In a somewhat desperate attempt to improve my play I have analysed some of my games with Crafty 22.8 and have some success gettting a clue on what the *beeep* is going on when im playing
But in this position the score really bumps around - i think its getting stable at ply 19
[d] r2qkb1r/3bpp1p/3P1P2/1p4Pp/np1P3P/p3BN2/P1PQ4/2KR1B1R b kq - 0 21
My question is this : isnt it more usual for programs to have a more consistent score than this(look below for a lengthy analysis from Crafty) ?
Other engines doesnt seem to fluctuate quite as much as Crafty ...
What could be the reason for this ?
This is about evaluation and search differences. Some programs will like one side or the other, depending on their evaluation of the black king in the center. In positions like this, many programs will jump around in the scores as the search uncovers things that change the circumstances significantly, and the evaluation then changes in a sometimes big way.
A suggested strategy for analysis of games:
Have the computer analyze when you are sleeping. Typically, a computer is unused for at least ten hours per day. So take a game like this and have the computer analyze for 10 hrs * 3600 sec/hour = 36000 seconds total, divided by 40 {in this case} distinct white positions:
rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - bm d4;
rnbqkbnr/pppppp1p/6p1/8/3P4/8/PPP1PPPP/RNBQKBNR w KQkq - bm e4;
rnbqk1nr/ppppppbp/6p1/8/3PP3/8/PPP2PPP/RNBQKBNR w KQkq - bm Nc3;
rnbqk1nr/ppp1ppbp/3p2p1/8/3PP3/2N5/PPP2PPP/R1BQKBNR w KQkq - bm Be3;
r1bqk1nr/pppnppbp/3p2p1/8/3PP3/2N1B3/PPP2PPP/R2QKBNR w KQkq - bm Qd2;
r1bqk2r/pppnppbp/3p1np1/8/3PP3/2N1B3/PPPQ1PPP/R3KBNR w KQkq - bm f3;
r1bqk2r/1ppnppbp/p2p1np1/8/3PP3/2N1BP2/PPPQ2PP/R3KBNR w KQkq - bm O-O-O;
r1bqk2r/2pnppbp/p2p1np1/1p6/3PP3/2N1BP2/PPPQ2PP/2KR1BNR w kq - bm h4;
r1bqk2r/2pnppbp/p2p2p1/1p5n/3PP2P/2N1BP2/PPPQ2P1/2KR1BNR w kq - bm Nce2;
r1bqk2r/2p1ppbp/pn1p2p1/1p5n/3PP2P/4BP2/PPPQN1P1/2KR1BNR w kq - bm b3;
r1bqk2r/2p1ppbp/1n1p2p1/pp5n/3PP2P/1P2BP2/P1PQN1P1/2KR1BNR w kq - bm g4;
r1bqk2r/2p1ppbp/1n1p1np1/pp6/3PP1PP/1P2BP2/P1PQN3/2KR1BNR w kq - bm Ng3;
r2qk2r/2pbppbp/1n1p1np1/pp6/3PP1PP/1P2BPN1/P1PQ4/2KR1BNR w kq - bm g5;
r2qk2r/2pbppbp/1n1p2p1/pp4Pn/3PP2P/1P2BPN1/P1PQ4/2KR1BNR w kq - bm Nxh5;
r2qk2r/2pbppbp/1n1p4/pp4Pp/3PP2P/1P2BP2/P1PQ4/2KR1BNR w kq - bm f4;
r2qk2r/2pbppbp/1n1p4/1p4Pp/p2PPP1P/1P2B3/P1PQ4/2KR1BNR w kq - bm b4;
r2qk2r/2pbppbp/1n1p4/1p4Pp/1P1PPP1P/p3B3/P1PQ4/2KR1BNR w kq - bm f5;
r2qk2r/2pbppbp/3p4/1p3PPp/nP1PP2P/p3B3/P1PQ4/2KR1BNR w kq - bm Nf3;
r2qk2r/3bppbp/3p4/1pp2PPp/nP1PP2P/p3BN2/P1PQ4/2KR1B1R w kq - bm e5;
r2qk2r/3bppbp/3p4/1p2PPPp/np1P3P/p3BN2/P1PQ4/2KR1B1R w kq - bm f6;
r2qkb1r/3bpp1p/3p1P2/1p2P1Pp/np1P3P/p3BN2/P1PQ4/2KR1B1R w kq - bm exd6;
r2qkb1r/3b1p1p/3P1p2/1p4Pp/np1P3P/p3BN2/P1PQ4/2KR1B1R w kq - bm Bf4;
r3kb1r/3b1p1p/3P1p2/qp4Pp/np1P1B1P/p4N2/P1PQ4/2KR1B1R w kq - bm Bh3;
r3kb1r/3b1p1p/3P1p2/qp4Pp/1p1P1B1P/p1n2N1B/P1PQ4/2KR3R w kq - bm Bxd7+;
r4b1r/3k1p1p/3P1p2/qp4Pp/1p1P1B1P/p1n2N2/P1PQ4/2KR3R w - - bm gxf6;
