new release of ChessDB
That is a very cool idea for a feature. I haven't used databases much
in general, so I don't know much about existing features, but I could see how something like that could be very useful to a master trying to refine their opening preparation. At my patzer level (1400ish USCF), it's something that could be cool just for the sake of curiousity. So is this a free program? As I said, I don't know much about chess databases (which is ironic, since I specialize in SQL databases in my profession). I'd like to get a database program and a large database of master and GM games eventually, so I can see how better players than me handle certain openings and the positions that result from them. For now, I tend to just go to chesslab.com and look at games there in the openings I play. Fromper 
#4




Statistical significance of score differences  new release ofChessDB
Well, if you like chess databases with a lot of features you can grab
Scid at http://scid.sourceforge.net or http://prolinux.free.fr/scid (the lattest with some training features, and the ability to play against various engines). Richard a écrit : That is a very cool idea for a feature. I haven't used databases much in general, so I don't know much about existing features, but I could see how something like that could be very useful to a master trying to refine their opening preparation. At my patzer level (1400ish USCF), it's something that could be cool just for the sake of curiousity. So is this a free program? As I said, I don't know much about chess databases (which is ironic, since I specialize in SQL databases in my profession). I'd like to get a database program and a large database of master and GM games eventually, so I can see how better players than me handle certain openings and the positions that result from them. For now, I tend to just go to chesslab.com and look at games there in the openings I play. Fromper 
#5




Statistical significance of score differences  new release ofChessDB
At least other people around the world clearly got who is that guy !!
Pascal http://prolinux.free.fr/alex_guestbook/ Dave a écrit : I've made a new release of ChessDB, a chess database based on Scid from Shane Hudson. There is a fork too of ChessDB from the lying plagiarist Pascal Georges who passes of work of mine as his own, as I have documented at: http://groups.google.co.uk/group/rec...b9e3c5e4e7266a Anyway, the main reason for my post it to introduce a new feature in ChessDB and I would be interested in comments from others about this. Basically ChessDB has a tree window, like many databases (Scid, ChessBase, Chess Assistant etc). But I've added code that will determine if the difference in score between two moves is really real ('statistically significant') , or if it could be due to chance. (If you toss a coin 20 times and it lands on heads 12 times and tails 8 times, you can't deduce the coin is biased)  such a small difference can be due to chance with only 20 tosses. In contrast, if it landed on the head 19 times and the tail only once, you be pretty sure it is biased. It is assumed that the difference in score between two moves is not due to chance if the probability of the observed (or any larger) difference being due to chance, with no underlying reason, is less than 0.05. See: http://chessdb.sourceforge.net/tutor...earch_tree.php Some interesting observations can be seen looking at my database of 3.5 million games 1) 1.d4 scores better than 1.e4 with a pvalue of less than 0.01. In other words, the chance of the observed or any large score difference being due to chance is less than 1%. 2) In my database, the opening move with the highest score is 1.Na3. Despite the score being a lot higher than 1.e4 or 1.d4, this is *not* statistically significant. In other words, whilst we can't say it 1.Na3 is any better or worst than 1.e4 or 1.d4, we can say that that there is a high probability that the observed difference is due to chance. As such, we should pay very little attention to the relative scores. 3) In my database, 3.Nd2 (Tarrash variation) in the French (1.e4 e6 2.d4 d5) scores higher than 3.Nc3 (main line) and is statistically significant at the 5% level, but not at the 1% level. In other words, we can be 95% sure there is a real difference in score between 3.Nf3 and 3.Nd2 in my database, but we can't be 99% sure. In contrast, the difference in scores of 3.Nf3 (or 3.Nd2) to the exchange (3.exd5) or advance (3.e5) variations is statistically significant at the p=0.01 level, so there is less than a 1% chance the observed difference in score is due to chance and can be more than 99% sure there is an underlying reason. (The reason can't be determined in ChessDB, but one might strongly suspect the advance or exchange are inferior for white than the main line (3.Nc3) or Tarrasch (3.Nd2). (I personally have a much better success rate with the Tarrasch than the advance too. I will not contemplate the exchange as it is too boring and while it is drawish, it scores pretty low for white.) Anyone with a reasonable knowledge of statistics might guess I am using a chisquared test, which is what I am doing. Chisquared is calculated then the pvalue determined from that, using an algorithm good to 4 decimal places. I intend changing that to a more accurate approximation soon. Other changes in ChessDB include * Native support for UCI engines (using some code from P. Georges, which I fully acknowledge, unlike him when he uses my code). * The facility to download a database of either 100,000 or 3.5 million games. The database is split into multiple parts for easy downloading, then reconstructed by ChessDB and an MD5 checksum used to verify the database has not been corrupted in transmission. * Quickly download games from the history of anyone on ICC or FICS. * Numerous other changes documented at: http://chessdb.sourceforge.net/Scid/ I'm interested in what others think of the idea of testing the statistical significance in the difference of two moves. To the best of my knowledge, no other chess database does this, yet it seems to me quite logical. 
#6




