Leela is Bad at Endgames? Not So Fast.06 Jul 2018
[Here’s an analysis of Leela’s grasp of a fairly strategic endgame.
When the author mentions lc0, he’s actually talking about lczero v0.10.]
This position is an extremely deep and subtle ending that traditional engines have trouble with for several reasons. First, we can note that there are too many pieces on the board initially for current tablebases to provide a clear answer, though as we progress into the variations, tablebase hits will occur and provide us with clarity. The other reasons are all synergystic with one another: material stands at a significant imbalance in the traditional point count, it is not a highly tactical position with only one major piece for each side and no minor pieces, the pawn structure is relatively locked (white will not be keen to push the f-pawn), and in fact, white does in some variations threaten to set up a surprising, but known, theoretical fortress (for example, if both pairs of the g- and h-pawns are exchanged, and the f-pawn remains on f2, this is impenetrable).
We know that SF 8 and 9 are currently absolutely stronger than Lc0, though these traditional engines are relatively weak positionally (compared to their tactical strength) and Lc0 is known to be strong positionally (relative to its tactical strength). All of the factors in the present position make it a fertile ground for Lc0 to attempt to prove that it is not bad in “endgames” per se. Here we will, at various positions, present analysis from up to 4 different sources (though not all of them for all positions): 1) Dvoretsky’s Endgame Manual by Mark Dvoretsky, from which I was introduced to this position in my studies, and I reference it here, where it should be noted that Dvoretsky introduces this position by mentioning that “V. Khenkin supplied a detailed analysis for this ending.”; 2) Stockfish 8; 3) Stockfish 9 with 6-man Syzygy Tablebases, which I utilize to determine the only-move status of many of the presented moves; 4) Lc0 449, searching an appropriate fraction of the nodes of the Stockfish engines.