r4b1r/3k1p1p/3P1P2/qp5p/1p1P1B1P/p4N2/n1PQ4/2KR3R w - - bm Kb1;
r4b1r/3k1p1p/3P1P2/qp5p/1p1P1B1P/p1n2N2/2PQ4/1K1R3R w - - bm Ka1;
r6r/3k1p1p/3b1P2/qp5p/1p1P1B1P/p1n2N2/2PQ4/K2R3R w - - bm Bxd6;
r6r/5p1p/3k1P2/qp5p/1p1P3P/p1n2N2/2PQ4/K2R3R w - - bm Qf4+;
r6r/3k1p1p/5P2/qp5p/1p1P1Q1P/p1n2N2/2P5/K2R3R w - - bm Qf5+;
r2k3r/5p1p/5P2/qp3Q1p/1p1P3P/p1n2N2/2P5/K2R3R w - - bm Rde1;
3k3r/r4p1p/5P2/qp3Q1p/1p1P3P/p1n2N2/2P5/K3R2R w - - bm Re7;
3k3r/4rp1p/5P2/qp3Q1p/1p1P3P/p1n2N2/2P5/K6R w - - bm fxe7+;
7r/4kp1p/8/qp3Q1p/1p1P3P/p1n2N2/2P5/K6R w - - bm Qe5+;
7r/3k1p1p/8/qp2Q2p/1p1P3P/p1n2N2/2P5/K6R w - - bm Qxh8;
7Q/3k1p1p/8/qp5p/3P3P/ppn2N2/2P5/K6R w - - bm Ne5+;
7Q/2k2p1p/8/qp2N2p/3P3P/ppn5/2P5/K6R w - - bm cxb3;
7Q/2k2p1p/1q6/1p2N2p/3P3P/pPn5/8/K6R w - - bm Nxf7;
7Q/2k2N1p/6q1/1p5p/3P3P/pPn5/8/K6R w - - bm Qd8+;
3Q4/5N1p/2k3q1/1p5p/3P3P/pPn5/8/K6R w - - bm Ne5+;
giving 900 seconds per white move (in this case).
This should yield 3-4 swing points where a particularly interesting move was made (either in a positive or negative sense). Take these (perhaps 4) positions and analyze them the next night at 2 hours/position. I find it is helpful to to examine not only the last predicted variation (which probably contains the ideal trajectory of play) but also the intervening ones where the key move has changed. This is because I often like one of the intervening moves and do not fully understand the final best path. The last pv for the intervening move will for one of these places where the predicted move has changes will often show the refutation. Finally, take the best choice pv and play through it using some visual tool like Winboard. This occasionally results in one of those "Aha!" moments where you can see the reason for the plan.
Dann Corbit wrote:
This should yield 3-4 swing points where a particularly interesting move was made (either in a positive or negative sense). Take these (perhaps 4) positions and analyze them the next night at 2 hours/position. I find it is helpful to to examine not only the last predicted variation (which probably contains the ideal trajectory of play) but also the intervening ones where the key move has changed. This is because I often like one of the intervening moves and do not fully understand the final best path. The last pv for the intervening move will for one of these places where the predicted move has changes will often show the refutation. Finally, take the best choice pv and play through it using some visual tool like Winboard. This occasionally results in one of those "Aha!" moments where you can see the reason for the plan.
Actually this sounds like a very good idea .... are there some GUI's that are easier to do this in than others ?
It seems like Rybka is really struggling with this also - but the position is both very unbalanced and complicated so maybe it throws the programs off ?
Im going to let Crafty look at this while im go into deep think mode(sleep) ill bump up the hash for crafty and let it run for 6-10 hours and get back with the analysis.
