Understanding Chess Complexity and Feature Detection by Player Elo

Abstract

In the world of chess programming, not only are researchers and engineers interested in creating the strongest possible chess engine, but also in exploring the relations and parallels between human cognition and how neural networks learn. One significant area of research is the construction of 'human-like' engines, or chess engines that make moves much like those of a similarly-rated human. Similarities in how some engines 'think' about the game and how humans do may give insight into how both humans and algorithms alike may improve in 'cognitive capacity,' or adeptness at the game. As perhaps the most famous chess engine, and typically the strongest, Stockfish provides a great example of how engines evaluated positions and arrived at moves before the widespread usage of neural networks. This paper seeks to explore how evaluation features that Stockfish 8, an older version of Stockfish, may reflect what humans look for in a chess position by running regression and random forests on a Lichess data set. Attempting to predict whether or not players would make the correct move based on Stockfish-detected features did not result in any major insights into specific features that aid players across elo ratings, perhaps pointing to the relatively restrictive nature of Stockfish 8's native evaluation output and the large number of potential variables that humans may face when choosing their next move.