AAI_2025_Capstone_Chronicles_Combined

Evaluating Deep Learning Model Convergence in Chess via Nash Equilibria

9

Figure 5: The piece density gifs showcase how the hero and villain pieces evolve with total piece count on the board. As pieces are removed from the board, via the natural progression of the chess games, we observe how pawns move up the board and how the king's castle and activate themselves in the later stages of the game. This behavior in pawns and kings is characteristic of strong chess play and acts as a sanity check for the dataset. Despite the perfect information nature of the chess position, there are known complex relationships that can be observed throughout millions of high-level chess games. The tensor representation created for the chess Convolutional Neural Network encodes a variety of these features explicitly and implicitly. Explicitly the position of the pieces and the rules are encoded, while implicitly the data is structured in a way to show who’s turn it is and to easily allow the accounting for different piece types per square. Each plane of the 8x8x36 tensor should be

83