AAI_2025_Capstone_Chronicles_Combined

Evaluating Deep Learning Model Convergence in Chess via Nash Equilibria

7

Figure 3: The column chart denotes the piece count for each piece type given a class. Chess is known to become more drawish as pieces leave the board, and this notion is seen in the dataset. Draws usually have less of each type of piece relative to the decisive classes.

Chess is known to be drawish when the “heavy” pieces are off the board and the number of pieces in general are gone from the board. The bar chart above splits hero piece count into the three possible game result classes. As you can see, drawn games generally have less of every piece (except for the king of course). Even in games where the hero is lost, there is usually a larger number of pieces. This is often thought to be the case since positions with more pieces are more complicated, allowing for more mistakes to be made. Density charts for piece locations based where produced per game results. For all three game results, these charts look fairly similar qualitatively and quantitatively. One notable observation is that the pawns in front of the king do not move in general. Paired with the fact that the density chart shows that kings are often castle kingside, this shows the tendency of strong players to withhold advancing pawns in front of their own king. The pawn density chart also

81