Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / To err is human, to really foul things up requires a NN
- - By user923005 (****) Date 2020-01-24 03:12 Upvotes 2
I have a collection of 1081 positions I have analyzed at fairly long time control both with Stockfish and with LC0.
Due (most likely) to the long time control, only 96 times out of 1081 positions do the engines prefer a different best move.

Even when they prefer a different best move, they almost always agree fairly closely on the score for the position (to within ten centipawns or less for all except 38 positions out of 1081).

The interesting thing is that when they differ, the NN engine usually has the better answer.  But when they differ by a LOT it is the alpha-beta engine that has it right.
Attachment: nn-ab.epd - The EPD positions from the SQL query (2k)
Parent - By dorszcz (**) Date 2020-01-24 07:42
Stockfish and Lc0 have similar strenght. When Stockfish is better in tactics, Lc0 must get points in quieter positions.
Stockfish usually has much higher eval than any other alpha-beta engine when it sees a clear win.
Parent - By Venator (Silver) Date 2020-01-25 09:48
LC0 overlooks tactics now and then.
Stockfish overlooks good positional ideas now and then.

Both have (different) blind spots. Both are still far from perfect. But in strength it seems they are very, very close.
Parent - - By Uly (Gold) Date 2020-01-26 21:51
Do you have a position that Stockfish says it's =>2.00 but Leela doesn't see the side is winning and it turns out Leela is right? I have a theory that at some point in score Stockfish will always be right, so you could just ignore Leela entirely on those variations. I have no idea what this score would be, but if such a 2.00 score non-win position existed it'd prove my premise wrong (ignoring Leela would come much later.)
Parent - - By user923005 (****) Date 2020-01-27 19:27 Upvotes 1
I guess that Fortress positions might be an example, but I did not check for that.  Stockfish often needs 100 plies to see the fortress.
Parent - By Uly (Gold) Date 2020-01-28 21:30
Okay, but usually in those the score remains the same iteration after iteration, so they'd be easy to spot. Anyway, I realized yesterday that I've been using a CPU Leela that abnormally runs some 250 times slower than it's supposed to, so I probably can't conjecture anything about it (as it'd be as if I was talking about depth 6 Stockfish or something.)
- - By rov (*) Date 2020-01-25 14:18
@user923005

Is it better to cope with epd positions and generally with chess databases by using sql studio ? Are Your data in txt files or in sql bases ?
Parent - By user923005 (****) Date 2020-01-26 05:25
It is better for me because I use SQL a lot.  I can more or less think in SQL, so it is a natural fit for me to search for ideas with it
Up Topic The Rybka Lounge / Computer Chess / To err is human, to really foul things up requires a NN

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill