Abstract:Moves in chess games are usually analyzed on a case-by-case basis by professional players, but thanks to the availability of large game databases, we can envision another approach of the game. Here, we indeed adopt a very different point of view, and analyze moves in chess games from a statistical point of view. We first focus on spatial properties and the location of pieces and show that the number of possible moves during a game is positively correlated with its outcome. We then study heatmaps of pieces and show that the spatial distribution of pieces varies less between human players than with engines (such as Stockfish): engines seem to use pieces in a very different way as human did for centuries. These heatmaps also allow us to construct a distance between players that characterizes how they use their pieces. In a second part, we focus on the best move and the second best move found by Stockfish and study the difference $\Delta$ of their evaluation. We found different regimes during a chess game. In a `quiet' regime, $\Delta$ is small, indicating that many paths are possible for both players. In contrast, there are also `volatile' regimes characterized by a `tipping point', for which $\Delta$ becomes large. At these tipping points, the outcome could then switch completely depending on the move chosen. We also found that for a large number of games, the distribution of $\Delta$ can be fitted by a power law $P(\Delta)\sim \Delta^{-\beta}$ with an exponent that seems to be universal (for human players and engines) and around $\beta\approx 1.8$. The probability to encounter a tipping point in a game is therefore far from being negligible. Finally, we conclude by mentioning possible directions of research for a quantitative understanding of chess games such as the structure of the pawn chain, the interaction graph between pieces, or a quantitative definition of critical points.

Estimating the number of reachable positions in Minishogi

Uct Based Search In Phantom Go

Proof Number Based Monte-Carlo Tree Search

An optimal MOO strategy

Upper Bounds for the Number of Solutions to Spatially Coupled Sudokus

Construction, observation and knowledge abstraction for go endgames on small boards

Belief-state Monte-Carlo Tree Search for Phantom Games

Evaluation of Monte Carlo tree search and the application to Go

Benchmark Instances and Branch-and-Cut Algorithm for the Hashiwokakero Puzzle

Gardner's Minichess Variant is solved

Solving 7x7 Killall-Go with Seki Database

Expected Work Search: Combining Win Rate and Proof Size Estimation

Strategy Complexity of Reachability in Countable Stochastic 2-Player Games

Shogi and Frieze group

Exact Method for Generating Strategy-Solvable Sudoku Clues

$ε$-Optimally Solving Zero-Sum POSGs

Mathematical analysis of the Royal Game of Ur

Statistical analysis of chess games: space control and tipping points

Approximate Integer Solution Counts over Linear Arithmetic Constraints

AlphaZero Gomoku