Chess Cloud Database Query Interface

[Master Om]

In that url by noobftw in this position after
1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6
[d]rnbqkb1r/1p2pppp/p2p1n2/8/3NP3/2N5/PPP2PPP/R1BQKB1R w KQkq - 0 6

It suggests 6.Be2 surprisingly.
After 6.Be2 e5 7.Nb3 Be7 it suggests 8.g4??
[d]rnbqk2r/1p2bppp/p2p1n2/4p3/4P1P1/1NN5/PPP1BP1P/R1BQK2R b KQkq - 0 8

Now after that the best moves it suggests are 8. g4 h6 9. Be3 Be6 10. h4 Nbd7 11. Rg1 Nh7

[d]r2qk2r/1p1nbppn/p2pb2p/4p3/4P1PP/1NN1B3/PPP1BP2/R2QK1R1 w Qkq - 0 12

Now after That it again suggests Rh1. Now after Nf6 again its Rg1 and so on. Whats the point of suggesting a draw line as best ?
[pgn][Event "4670"]
[Site "AICCF"]
[Date "????.??.??"]
[Round "?"]
[White "Om Prakash"]
[Black "Sethuraman"]
[Result "*"]
[BlackElo "1642"]
[Classes "?"]
[ECO "B92"]
[TimeControl "?"]
[WhiteElo "2106"]

{1|46700403} 1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Be2 e5 7.
Nb3 Be7 8. g4 h6 9. Be3 Be6 10. h4 Nbd7 11. Rg1 Nh7 12. Rh1 Nhf6 13. Rg1 *[/pgn]

[jdart]

Sometimes a draw line is best, according to the scores, in other words, other moves appear to have negative scores.

One of the problems in actual game play though is that the value of a draw should really depend on the contempt setting.

[Master Om]

Sometimes a draw line is best, according to the scores, in other words, other moves appear to have negative scores.

One of the problems in actual game play though is that the value of a draw should really depend on the contempt setting.

There are better moves than Be2. Like Be3 and h3 perhaps.

[Ovyron]

Now after That it again suggests Rh1. Now after Nf6 again its Rg1 and so on. Whats the point of suggesting a draw line as best ?

It's a problem with the way CDBCN fudges scores. If the leaf position is 0.00, the 0.00 will not be propagated to a previous position, because the thing will guess that the defending side has some chance to not play the best line, so it'd give a different score depending on how likely it thinks it'll be played (the root scores that it shows do NOT assume best play.)

When this was discussed before being implemented, I suggested two ideas, one was to offset all the scores by 0.10 if white has the advantage and -0.10 if black does, then the whole -0.10 to 0.10 is left for this score fudging of 0.00 positions. The other idea would be to go subscoring like -0.001 to 0.001, and use this extra digit for fudging between scores. But since CDBCN already saves space by using 1 instead of 0.01 there's no way this is happening.

But your example might show how the fudging is too strong (it shouldn't suggest a move that leads to 0 if the other 2 lead to an advantage.)

[Dicaste]

Code: Select all

rnbqkb1r/1p2pppp/p2p1n2/8/3NP3/2N5/PPP2PPP/R1BQKB1R w KQkq - 0 1

Analysis by Lc0 LS:
1. +/- (0.72): 6.h3 e5 7.Nde2 Be6 8.g4 h6 9.Bg2 Nbd7 10.a4 Rc8 11.0-0 Be7 12.Ng3 g6 13.Nd5 Bxd5 14.exd5 Kf8 
2. +/= (0.68): 6.f3 e5 7.Nb3 Be6 8.Be3 Be7 9.Qd2 h5 10.Nd5 Nxd5 11.exd5 Bf5 12.Na5 b6 13.Nc4 Nd7 14.a4 0-0 15.Be2 Bh4+ 16.g3 Be7 17.0-0 Qc7 18.a5 b5 19.Nb6 Rab8 20.c4 bxc4 21.Nxc4 Rb3 22.Rfc1 
3. +/= (0.67): 6.Be3 e5 7.Nb3 Be7 8.f3 Be6 9.Qd2 h5 10.Nd5 Nxd5 11.exd5 Bf5 12.Na5 b6 13.Nc4 Nd7 14.a4 0-0 15.Be2 Bh4+ 16.g3 Be7 17.0-0 Qc7 18.a5 b5 19.Nb6 Rab8 20.c4 bxc4 21.Nxc4 Rb3 22.Rfc1 Qb7 23.Ra3 Rxa3 
White is better

[noobpwnftw]

The back propagation of scores consider many factors: it does not prefer draws, a small penalty is given for non-progression, the number of counter-moves and their score contribution to the final weighted average score. White and black are not strictly treated equally to prevent those effects to cascade. It will very likely to produce slightly different move evaluations than any engine would do, since some "tweaks" to the score are invented by myself while some of those calculations are simply too expensive to perform for an engine.

Playing for some borderline advantage that may not convert and may backfire is just as "bad" as playing safe, however once a line is learned/discovered by CDB it will adjust scores accordingly.

[Ovyron]

White and black are not strictly treated equally to prevent those effects to cascade.

Now I'm strongly against this. If there's some English variation that reaches a position exactly like a Reversed Sicilian (colors are swapped), then not only the analysis of both should be exactly the same, you can just produce a single analysis and link it so it's valid for both colors playing the positions (so any node that you analyze on the Sicilian makes it to that variation of the English with colors reversed, and vice versa.)

The Houdini chess engine does something like this, but instead of using colors it checks who has the advantage, so white and black are not strictly treated equally, but in positions where black has the advantage, white is treated like black and black is treated like white. Perhaps this would be an improvement.

[noobpwnftw]

Flipping colors of the board is not exactly the same as "reaching certain position with which side to move", they are different concepts.
Positive scores and negative scores have different weighted averaging windows, it may be similar to what some engines did. Internally, CDB counts the shortest line to reach a position from start position, this is a property of the position itself, in order to measure "progression" of certain lines and "degrade" its scores on their parent, which may solve some fortress issues.

So if you reach a certain position from different lines, CDB will provide the exactly same results, on the other hand, if CDB discovers a shorter line that leads to the same position, then this property is updated and all subsequent scores will be recalculated.

[Ovyron]

I see, that's very cool and I think ChessDBCN is the only database with those features (I should have checked, 1.e4 e5 and 1.e3 e5 2.e4 show the same analysis flipped and are scored the same so white and black are treated equally in such cases). I still don't like those highly scored variations that lead to a 0 position if you play the best moves, but I can't think of a way to avoid that.

[Master Om]

The back propagation of scores consider many factors: it does not prefer draws, a small penalty is given for non-progression, the number of counter-moves and their score contribution to the final weighted average score. White and black are not strictly treated equally to prevent those effects to cascade. It will very likely to produce slightly different move evaluations than any engine would do, since some "tweaks" to the score are invented by myself while some of those calculations are simply too expensive to perform for an engine.

Playing for some borderline advantage that may not convert and may backfire is just as "bad" as playing safe, however once a line is learned/discovered by CDB it will adjust scores accordingly.

I still dont understand how a move scores at top if black can make draw by repeating the move order again and again ?