AAI_2025_Capstone_Chronicles_Combined

Evaluating Deep Learning Model Convergence in Chess via Nash Equilibria

14

Figure 8: Results of a round-robin tournament played between 10 model snapshots sampled from training. Each match between models entailed 5 pre-determined starting positions: the Sicilian, King’s Pawn, Indian Defense, Queen’s Pawn, and English Openings. 10 games were played each match, with each model getting a chance to play white and black on all 5 starting positions. Notice the strong winrates of the very first model, this model is the least trained, with only 100k gradient steps.

positions, resulting in a fair tournament with no repeated starting positions from the model’s perspective per match. The results of this tournament are seen in Figure 8. The winrate matrix already alludes to dramatic underperformance from models later in training.

88