AAI_2025_Capstone_Chronicles_Combined

Evaluating Deep Learning Model Convergence in Chess via Nash Equilibria

5 For example, the bishop can only move diagonally, but never through a piece. Meanwhile, the knight can travel through pieces while moving in an L-shape only. To make it easier for the model, I encoded these pseudo-legal moves into each position tensor. I say pseudo-legal since chess has many exceptions, but for the most part, the possible squares a piece a given hero or villain piece could traverse barring exceptions is encoded into the position. One limiting factor is that I could not quickly encode ALL legal moves into the position, since I found that was severely limiting the speed of the dataset generators during training. A more thorough implementation with more feature engineering should net even better results, since there is less information a model must learn to make informed decisions about the evaluation of the position.

Figure 1: The column chart showcases class distribution count weighted by importance. Importance is computed by taking a discount factor (0.99 in this experiment) and raising it to the number of ply (half-moves) from the final position of the game.

For my EDA, I created a set of visualizations to showcase how different game results correlate with types of positions, piece placements, and material imbalance. Given the class balance (from a small sample of 10k games) seemed fairly balanced, I quickly moved on to more thorough analysis of positional characteristics.

79