This is a companion discussion topic for the original entry at https://dphi.tech/challenges/138
The Target value “depth-of-win” is inaccurate. Provided the constraint “min-max optima” either the “depth-of-win” provided in the dataset is high (because there are much more optimal moves and hence game could have been ended sooner) or “depth-of-win” is low (because there are more optimal moves for the black king and hence game could have been laster longer). I run an analysis using one of the top chess engines “stockfish” and most of their values don’t align perfectly. Nevertheless, the dataset is a part of a human playing against a human so there is a chance of making a mistake, or quitting too early when you can anticipate the outcome making “depth-of-win” shorter (I suppose).
Here is the screengrab of the analysis result:
* Top cell shows the advantage score (depth of win) by stockfish (white perspective).
Here is the link to my analysis if anybody is interested: Data Sprint #39:Chess Endgame (ANALYSIS)
Does this analysis start from black side?
From some low depth data, I think first move is black side.
If the black king can not move, the depth is 0.
black move, white move then checkmate → 1.
black move, white move, black move, white move then checkmate → 2.
1 Like
Hi @uwi , By default white moves first (in the python “chess” module). I wish they have (competition organizers) mentioned the board state is “Black-to-move” (I should have read the original dataset source title carefully). I made black to move first and values are much more closer to the dataset values but are still not completely in sync.
I wonder if adjusting few parameter values could work (ie, depth, time limit, etc).
Even the original source of the dataset says "It should be noted that our database is not guaranteed correct, but the class distribution is the same as Clarke’s database. "
Thank you,
Stay safe.