FLBR-MWU achieves geometric last-iterate convergence, beating OGDA
After an open question, FLBR-MWU gets a concrete rate of convergence.
The paper, titled "Improved Last-iterate Convergence Properties for the FLBR-MWU Dynamics," revisits a variant of Multiplicative Weights Update (MWU) introduced by Fasoulakis et al. at AISTATS 2022. FLBR-MWU uses an extra-gradient-like approach but with different learning rates for the intermediate and final updates, giving it a forward-looking best-response flavor. While asymptotic last-iterate convergence was previously established, no explicit rate was known. The authors answer that open question by proving a geometric convergence rate: O(c^t), where c<1 depends on game parameters like the maximum eigenvalue of the Jacobian matrix. This bridges a critical gap between theory and practice.
To validate their analysis, the authors compare FLBR-MWU against Optimistic Gradient Descent-Ascent (OGDA), a state-of-the-art method for solving zero-sum games. Across their experiments, FLBR-MWU matches or, in some cases, outperforms OGDA, making it a strong candidate for practical applications. For professionals in AI and game theory, this means a more reliable and provably fast algorithm for training adversarial models, computing Nash equilibria, and building multi-agent systems.
- FLBR-MWU now has a geometric convergence rate of O(c^t) with c<1, solving an open problem from Fasoulakis et al. 2022.
- The convergence rate depends on game parameters such as the maximum eigenvalue of the Jacobian matrix, enabling tuning.
- In experiments, FLBR-MWU matches or outperforms OGDA, a leading last-iterate method for zero-sum games.
Why It Matters
Faster last-iterate convergence means more efficient and stable algorithms for adversarial training, GANs, and multi-agent learning.