Decoupled SGDA for Games with Intermittent Strategy Communication

Ali Zindari, Parham Yazdkhasti, Anton Rodomanov, Tatjana Chavdarova, Sebastian U Stich

Abstract: We introduce *Decoupled SGDA*, a novel adaptation of Stochastic Gradient Descent Ascent (SGDA) tailored for multiplayer games with intermittent strategy communication. Unlike prior methods, Decoupled SGDA enables players to update strategies locally using outdated opponent strategies, significantly reducing communication overhead. For Strongly-Convex-Strongly-Concave (SCSC) games, it achieves near-optimal communication complexity comparable to the best-known GDA rates. For *weakly coupled* games where the interaction between players is lower relative to the non-interactive part of the game, Decoupled SGDA significantly reduces communication costs compared to standard SGDA. Additionally, *Decoupled SGDA* outperforms federated minimax approaches in noisy, imbalanced settings. These results establish *Decoupled SGDA* as a transformative approach for distributed optimization in resource-constrained environments.

Lay Summary:

Our research introduces a new method called Decoupled SGDA (Stochastic Gradient Descent Ascent) designed to improve how players update their strategies in online games, especially when communication between them is limited or intermittent.

Traditional methods require players to constantly exchange information, leading to high communication costs. Decoupled SGDA allows players to make their moves locally, even using slightly outdated information from opponents. This significantly reduces the amount of communication needed, making it more efficient for games where resources are constrained.

We show that for certain types of games (Strongly-Convex-Strongly-Concave games), our method is almost as efficient in terms of communication as the best existing approaches. For "weakly coupled" games, where players' interactions are less critical, Decoupled SGDA drastically cuts down communication costs compared to standard methods.

Furthermore, Decoupled SGDA performs better than other similar approaches in challenging conditions, such as when there's a lot of noise or imbalances in the game. These findings suggest that Decoupled SGDA is a promising and adaptable solution for optimizing distributed systems in various real-world scenarios with limited resources.

Application-Driven Machine Learning: This submission is on Application-Driven Machine Learning.

Primary Area: Optimization

Keywords: minimax games, Distributed Optimization

Submission Number: 11389