Trusting Databases

[Richard Allbert]

Hi,

Just a silly point. I finshed writing the bare bones of a class to analyse databases of games, to use the statistics to help determine evaluation weights.

I've been debugging using a "2600.pgn" I found (I think on Dann's site) of 14000+games.

First test for statistics was one side minor piece up (conditions are less than 7 pawns per side and 5 plys since capture). I was shocked to see barely over 50% score for the side a piece up.... so I ended up printing the valid games to have a look.

An example...

Code: Select all

[Event "EU-ch blitz"]
[Site "Panormo"]
[Date "2002.10.02"]
[Round "3"]
[White "Volkov, Sergey"]
[Black "Ehlvest, Jaan"]
[Result "1-0"]
[ECO "E20"]
[WhiteElo "2642"]
[BlackElo "2600"]
[PlyCount "47"]
[EventDate "2002.10.02"]
[EventType "tourn (blitz)"]
[EventRounds "39"]
[EventCountry "GRE"]
[Source "ChessBase"]
[SourceDate "2002.11.06"]

1.d4 Nf6 2.c4 e6 3.Nc3 Bb4 4.f3 d5
5.a3 Bxa3 6.bxa3 c5 7.cxd5 Nxd5 8.dxc5 Qa5
9.e4 Nf6 10.Be3 O-O 11.Kf2 Nfd7 12.Rb1 b6
13.cxb6 axb6 14.Bb5 Nc5 15.Nge2 Ba6 16.a4 Bxb5
17.axb5 Nbd7 18.Qd6 Qa2 19.Ra1 Qc2 20.Rxa8 Rxa8
21.Bxc5 Nxc5 22.Ra1 Rc8 23.Qc6 Rf8 24.Ra8 1-0

A blitz game with an unusual variation from move 5. as black !!



Ah well, back to the drawing board...

Btw, I've been scoring each game as follows....

Each game that has a position meeting the defined criteria (e.g up a minor piece) has 100 added to the running total.

If the side up a minor piece wins, another score gets 100. If draw gets 50. If loss zero.

Then I compare the scores together.

Is this flawed? Most likely...

Regards

Richard

[Dr.Ex]

You should use a big computer database to determine evaluation weights.
The human databases, even the commercial ones, contain lots of silly moves that actually never happened.
The main reasons are unreadable Score sheets.

[Richard Allbert]

Yes, I will.... like I said above, I was using the database for debugging.

It's faster than using 2million plus game databases for debugging!

[Mike S.]

This is a strange transmission error or input error. 5...Bxa3 6.bxa3 has to be replaced with the very obvious 5...Bxc3 6.bxc3 as it's common in the French. Then we have a known variation up to 12.Rb1 and beyond.

[D]rnbqk2r/ppp2ppp/4pn2/3p4/1bPP4/P1N2P2/1P2P1PP/R1BQKBNR b KQkq - 0 5

[Event "EU-ch blitz"]
[Site "Panormo"]
[Date "2002.10.02"]
[Round "3"]
[White "Volkov, Sergey"]
[Black "Ehlvest, Jaan"]
[Result "1-0"]
[ECO "E20"]
[WhiteElo "2642"]
[BlackElo "2600"]
[PlyCount "47"]
[EventDate "2002.??.??"]

1. d4 Nf6 2. c4 e6 3. Nc3 Bb4 4. f3 d5 5. a3 Bxc3+ 6. bxc3 c5 7. cxd5 Nxd5 8.
dxc5 Qa5 9. e4 Nf6 10. Be3 O-O 11. Kf2 Nfd7 12. Rb1 b6 13. cxb6 axb6 14. Bb5
Nc5 15. Ne2 Ba6 16. a4 Bxb5 17. axb5 Nbd7 18. Qd6 Qa2 19. Ra1 Qc2 20. Rxa8 Rxa8
21. Bxc5 Nxc5 22. Ra1 Rc8 23. Qc6 Rf8 24. Ra8 1-0

(I guess Black lost on time?)

[Richard Allbert]

Ok, thanks for the input!