Tradeoffs Between Near-Nash Proximity and Reduced-Uncertainty Gain in TIOLI-GAN
Tetsuya Saito · 2025
This study investigates the equilibrium stability of TIOLI-GAN by analyzing its training dynamics using the difference in Jensen-Shannon divergence (∆JSD). Experimental results demonstrate that TIOLI achieves a local ϵ-Nash equilibrium with greater proximity than minimax GAN, stabilizing efficiently, while minimax GAN exhibits persistent fluctuations. The trade-off between ϵ and reduced-uncertainty gains (RUG) is empirically verified, confirming that a smaller ϵ restricts RUG, while a larger ϵ allows greater RUG at the cost of proximity to equilibrium. The leader-follower bargaining mechanism in TIOLI-GAN mitigates uncertainty, ensuring a structured and predictable learning process. These findings suggest that structured bargaining enhances convergence stability in adversarial learning.