Negative Momentum for Improved Game Dynamics

• DOI: 10.48550/arxiv.1807.04740
• Publication Date: 2018-01-01
• Authors: Gidel, Gauthier; Hemmat, Reyhane A...; Pezeshki, Mohammad; Lepriol, Remi; Huang, Gabriel; Lacoste-Julien, S...; Mitliagkas, Ioannis
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Statistics - Machine Learning; I.2.6; G.1.6; 0202 electrical engineering, electronic engineering, information engineering; Machine Learning (stat.ML); 02 engineering and technology; I.2.6; G.1.6; Machine Learning (cs.LG)

Appears in: Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS 2019). Minor changes with respect to the AISTATS version: typo corrected in Thm. 6 (squared condition number instead of condition number; and small change in constant) and dependence in $\beta$ changed in Theorem 5 for the formal statement; not changing the conclusions. 28 pages<br/>Games generalize the single-objective optimization paradigm by introducing different objective functions for different players. Differentiable games often proceed by simultaneous or alternating gradient updates. In machine learning, games are gaining new importance through formulations like generative adversarial networks (GANs) and actor-critic systems. However, compared to single-objective optimization, game dynamics are more complex and less understood. In this paper, we analyze gradient-based methods with momentum on simple games. We prove that alternating updates are more stable than simultaneous updates. Next, we show both theoretically and empirically that alternating gradient updates with a negative momentum term achieves convergence in a difficult toy adversarial problem, but also on the notoriously difficult to train saturating GANs.<br/>