Questioning the analysis

wollyhood wrote:

It ( analysis / report) says that @ 20...bxe3 I am up as black -1.38, then they make the best move and take me, and then I make the best move at 21...f5 and I am still winning but only at -1.08

Yes, that's the horizon effect described above. The analysis only goes a certain number of half-moves deep and then the position is evaluated. So when you make a move on the board, that move is cleared from the analysis queue, the analysis extends one move deeper, and a new position is evaluated (often, changing the value).


Hm. OK. It still seems incorrect if all best moves lead to a draw.


ok, i see my error - i had the board reversed with respect to the black and white sides of the board.

Yes, it's a draw according to the endgame tablebase.


i remember reading a article that explained that even high-powered chess engines like stockfish will change their evaluation from move to move and sometimes the change in evaluation can be rather severe.  Apparently it is in the nature of the game.  Not sure about AlphaZero or Leela Chess Zero.


I guess my point for this thread is that the analysis is wrong :\ seems a little concerning, but at least PCs don't have chess completely sussed yet : )


It is important to understand the limitations of computer analysis, such as the "horizon effect" that blueemu has mentioned, where the computer can't analyze one move further without devoting too many resources to the analysis.


There are also many positions where a computer is simply not good at solving them.  These positions tend to be tactically quiet and generally subtle positions - the kind of positions where the best move may be Kb2, but the strategic difference between playing Kb2 and playing Kb1 might not be evident for 10-15+ moves.


Related to this, another issue with computer analysis is that there are some positions where a human can immediately say "this is dead drawn," but that a computer might not evaluate as well.  Take, for example, a position with the following - white pawns on a3, c3, e3, and g3 and black pawns on a4, c4, e4, and g4; white king on a2 and black king on g7; black bishop on b5.  There is no way for either side to advance, but a computer may give a major edge to black due to the extra space and extra piece.  A human can scan the board and quickly see that there is no way for a king or the bishop to take any of the pawns or otherwise get past them.  A computer will have to exhaustively search every possible move/move order to make the same determination, since there are no 11-man tablebases.  Since that kind of search would take way too long, it will eventually stop looking. and it'll see the so-called plusses in the black position that I mentioned and give them the edge, when in reality, that edge is purely imaginary.





Very good, thanks for the explanations and visual.