Statistical significance of score differences  new release of ChessDB
Den 20070817 14:52:16 skrev Dave :
I've made a new release of ChessDB, a chess database based on Scid from Shane Hudson. There is a fork too of ChessDB from the lying plagiarist Pascal Georges who passes of work of mine as his own, as I have documented at:.......... Befrore I download it I would like to know whether it handles transpositions, i.e., are *unplayed* moves visible in the tree that lead to a played position? Mats 
#7




Statistical significance of score differences  new release ofChessDB
17.08.2007 14:52, Dave:
But I've added code that will determine if the difference in score between two moves is really real ('statistically significant') , or if it could be due to chance. Two important factors are completely ignored in this calculation: development over time and strength of the involved players. The first is important because once a refuation or at least a very strong answer for a move is found, its frequency drops. So, the old statistics of this move stay unchanged over a long time  possibly with a favourable result for this move, although it might be well known that this move should be avoided. The second factor is quite obvious: games of higher rated players tend to be less erratic, so that those results are more expressive. Therefore, if I look at numbers, I check the average Elo and the performance. Greetings, Ralf 
#8




Statistical significance of score differences  new release ofChessDB
Anonymous wrote:
Well, if you like chess databases with a lot of features you can grab Scid at http://scid.sourceforge.net or http://prolinux.free.fr/scid (the lattest with some training features, and the ability to play against various engines). And the latter of which has code taken from ChessDB but not acknowledged. 
#9




Statistical significance of score differences  new release ofChessDB
Richard wrote:
That is a very cool idea for a feature. I'm glad you like it. I haven't used databases much in general, so I don't know much about existing features, but I could see how something like that could be very useful to a master trying to refine their opening preparation. At my patzer level (1400ish USCF), it's something that could be cool just for the sake of curiousity. So is this a free program? Yes, its open source and free. homepage http://chessdb.sourceforge.net/ tutorial http://chessdb.sourceforge.net/tutorial/ download page: http://chessdb.sourceforge.net/downloads/ As I said, I don't know much about chess databases (which is ironic, since I specialize in SQL databases in my profession). I'd like to get a database program and a large database of master and GM games eventually, so I can see how better players than me handle certain openings and the positions that result from them. If you download it, then go to the Tools menu, Select "Download games from" then select "3.5 million games site #1" it will download you a 3.5 million game database. For now, I tend to just go to chesslab.com and look at games there in the openings I play. Well with 3.5 million you have quite a few. ChessDB also has the facility to download from The Week In Chess (TWIC), so you can update the database every week (usually a Monday), when new games are added to TWIC. See: http://chessdb.sourceforge.net/tutor...retriveal.php (The program has an http client to connect to the external sources of data. There is also a telnet client which is used to download games from FICS and ICC). I do have a larger database, which I could make available, but as databases get larger, the quality of the games goes down. 
#10




Statistical significance of score differences  new release ofChessDB
M Winther wrote:
Befrore I download it I would like to know whether it handles transpositions, i.e., are *unplayed* moves visible in the tree that lead to a played position? Mats Yes. It shown positions, not moves. Hence sometimes you will find there are no games in the database at move 5, but by move 6 there are thousands. 