Stockfish or Leela Chess Zero? A Comparison Against Endgame Tablebases

The game of chess has long been used as a benchmark for testing human creativity and intelligence. With the advent of powerful chess engines, such as Stockfish and Leela Chess Zero (Lc0), endgame studies have also become a tool for evaluating the capabilities of machine chess engines. In this work, we conduct a detailed study of Stockfish and Lc0, two leading chess engines with distinct methods of play, using chess endgames with varying numbers of remaining pieces. We evaluate the programs’ move decision errors when using only the raw policy network as well as when using a small amount of search. We provide insights into the strengths and weaknesses of Stockfish and Lc0 in handling complex endgame positions by exploring common mistakes and identifying interesting behaviours of the engines based on the position of the opponent’s last pawn remaining on the board.