Dann Corbit wrote:
This should yield 3-4 swing points where a particularly interesting move was made (either in a positive or negative sense). Take these (perhaps 4) positions and analyze them the next night at 2 hours/position. I find it is helpful to to examine not only the last predicted variation (which probably contains the ideal trajectory of play) but also the intervening ones where the key move has changed. This is because I often like one of the intervening moves and do not fully understand the final best path. The last pv for the intervening move will for one of these places where the predicted move has changes will often show the refutation. Finally, take the best choice pv and play through it using some visual tool like Winboard. This occasionally results in one of those "Aha!" moments where you can see the reason for the plan.
Actually this sounds like a very good idea .... are there some GUI's that are easier to do this in than others ?
I use SCID right now for analysing.
Nothing wrong with SCID.
My choice would be the following path:
1. pgn2epd < krj.pgn > krj.epd {This step creates EPD records from the PGN}
2. grep " w " krj.epd > krjw.epd {This step extracts only the white records}
3. Count the records
4. Calculate how much time you have per record
5. Analyze using one of:
A. Arena GUI {Free}
B. Shredder GUI {$$$}
C. ChessAssistant GUI {$$$}
Some other GUIs that I like but are currently a little troublesome for EPD analysis are Aquarium and ChessGUI, and I have hopes for big things from these in the near future.
After the above steps, I would reanalyze the interesting records at much longer time control.
In a somewhat desperate attempt to improve my play I have analysed some of my games with Crafty 22.8 and have some success gettting a clue on what the *beeep* is going on when im playing
But in this position the score really bumps around - i think its getting stable at ply 19
[d] r2qkb1r/3bpp1p/3P1P2/1p4Pp/np1P3P/p3BN2/P1PQ4/2KR1B1R b kq - 0 21
My question is this : isnt it more usual for programs to have a more consistent score than this(look below for a lengthy analysis from Crafty) ?
Other engines doesnt seem to fluctuate quite as much as Crafty ...
Dann Corbit wrote:
This should yield 3-4 swing points where a particularly interesting move was made (either in a positive or negative sense). Take these (perhaps 4) positions and analyze them the next night at 2 hours/position. I find it is helpful to to examine not only the last predicted variation (which probably contains the ideal trajectory of play) but also the intervening ones where the key move has changed. This is because I often like one of the intervening moves and do not fully understand the final best path. The last pv for the intervening move will for one of these places where the predicted move has changes will often show the refutation. Finally, take the best choice pv and play through it using some visual tool like Winboard. This occasionally results in one of those "Aha!" moments where you can see the reason for the plan.
Actually this sounds like a very good idea .... are there some GUI's that are easier to do this in than others ?
I use SCID right now for analysing.
Nothing wrong with SCID.
My choice would be the following path:
1. pgn2epd < krj.pgn > krj.epd {This step creates EPD records from the PGN}
2. grep " w " krj.epd > krjw.epd {This step extracts only the white records}
2.5 edit krjw.epd and remove records from the top that are opening theory and records from the bottom (if any) from where the result is clear to the end of the game.
Dann Corbit wrote:3. Count the records
4. Calculate how much time you have per record
5. Analyze using one of:
A. Arena GUI {Free}
B. Shredder GUI {$$$}
C. ChessAssistant GUI {$$$}
Some other GUIs that I like but are currently a little troublesome for EPD analysis are Aquarium and ChessGUI, and I have hopes for big things from these in the near future.
After the above steps, I would reanalyze the interesting records at much longer time control.
kiroje wrote:It seems like Rybka is really struggling with this also - but the position is both very unbalanced and complicated so maybe it throws the programs off ?
Im going to let Crafty look at this while im go into deep think mode(sleep) ill bump up the hash for crafty and let it run for 6-10 hours and get back with the analysis.
There is a family of pathological positions that cause this kind of behavior. Basically, each time you go a ply deeper, you see more and different tactics. Which can affect the score at the root. Sometimes the positions are balanced on the edge so that a horizon-effect is used to hide something in this iteration, but the next goes one ply deeper and now you see something else. And you swap to a move that has a slight horizon-effect blindness itself until the next iteration exposes that. Flip-flopping back and forth is not that uncommon in wild tactical positions, and in certain types of endgames where a single tempi is critical.
